Sport practice can take advantage from the quantitative assessment of task execution, which is strictly connected to the implementation of optimized training procedures. To this aim, it is interesting to explore the effectiveness of biofeedback training techniques. This implies a complete chain for information extraction containing instrumented devices, processing algorithms and graphical user interfaces (GUIs) to extract valuable information (i.e. kinematics, dynamics, and electrophysiology) to be presented in real-time to the athlete. In cycling, performance indexes displayed in a simple and perceivable way can help the cyclist optimize the pedaling. To this purpose, in this study four different GUIs have been designed and used in order to understand if and how a graphical biofeedback can influence the cycling performance. In particular, information related to the mechanical efficiency of pedaling is represented in each of the designed interfaces and then displayed to the user. This index is real-time calculated on the basis of the force signals exerted on the pedals during cycling. Instrumented pedals for bikes, already designed and implemented in our laboratory, have been used to measure those force components. A group of subjects underwent an experimental protocol and pedaled with (the interfaces have been used in a randomized order) and without graphical biofeedback. Preliminary results show how the effective perception of the biofeedback influences the motor performance.

In this work a novel technique for detecting and segmenting textured areas in natural images is presented. The method is based on the circular harmonic function, and, in particular, on the Laguerre Gauss functions. The detection of the textured areas is performed by analyzing the mean, the mode, and the skewness of the marginal densities of the Laguerre Gauss coefficients. By using these parameters a classification of the patch and of the pixel, is performed. The feature vectors representing the textures are built using the parameters of the Generalized Gaussian Densities that approximate the marginal densities of the Laguerre Gauss functions computed at three different resolutions. The feature vectors are clustered by using the K-means algorithm in which the symmetric Kullback-Leibler distance is adopted. The experimental results, obtained by using a set of natural images, show the effectiveness of the proposed technique.

Color images formed by modern digital cameras are often noisy, especially if they are captured in bad illumination conditions. This makes desirable to remove the noise by image pre-filtering. A specific feature of the noise observed for the considered application is that it can be spatially correlated. Filters to be applied have to effectively suppress noise introducing only negligible distortions into processed images. Moreover, such filters have to be fast enough and tested for a variety of natural images and noise properties. Another specific requirement is that a visual quality of processed images has to be paid a specific attention. To carry out intensive testing of some denoising approaches, a recently designed database TID2008 of distorted images provides a good opportunity since it contains 25 different images corrupted by i.i.d. and spatially correlated noise with several levels of variances. Taking into account the known fact that the color components are highly correlated, both modern 2D (component-wise) and 3D (vector) filtering techniques are studied. It is demonstrated that the use of 3D filters that allow exploiting inter-channel correlation provides considerably better results in terms of conventional and visual quality metrics. It is also shown how 3D filter based on discrete cosine transform (DCT) can be adapted to a spatial correlation of noise. This adaptation produces sufficient increase of the filter's efficiency. Examples of filter's performance are presented.

Text detection and recognition in natural images have conventionally been seen in the prior art as autonomous tasks executed in a strictly sequential processing chain with limited information sharing between sub-systems. This approach is flawed because it introduces (1) redundancy in extracting the same text properties multiple times and (2) error by prohibiting verification of hard (often binarized) detection results at later stages. We explore the possibilities for integration of detection and recognition modules by a feedforward multidimensional information stream. Integration involves suitable characterization of the text string at detection and application of the knowledge to ease recognition by a given OCR system. The choice of characterization properties generally depends on the OCR system, although some of them have proven universally applicable. We show that the proposed integration measures enable more robust recognition of text in complex, unconstrained natural environments. Specifically, integration by the proposed measures (1) eliminates textual input irregularities that recognition engines cannot handle and (2) adaptively tunes the recognition stage for each input image. The former function boosts correct detections, while the latter mainly reduces the number of false positives. Our validation experiments on a set of low-quality natural images show that adaptively tuning the OCR stage to the typical text-to-background transitions in the input image (gradient significance profiling) allows to attain an improvement of 29% in the precision-recall performance, mostly through boosting precision.

The use of ear information for people identification has been under testing at least for 100 years. However, it is still an open issue if the ears can be considered unique or unique enough to be used as biometric feature. In this paper a biometric system for human identification based on ear recognition is presented. The ear is modeled as a set of contours extracted from the ear image with an edge potential function. The matching algorithm has been tested in presence of several image modifications. Two human ear databases have been used for the tests. The experimental results show the effectiveness of the proposed scheme.

In this paper we present a method for extracting feature information from ladar data presented in the form of a spatial point cloud. The method exploits a modified version of Generalized Principal Component Analysis (GPCA) to extract planar, or even non-linear, surface elements from this sort of data. The essential difficulty is that, depending on the aspect of the object, certain surfaces will be minimally exposed. As a result we cannot say in advance how many surfaces we are looking for, and we cannot reliably detect surfaces that are hit by only a few of the points in the cloud. An additional difficulty occurs when reconstructing the surface normal at points near where two surfaces join. The algorithm handles both issues and captures enough essential surface features to allow accurate alignment to say a CAD model for detailed recognition. One can also use the extracted planar facets as a kind of partial bounding polyhedron (modified partial bounding box) as input to an initial identification algorithm that works off of the invariants of the planar arrangement.

