Compressed sensing (CS) has been exploited for accelerating data acquisition in magnetic resonance imaging (MRI). MR images can be then reconstructed from significantly fewer measurements, i.e., drastically lower than that required by the Nyquist sampling criterion. However, the compressed sensing method usually produces images with artifacts, particularly at high reduction rates. In this paper, we propose a novel compressed sensing MRI method, called CS-NLTV that exploits curvelet sparsity (CS) and nonlocal total variation (NLTV) regularization. The curvelet transform is optimal sparsifying transform with the excellent directional sensitivity than that of wavelet transform. The NLTV, on the other hand extends the total variation regularizer to a nonlocal variant which can preserve both textures and structures and produce sharper images. We have explored a new approach of combining alternating direction method of multiplier (ADMM), adaptive weighting, and splitting variables technique to solve the formulated optimization problem. The proposed CSNLTV method is evaluated experimentally and compared to the previously reported high performance methods. Results demonstrate a significant improvement of compressed MR image reconstruction on four medical MRI datasets.

This work proposes a method for an effective and quick monitoring of video contents produced by TV Broadcasters by means of a fully automatic system. The proposed system performs acquisition, recognition and classification of logos labeling video contents hosted by video-sharing platforms. This challenge is addressed in the Laguerre-Gauss wavelet domain; as soon as a logo is located, in any area of the video screen, a detection strategy, based on the analysis of local Fisher information of the selected logo region, is applied. A distance metric on the LG saliency maps, based nearest neighbor algorithm, is defined, to classify the logo in the relevant video portion. A preliminary test on a dataset of 300 heterogeneous videos, produced by several European Broadcasters, was performed, to verify the effectiveness of the proposed method. The experimental results proved the robustness of the implemented logo recognition and classification method, also for video content labeled with different logo sizes and shapes and for video content corrupted by geometric transformations and/or coding degradations.

This paper introduces a texture representation suitable for image synthesis of textured surfaces. An efficient representation for natural images is of fundamental importance in image processing and analysis. The automated analysis of texture is widely applied in a number of real-world applications, e.g., image and video retrieval, object recognition and classification. For texture representation we consider the orthogonal decomposition of two-dimensional signals (images) using spectral transform in the different basis functions. This paper focuses on the analysis of the following basis functions Fourier, Walsh, Haar, Hartley and cosine transform using system criteria analysis. This criterion includes error signal representation and computational cost. For correct calculation of the components of the system criterion we use statistical averaging. It is shown that the Haar transform can represent textural patches more efficiently with smaller average risk than other basis functions. The texture representations results compare favourably against other state-of-the-art directional representations.

In color imaging, the two-dimension discrete Fourier transform (2-D DFT) is used to process separately color channels and in the quaternion space the processing of the image with 2-D quaternion Fourier transforms do consider interactions between the color channels. The color image can be also considered in different models with transformation to the octonion space with following processing in the 8-D frequency domain. In this work, we describe the algorithm for the 2-D two-side octonion DFT (ODFT), by using two-side 