The Fourier descriptors paradigm is a well-established approach for affine-invariant characterization of shape contours. In the work presented here, we extend this method to images, and obtain a 2D Fourier representation that is invariant to image rotation. The proposed technique retains phase uniqueness, and therefore structural image information is not lost. Rotation-invariant phase coefficients were used to train a single multi-valued neuron (MVN) to recognize satellite and human face images rotated by a wide range of angles. Experiments yielded 100% and 96.43% classification rate for each data set, respectively. Recognition performance was additionally evaluated under effects of lossy JPEG compression and additive Gaussian noise. Preliminary results show that the derived rotation-invariant features combined with the MVN provide a promising scheme for efficient recognition of rotated images.

We explore the one-to one correspondence between parametric surfaces in 3D and two dimensional color images in the RGB color space. For the case of parametric surfaces defined on general parametric domains recently a new approximate geometric representation has been introduced¹ which also works for manifolds in higher dimensions. This new representation has a form which is a generalization to the B´ezier representation of parametric curves and tensorproduct surfaces. The main purpose of the paper is to discuss how the so generated technique for modeling parametric surfaces can be used for respective modification (re-modeling) of images. We briefly consider also some of the possible applications of this technique.

The curvelet transform is a recently introduced non-adaptive multi-scale transform that have gained popularity in the image processing field. In this paper, we study the effect of customized tiling of frequency content in the curvelet transform. Specifically, we investigate the effect of the size of the coarsest level and its relationship to denoising performance. Based on the observed behavior, we introduce an algorithm to automatically choose the optimal number of decompositions. Its performance shows a clear advantage, in denoising applications, when compared to default curvelet decomposition. We also examine how denoising is affected by varying the number of divisions per scale.

We derive colour spaces of the hue-colourfulness-luminance type, on the basis of a four-dimensional hypercube I⁴ (I = [0, 1]). The hypercube corresponds to a tetrachromatic colour system, analogous to the three-dimensional RGB cube. In the first derived space the colourfulness is chromatic saturation while in the second one, colourfulness refers to the vividness of the colour, even if it is achromatic. The hue is defined on the basis of an icositetrahedron of 24 triangles that is embedded in the boundary of the hypercube. The boundary of the hypercube is the polytope {4 3 3} (in Sclafli notation) that is a topological 3-sphere. Out of the 24 square faces in the boundary of the hypercube, 6 meet the black vertex and 6 meet the white vertex; the remaining 12 faces form a dodecahedron which is a topological 2-sphere. This equatorial or chromatic dodecahedron is used to define a hue for each point in the hypercube that is not on the achromatic segment; the icositetrahedron results from a division of each of the square faces of the dodecahedron into two triangles. In addition, a hexdecahedron of 16 square faces with the topology of a torus that is also embedded in the boundary of the hypercube, is used to define an alternate two-dimensional hue space.

This paper introduces a new spatial edge oriented algorithm for automatic digital inpainting. The approach is based on the Laguerre Gauss analysis of the structure information of the regions surrounding the damaged portions of the image, extrapolating in automatic way the gradient of the luminance and color in missing areas this estimation is made of a least square fitting algorithm from simplified edge lines that stood on the boundary of missing region. The reconstruction of the unknown parts is automatically obtained by a variational method that uses the predicted gradient information imposing smoothing constraints on luminance and color level. Experiments on a number of images show the effectiveness of the proposed algorithm in smooth areas, as well as in areas with edges and/or textured.

In this paper an image inpainting approach based on the construction of a composite curve for the restoration of the edges of objects in an image using the concepts of parametric and geometric continuity is presented. It is shown that this approach allows to restore the curved edges in damaged image by interpolating the boundaries of objects by cubic splines. A tensor analysis is used for classification of texture and non texture regions. After edge restoration stage, a texture restoration based on exemplar based method is carried out. It finds the best matching patch from another source region and copies it into the damaged region. For non texture regions a Telea method is applied.

This paper is concerned with the development and implementation of a registration and stabilization method in conjunction with airborne imaging applications. We consider the situations for which the camera motion and vibration collectively affect the noisy image sequence. The general routine presented in this work is a combination of two algorithms for global image registration and image stabilization. We use and present experiments with real image sequences to track a moving object in the direction of its motion trajectory.

During the last decades, statistical models, such as the Ising model, have become very useful in describing solid state systems. These models excel in their simplicity and versatility. Furthermore, their results get quite often accurate experimental proofs. Leading researchers have used them successfully during the last years to restore images. A simple method, based on the Ising model, was used recently in order to restore B/W and grayscale images and achieved preliminary results. In this paper we outline first the analogy between statistical physics and image processing. Later, we present the results we have achieved using a similar, though more complex iterative model in order to get a better restoration. Moreover, we describe models which enable us to restore colored images. Additionally, we present the results of a novel method in which similar algorithms enable us to restore degraded video signals. We confront our outcomes with the results achieved by the simple algorithm and by the median filter for various kinds of noise. Our model reaches PSNR values which are 2-3 dB higher, and SSIM values which are 15%-20% higher than the results achieved by the median filter for video restoration.

Most of cameras follow pinhole camera model. However, result of this model makes some undesirable effects in wide angle lens. The most serious problem among these effects is radial distortion which appears heavily in fish-eye images. Several geometric models for correcting radial distortion of fish-eye lens are developed. Most of these models require only one parameter. However, correcting with one parameter is limited to correct both central and outer part simultaneously. Aim of this paper is to solve this problem. The proposed method is able to correct radial distortion of both areas using region adaptive distortion parameter. Each parameter is determined by considering amount of distortion in each region respectively. Also, the proposed method modifies the existing division model to correct radial distortion of both regions. Experimental results show that radial distortions in both areas are corrected.

Digital images and videos are used in many digital devices recently. Also, the resolution of display became larger than that of previous years. Image up-scaling algorithm is important issue since original input source is limited in transferring within data bandwidth. Among various up-scaling algorithms, Super-Resolution (SR) image reconstruction method is able to estimate high-resolution (HR) image using multiple low-resolution (LR) images. Conventional approaches to estimate HR image with Least Square (LS) method and Weighted Least Square (WLS) method are not able to reconstruct high-frequency region effectively in case its blur kernel is assumed Gaussian kernel in unknown system. Also, these methods produce jagging artifacts from the deficiency of LR frames. The proposed SR algorithm uses edge adaptive WLS method to reconstruct high-frequency region considering local properties and is applied to video sequence with block process to cope with local motions. Moreover, to apply video sequence with complex motions, we use selectively the correct information of reference frame to avoid errors from incorrect information. For accurate additional information from reference frames, the proposed algorithm determines additional information in reference frame by comparing with current frame and reference frame. The experiments demonstrate the superior performance of the proposed algorithm.

In this paper, new methods are addressed for impulse and speckle noise removal in images. The approach is based on the fusion of noise detection and image inpainting techniques. To avoid destroying the real structures of the image, the noise areas are first recognized to be repaired by an inpainting algorithm, subsequently. To distinguish the impulse noise areas in the image, a Neuro-Fuzzy model is employed and, to extract the speckled regions an algorithm is proposed based on Frost filtering and image resizing. The advantage of inpainting technique over the regular filtering methods is that it will be easier to generalize to all types of noise. Once we detect the damaged pixels in the image, the inpainting algorithm will be able to repair them. Various types of images under three levels of noise are tested using PSNR and SSIM measures. The experimental results demonstrate the great ability of the new approaches to suppress the noise properly, while preserving critical details of the image.

We have developed a novel interpolation method for images containing text, graphics and natural scenes. The method allows us to select the best interpolation algorithm for different regions of an image. In particular, we segment the image into graphical and natural regions and use the appropriate algorithm for each region. The natural regions are interpolated using a current state-of-the-art algorithm. However, when applied to graphical images, the current state-of-the-art interpolators tend to produce artifacts at edge discontinuities. Thus, we developed a novel approach which we call Low Entropy Interpolation (LEI) algorithm for the graphical images. The LEI algorithm is highly non-linear and produces very sharp edges with very few defects necessary for good quality interpolation of graphical images.

In this paper, we present a new image thresholding algorithm based on fractional filter (FF). Our experiments showed that a good segmentation result corresponds to an optimal order of the filter. Then, we propose to use geometric moments to find the optimal order. The proposed algorithm, called FLM, allows including contextual information such as the global object shape and uses the properties of the two-dimensional fractional integration. The efficiency of FLM was illustrated by the comparison to other six competing methods recently published and it was tested on real-world problem.

Image enhancement algorithms attempt to improve the visual quality of images for human or machine perception. Most direct multi-scale image enhancement methods are based on enhancing either absolute intensity changes or the Weber contrast at each scale, and have the advantage that the visual contrast is enhanced in a controlled manner. However, the human visual system is not adapted to absolute intensity changes, while the Weber contrast is unstable for small values of background luminance and potentially unsuitable for complex image patterns. The Michelson contrast measure is a bounded measure of contrast, but its expression does not allow a straightforward direct image enhancement formulation. Recently, a second derivative-like measure of contrast has been used to assess the performance of image enhancement algorithms. This measure is a Michelson-like contrast measure for which a direct image enhancement algorithm can be formulated. Accordingly, we propose a new direct multi-scale image enhancement algorithm based on the SDME in this paper. Experimental results illustrate the potential benefits of the proposed algorithm.

Kernel-design based method such as Bilateral filter (BIL), non-local means (NLM) filter is known as one of the most attractive approaches for denoising. We propose in this paper a new noise filtering method inspired by BIL, NLM filters and principal component analysis (PCA). The main idea here is to perform the BIL in a multidimensional PCA-space using an anisotropic kernel. The filtered multidimensional signal is then transformed back onto the image spatial domain to yield the desired enhanced image. In this work, it is demonstrated that the proposed method is a generalization of kernel-design based methods. The obtained results are highly promising.

In this paper, we solve the impulse noise detection problem using an intelligent approach. We use a multilayer neural network based on multi-valued neurons (MLMVN) as an intelligent impulse noise detector. MLMVN was already used for point spread function identification and intelligent edge enhancement. So it is very attractive to apply it for solving another image processing problem. The main result, which is presented in the paper, is the proven ability of MLMVN to detect impulse noise on different images after a learning session with the data taken just from a single noisy image. Hence MLMVN can be used as a robust impulse detector. It is especially efficient for salt and pepper noise detection and outperforms all competitive techniques. It also shows comparable results in detection of random impulse noise. Moreover, for random impulse noise detection, MLMVN with the output neuron with a periodic activation function is used for the first time.

In modern ultrasound imaging systems, the spatial resolution is severely limited due to the effects of both the finite aperture and overall bandwidth of ultrasound transducers and the non-negligible width of the transmitted ultrasound beams. This low spatial resolution remains the major limiting factor in the clinical usefulness of medical ultrasound images. In order to recover clinically important image details, which are often masked due to this resolution limitation, an image restoration procedure should be applied. To this end, an estimation of the Point Spread Function (PSF) of the ultrasound imaging system is required. This paper introduces a novel, original, reliable, and fast Maximum Likelihood (ML) approach for recovering the PSF of an ultrasound imaging system. This new PSF estimation method assumes as a constraint that the PSF is of known parametric form. Under this constraint, the parameter values of its associated Modulation Transfer Function (MTF) are then efficiently estimated using a homomorphic filter, a denoising step, and an expectation-maximization (EM) based clustering algorithm. Given this PSF estimate, a deconvolution can then be efficiently used in order to improve the spatial resolution of an ultrasound image and to obtain an estimate (independent of the properties of the imaging system) of the true tissue reflectivity function. The experiments reported in this paper demonstrate the efficiency and illustrate all the potential of this new estimation and blind deconvolution approach.

In recent years, various gesture recognition systems have been studied for use in television and video games[1]. In such systems, motion areas ranging from 1 to 3 meters deep have been evaluated[2]. However, with the burgeoning popularity of small mobile displays, gesture recognition systems capable of operating at much shorter ranges have become necessary. The problems related to such systems are exacerbated by the fact that the camera's field of view is unknown to the user during operation, which imposes several restrictions on his/her actions. To overcome the restrictions generated from such mobile camera devices, and to create a more flexible gesture recognition interface, we propose a hybrid hand gesture system, in which two types of gesture recognition modules are prepared and with which the most appropriate recognition module is selected by a dedicated switching module. The two recognition modules of this system are shape analysis using a boosting approach (detection-based approach)[3] and motion analysis using image frame differences (motion-based approach)(for example, see[4]). We evaluated this system using sample users and classified the resulting errors into three categories: errors that depend on the recognition module, errors caused by incorrect module identification, and errors resulting from user actions. In this paper, we show the results of our investigations and explain the problems related to short-range gesture recognition systems.

This paper proposes a method for tracking and recognizing the white line marked in the surface of the road from the video sequence acquired by the camera attached to a walking human, towards the actualization of an automatic navigation system for the visually handicapped. Our proposed method consists of two main modules: (1) Particle Filter based module for tracking the white line, and (2) CLAFIC Method based module for classifying whether the tracked object is the white line. In (1), each particle is a rectangle, and is described by its centroid's coordinates and its orientation. The likelihood of a particle is computed based on the number of white pixels in the rectangle. In (2), in order to obtain the ranges (to be used for the recognition) for the white line's length and width, Principal Component Analysis is applied to the covariance matrix obtained from valid sample particles. At each frame, PCA is applied to the covariance matrix constructed from particles with high likelihood, and if the obtained length and width are within the abovementioned ranges, it is recognized as the white line. Experimental results using real video sequences show the validity of the proposed method.

In an attempt to further develop and evaluate the optical recognition systems for distinguishing between driver and frontseat passenger during their interactions with dual-view touch screen integrated to the automobile centre console, this work focuses on the enhancement of both image processing algorithms and experimental environment. In addition to the motion based forearm and hand segmentation and the texture based arm direction analysis, the adaptive boosting classifiers with Haar-like features have been engaged for the learning of driver's and passenger's hand patterns. The user discrimination system was completely reproduced in a laboratory, including passenger compartment with genuine dashboard, touch screen, camera and near-infrared lamps, so that different illumination conditions could be modeled. The new acquisition system allows automatic and unambiguous registration of all touch screen interactions and their synchronization with the video stream. This results in credible evaluation of the image processing routines. The adjustment of the camera position and the active infrared illumination made it possible to reduce the recognition error rates and to achieve superior discrimination performance compared to previous works.

In this paper, we propose a new technique to automatically restore the sharpness of blurred documents by equalizing the frequency response of given scanners using linear filters. To measure the blur characteristics of a scanning device, we measure its both horizontal and vertical Spatial Frequency Response (SFR). Starting from the measured SFR of the scanning device, we design an equalizing filter so that the combined SFR of the equalizing filter and the scanner resembles a perfect SFR. The desired 2D frequency response of the filter is computed using linear interpolation of the horizontal and vertical responses derived from the corresponding SFRs of the scanner. The filter design technique is two steps. First, a linear system of equations is constructed using the unknown filter coefficients and the desired filter 2D frequency response. The linear least squares method is used to solve the linear system of equations. The second step of the filter design uses a nonlinear optimization technique to refine the results of the first step. Our experimental results show that this automated process can be applied to different document scanning devices to equalize their spatial frequency response resulting in consistent output sharpness levels.

Pixels' edges can yield useful information on physical properties of objects featured on satellite images. These properties can be derived through the use of the imagery spatial contrast techniques. To differentiate various cloud types based on their shapes, one of these techniques is applied on thermal products from a polar orbiting satellite, the National Oceanic and Atmospheric Administration/Advanced Very-High-Resolution Radiometer (NOAA-AVHRR). Edge gradients extracted from daily global cloud temperature images of this satellite and the spatial relationship between these gradients permit the distinction of nine major cloud shapes distributed along three cloud pressure levels (high, middle and low). The cloud shape differentiation method utilized is a histogram-based gradient scheme describing the occurrence of different gradients' levels (high, middle and low) in each block of pixels. A detailed analysis of the distribution of the cloud shapes obtained is conducted, and the frequency of each cloud shape is evaluated with another cloud classification method (based on cloud optical properties) for validation purposes. Finally, implications of the results obtained, on the estimation of the impact of cloud shapes variations on the recent climate are discussed.

The number of people being affected by Diabetes mellitus worldwide is increasing at an alarming rate. Monitoring of the diabetic condition and its effects on the human body are therefore of great importance. Of particular interest is diabetic retinopathy (DR) which is a result of prolonged, unchecked diabetes and affects the visual system. DR is a leading cause of blindness throughout the world. At any point of time 25 - 44% of people with diabetes are afflicted by DR. Automation of the screening and monitoring process for DR is therefore essential for efficient utilization of healthcare resources and optimizing treatment of the affected individuals. Such automation would use retinal images and detect the presence of specific artifacts such as hard exudates, hemorrhages and soft exudates (that may appear in the image) to gauge the severity of DR. In this paper, we focus on the detection of hard exudates. We propose a two step system that consists of a screening step that classifies retinal images as normal or abnormal based on the presence of hard exudates and a detection stage that localizes these artifacts in an abnormal retinal image. The proposed screening step automatically detects the presence of hard exudates with a high sensitivity and positive predictive value (PPV ). The detection/localization step uses a k-means based clustering approach to localize hard exudates in the retinal image. Suitable feature vectors are chosen based on their ability to isolate hard exudates while minimizing false detections. The algorithm was tested on a benchmark dataset (DIARETDB1) and was seen to provide a superior performance compared to existing methods. The two-step process described in this paper can be embedded in a tele-ophthalmology system to aid with speedy detection and diagnosis of the severity of DR.

Automatic surface inspection has been used in the industry to reliably detect all kinds of surface defects and to measure the overall quality of a produced piece. Structured light systems (SLS) are based on the reconstruction of the 3D information of a selected area by projecting several phase-shifted sinusoidal patterns onto a surface. Due to the high speed of production lines, surface inspection systems require extremely fast imaging methods and lots of computational power. The cost of such systems can easily become considerable. The use of standard PCs and Graphics Processing Units (GPUs) for data processing tasks facilitates the construction of cost-effective systems. We present a parallel implementation of the required algorithms written in C with CUDA extensions. In our contribution, we describe the challenges of the design on a GPU, compared with a traditional CPU implementation. We provide a qualitative evaluation of the results and a comparison of the algorithm speed performance on several platforms. The system is able to compute two megapixels height maps with 100 micrometers spatial resolution in less than 200ms on a mid-budget laptop. Our GPU implementation runs about ten times faster than our previous C code implementation.

Massively parallel computing (multi-core) chips offer outstanding new solutions that satisfy the increasing demand for high resolution and high quality video compression technologies such as H.264. Such solutions not only provide exceptional quality but also efficiency, low power, and low latency, previously unattainable in software based designs. While custom hardware and Application Specific Integrated Circuit (ASIC) technologies may achieve lowlatency, low power, and real-time performance in some consumer devices, many applications require a flexible and scalable software-defined solution. The deblocking filter in H.264 encoder/decoder poses difficult implementation challenges because of heavy data dependencies and the conditional nature of the computations. Deblocking filter implementations tend to be fixed and difficult to reconfigure for different needs. The ability to scale up for higher quality requirements such as 10-bit pixel depth or a 4:2:2 chroma format often reduces the throughput of a parallel architecture designed for lower feature set. A scalable architecture for deblocking filtering, created with a massively parallel processor based solution, means that the same encoder or decoder will be deployed in a variety of applications, at different video resolutions, for different power requirements, and at higher bit-depths and better color sub sampling patterns like YUV, 4:2:2, or 4:4:4 formats. Low power, software-defined encoders/decoders may be implemented using a massively parallel processor array, like that found in HyperX technology, with 100 or more cores and distributed memory. The large number of processor elements allows the silicon device to operate more efficiently than conventional DSP or CPU technology. This software programing model for massively parallel processors offers a flexible implementation and a power efficiency close to that of ASIC solutions. This work describes a scalable parallel architecture for an H.264 compliant deblocking filter for multi core platforms such as HyperX technology. Parallel techniques such as parallel processing of independent macroblocks, sub blocks, and pixel row level are examined in this work. The deblocking architecture consists of a basic cell called deblocking filter unit (DFU) and dependent data buffer manager (DFM). The DFU can be used in several instances, catering to different performance needs the DFM serves the data required for the different number of DFUs, and also manages all the neighboring data required for future data processing of DFUs. This approach achieves the scalability, flexibility, and performance excellence required in deblocking filters.

Processing and rendering of plenoptic camera data requires significant computational power and memory bandwidth. At the same time, real-time rendering performance is highly desirable so that users can interactively explore the infinite variety of images that can be rendered from a single plenoptic image. In this paper we describe a GPU-based approach for lightfield processing and rendering, with which we are able to achieve interactive performance for focused plenoptic rendering tasks such as refocusing and novel-view generation. We present a progression of rendering approaches for focused plenoptic camera data and analyze their performance on popular GPU-based systems. Our analyses are validated with experimental results on commercially available GPU hardware. Even for complicated rendering algorithms, we are able to render 39Mpixel plenoptic data to 2Mpixel images with frame rates in excess of 500 frames per second.

Gaussian Mixture Model (GMM) for background subtraction (BGS) is widely used for detecting and tracking objects in video sequences. Although the GMM can provide good results, low processing speed has become its bottleneck for realtime applications. We propose a novel method to accelerate the GMM algorithm based on graphics processing unit (GPU). As GPU excels at performing massively parallel operations, the novelty lies in how to adopt various optimization strategies to fully exploit GPU's resources. The parallel design consists of three levels. On the basis of first-level implementation, we employ techniques such as memory access coalescing and memory address saving to the secondlevel optimization and the third-level modification, which reduces the time cost and increases the bandwidth greatly. Experimental results demonstrate that the proposed method can yield performance gains of 145 frames per second (fps) for VGA (640*480) video and 505 fps for QVGA (320*240) video which outperform their CPU counterparts by 24X and 23X speedup respectively. The resulted surveillance system can process five VGA videos simultaneously with strong robustness and high efficiency.

In this paper, we study a plane-dependent technique that reduces dot-on-dot printing in color images, and apply this technique to a GPU-based error diffusion halftoning algorithm. We design image quality metrics to preserve mean color and minimize colorant overlaps. We further use randomized intra-plane error filter weights to break periodic structures. Our GPU implementation achieves a processing speed of 200 MegaPixels/second for RGB color images, and a speedup of 30 - 37x over a multi-threaded implementation on a dual-core CPU. Since the GPU implementation is memory bound, we essentially get the image quality benefits for free by adding arithmetic complexities for inter-plane dependency and error filter weights randomization.

In the development of commercial imaging based software applications there is the challenge of trying to provide high performance imaging algorithms that are utilized by multiple applications running on a range of hardware platforms. Many times the imaging algorithms will need to be run on workstations, smartphones, tablets, or other devices that may have different CPU and possibly GPU/DSP hardware. Implementing software on the cloud infrastructure can place limitations on the hardware capabilities imaging software can take advantage of. In the face of these challenges, OpenCL provides a promising framework to write imaging algorithms in. It promises that algorithms can be written once and then deployed on many different hardware configurations; GPU, DSP, CPU, etc... and take advantage of the computing features of particular hardware. In this paper we look at how well OpenCL delivers on this multi target promise for different image processing algorithms. Both GPU (Nvidia and AMD) and CPU (AMD and Intel) platforms are explored to see how OpenCL does in using the same code on different hardware. We also compare OpenCL with optimized CPU and GPU (CUDA) versions of the same imaging algorithms. Our findings are presented and we share some interesting observations in using OpenCL. The imaging algorithms include a basic CMYK to RGB color transformation, 25 x 25 floating point convolution, and visual attention saliency map calculation. The saliency map algorithm is complex and includes many different imaging calculations; difference of Gaussian, color features, image statistics, FFT filtering, and assorted other algorithms. Looking at such a complex set of algorithms gives a good real world test for comparing the different platforms with.

The Intensity Constrained Flat Kernel Filtering (ICFK) scheme is a dual domain (spatial and intensity) nonlinear framework which has been shown to generate useful filters for image processing. This paper proposes a new filter developed within the ICFK framework. Although local in nature the filter is designed to suppress large scale spatial features within the image. As in every other filter derived within the scheme the suppressed features are defined by two parameters: size of the kernel and intensity range. The filter, a single-step procedure, is applied to removal of hair artifacts in skin lesion epiluminescence microscopy images, the task essential in assisting in automated segmentation of imaged area into lesion and surrounding skin. Results of the experiments on 400 dermatoscopic images of lesions with hair indicate suitability of the method as an aid in lesion segmentation by suppressing hair or vascular features near the lesion borders.

In this paper, we build up a new decision support tool to improve treatment intensity choice in childhood ALL. The developed system includes different methods to accurately measure furthermore cell properties in microscope blood film images. The blood images are exposed to series of pre-processing steps which include color correlation, and contrast enhancement. By performing K-means clustering on the resultant images, the nuclei of the cells under consideration are obtained. Shape features and texture features are then extracted for classification. The system is further tested on the classification of spectra measured from the cell nuclei in blood samples in order to distinguish normal cells from those affected by Acute Lymphoblastic Leukemia. The results show that the proposed system robustly segments and classifies acute lymphoblastic leukemia based on complete microscopic blood images.

This paper presents a novel sharpness metric for color images. The proposed metric can be used for no-reference assessment of image visual quality. The metric basically relies on local power of wavelet transform high-frequency coefficients. It also takes into account possibility of presence of macrophotography and portrait photography effects in an image where the image part (usually central one) in sharp whilst the remained part (background) is smeared. Such effects usually increase subjective evaluation of image visual quality by humans. The effects are taken into consideration by joint analysis of wavelet coefficients with largest and smallest squared absolute values. Besides, we propose a simple mechanism for blocking artifact accounting (if an image is compressed by JPEG) and compensation of this factor contribution. Finally, the proposed sharpness metric is calculated in color space YCbCr as a weighted sum of sharpness components. Weight optimization has shown that a weight for intensity component Y is to be considerably smaller than weights for color components Cb and Cr. Optimization of weights for all stages of sharpness metric calculation is carried out for specialized database NRTID that contains 500 test images with previously determined MOS (Mean Opinion Score). Spearman rank order correlation coefficient (SROCC) determined for the designed sharpness metric and MOS is used as optimization criterion. After optimization, it reaches 0.71. This is larger than for other known available no-reference metrics considered at verification stage.

The aim of this study is to develop and evaluate a new set of color features and their performance on the classification of skin lesions. The proposed system introduces new features based on 2-dimensional color histograms, an automated segmentation method using a fusion of thresholding methods, classification procedures and is designed to be used by dermatologists as a complete integrated dermatological analysis tool to improve the rate of correct diagnosis above 90%. Simulations are implemented to show the measured features as well as classification results. The outcomes showed that the CAD model discussed in this paper has an improved classification performance and is an objective diagnostic tool that can be used in medical practice.

Spatial-temporal filters have been widely used in video denoising module. The filters are commonly designed for monochromatic image. However, most digital video cameras use a color filter array (CFA) to get color sequence. We propose a recursive spatial-temporal filter using motion estimation (ME) and motion compensated prediction (MCP) for CFA sequence. In the proposed ME method, we obtain candidate motion vectors from CFA sequence through hypothetical luminance maps. With the estimated motion vectors, the accurate MCP is obtained from CFA sequence by weighted averaging, which is determined by spatial-temporal LMMSE. Then, the temporal filter combines estimated MCP and current pixel. This process is controlled by the motion detection value. After temporal filtering, the spatial filter is applied to the filtered current frame as a post-processing. Experimental results show that the proposed method achieves good denoising performance without motion blurring and acquires high visual quality.

This paper discusses a procedure of filtering a three dimensional (3-D) polygonal mesh by utilizing the basic methods of finite impulse response (FIR), one dimensional (1-D) filtering. A method of linearizing a 3-D mesh was developed in order to apply the 1-D filter methods. With the development of low cost 3-D scanners, which physically scan and digitize real-world objects, the amount of "noise" that is found on these models has increased. This noise can come from various sources, including unwanted imperfections in the object itself, and from the device being used to scan the object. This newly developed filtering method will provide not only a way to decrease the noise in models, but increase details of the model as well, using low-pass and high-pass filters respectively.

CT (Computed tomography) is a widely employed imaging modality in the medical field. Normally, a volume of CT scans is prescribed by a doctor when a specific region of the body (typically neck to groin) is suspected of being abnormal. The doctors are required to make professional diagnoses based upon the obtained datasets. In this paper, we propose an automatic registration algorithm that helps healthcare personnel to automatically align corresponding scans from 'Study' to 'Atlas'. The proposed algorithm is capable of aligning both 'Atlas' and 'Study' into the same resolution through 3D interpolation. After retrieving the scanned slice volume in the 'Study' and the corresponding volume in the original 'Atlas' dataset, a 3D cross correlation method is used to identify and register various body parts.

Time/space varying filter banks (FBs) are proved to be useful in building signal adaptive transforms. Lifting factorization of FBs allows to spatially adapt between arbitrary FBs, avoiding the need to design border FBs to complete perfect reconstruction (PR) during the transition. However, lifting based switching between arbitrarily designed FBs induces spurious transients into the resulting subbands during the transition. In this paper we propose a boundary handling mechanism that maintains good frequency response and eliminates the transients during the transition. We successfully show spatial adaptation between JPEG2000 9/7 and 5/3 FBs to reduce the ringing artifacts in images.

Connected component labeling is a fundamental operation in binary image processing. A plethora of algorithms have been proposed for this low-level operation with the early ones dating back to the 1960s. However, very few of these algorithms were designed to handle color images. In this paper, we present a simple algorithm for labeling connected components in color images using an approximately linear-time seed fill algorithm. Experiments on a large set of photographic and synthetic images demonstrate that the proposed algorithm provides fast and accurate labeling without requiring excessive stack space.

Gray-level quantization (reduction) is an important operation in image processing and analysis. The Lloyd- Max algorithm (LMA) is a classic scalar quantization algorithm that can be used for gray-level reduction with minimal mean squared distortion. However, the algorithm is known to be very sensitive to the choice of initial centers. In this paper, we introduce an adaptive and deterministic algorithm to initialize the LMA for gray-level quantization. Experiments on a diverse set of publicly available test images demonstrate that the presented method outperforms the commonly used uniform initialization method.

Infrared Non-Destructive Testing (INDT) is known as an effective and rapid method for nondestructive inspection. It can detect a broad range of near-surface structuring flaws in metallic and composite components. Those flaws are modeled as a smooth contour centered at peaks of stored thermal energy, termed Regions of Interest (ROI). Dedicated methodologies must detect the presence of those ROIs. In this paper, we present a methodology for ROI extraction in INDT tasks. The methodology deals with the difficulties due to the non-uniform heating. The non-uniform heating affects low spatial/frequencies and hinders the detection of relevant points in the image. In this paper, a methodology for ROI extraction in INDT using multi-resolution analysis is proposed, which is robust to ROI low contrast and non-uniform heating. The former methodology includes local correlation, Gaussian scale analysis and local edge detection. In this methodology local correlation between image and Gaussian window provides interest points related to ROIs. We use a Gaussian window because thermal behavior is well modeled by Gaussian smooth contours. Also, the Gaussian scale is used to analyze details in the image using multi-resolution analysis avoiding low contrast, non-uniform heating and selection of the Gaussian window size. Finally, local edge detection is used to provide a good estimation of the boundaries in the ROI. Thus, we provide a methodology for ROI extraction based on multi-resolution analysis that is better or equal compared with the other dedicate algorithms proposed in the state of art.

Dynamic medical images may provide valuable information such as contraction rate, deformation and elasticity. For this purpose, it is fundamental to estimate the displacement of each point of interest. However, in ultrasound this task is hampered by speckle noise. The objective is the estimation of structure deformation and contraction using robust tracking of a set of representative points in a sequence of ultrasound images. The proposed approach is based on discrete optimization of joint displacement estimation via dynamic programming where the criteria involved joint intensity and morphology similarity. We also investigated the effect of initialization of the graph by maximization of Bhattacharyya coefficient. We evaluated in realistic numerical phantoms with speckle noise and compared with traditional approaches. Ten points were considered in the phantom and we applied several affine transformations to generate the deformed images as well as finite element based deformations. The average displacement error has decreased from 4.4 ± 6.6 pixels to 1.9 ± 2.5 pixels for the approach with proposed initialization with statistical significant difference at 5% level. In conclusion, we have shown that robust estimation of first point of contour provides a major improvement in the mapping of the contour points by dynamic programming.

Current face identification systems are not robust enough to accurately identify the same individual in different images with changes in head pose, facial expression, occlusion, length of hair, illumination, aging, etc. This is especially a problem for facial images that are captured using low resolution video cameras or webcams. This paper introduces a new technique for facial identification in low resolution images that combines facial structure with skin texture to accommodate changes in lighting and head pose. Experiments using this new technique show that combining facial structure features with skin texture features results in a facial identification system for low resolution images that is more robust to pose and illumination conditions than either technique used alone.

HuVisCam, a human vision simulation camera, that can simulate not only Purkinje effect for mesopic and scotopic vision but also dark and light adaptation, abnormal miosis and abnormal mydriasis caused by the influence of mydriasis medicine or nerve agent is developed. In this article, details of the system are described.

Despite major advances in x-ray sources, detector arrays, gantry mechanical design and special computer performances, computed tomography (CT) enjoys the filtered back projection (FBP) algorithm as its first choice for the CT image reconstruction in the commercial scanners [1]. Over the years, a lot of fundamental work has been done in the area of finding the sophisticated solutions for the inverse problems using different kinds of optimization techniques. Recent literature in applied mathematics is being dominated by the compressive sensing techniques and/or sparse reconstruction techniques [2], [3]. Still there is a long way to go for translating these newly developed algorithms in the clinical environment. The reasons are not obvious and seldom discussed [1]. Knowing the fact that the filtered back projection is one of the most popular CT image reconstruction algorithms, one pursues research work to improve the different error estimates at different steps performed in the filtered back projection. In this paper, we present a back projection formula for the reconstruction of divergent beam tomography with unique convolution structure. Using such a proposed approximate convolution structure, the approximation error mathematically justifies that the reconstruction error is low for a suitable choice of parameters. In order to minimize the exposure time and possible distortions due to the motion of the patient, the fan beam method of collection of data is used. Rebinning [4] transformation is used to connect fan beam data into parallel beam data so that the well developed methods of image reconstruction for parallel beam geometry can be used. Due to the computational errors involved in the numerical process of rebinning, some degradation of image is inevitable. However, to date very little work has been done for the reconstruction of fan beam tomography. There have been some recent results [5], [6] on wavelet reconstruction of divergent beam tomography. In this paper, we propose a convolution algorithm for the reconstruction of divergent beam tomography, which is simpler than wavelet methods and provides small reconstruction error. As the formula is approximate in nature, we prove an estimate for the error associated with the formula. Using the estimate, we deduce the condition that minimizes the approximation error.

Electro-optic (EO) images exhibit the properties of high resolution and low noise level, while it is a challenge to distinguish objects at night through infrared (IR) images, especially for objects with a similar temperature. Therefore, we will propose a novel framework of IR image enhancement based on the information (e.g., edge) from EO images, which will result in high resolution IR images and help us distinguish objects at night. Superimposing the detected edge of the EO image onto the corresponding transformed IR image is our principal idea for the proposed framework. In this framework, we will adopt the theoretical point spread function (PSF) proposed by Russell C. Hardie et al. for our IR image system, which is contributed by the modulation transfer function (MTF) of a uniform detector array and the incoherent optical transfer function (OTF) of diffraction-limited optics. In addition, we will design an inverse filter in terms of the proposed PSF to conduct the IR image transformation. The framework requires four main steps, which are inverse filter-based IR image transformation, EO image edge detection, registration and superimposing of the obtained image pair. Simulation results will show the superimposed IR images.

State-of-the-art results in image inpainting are obtained with patch-based methods that fill in the missing region patch-by-patch by searching for similar patches in the known region and placing them at corresponding locations. In this paper, we introduce a context-aware patch-based inpainting method, where the context is represented by texture and color features of a block surrounding the patch to be filled in. We use this context to recognize other blocks in the image that have similar features and then we constrain the search for similar patches within them. Such an approach guides the search process towards less ambiguous matching candidates, while also speeding up the algorithm. Experimental results demonstrate benefits of the proposed context-aware approach, both in terms of inpainting quality and computation time.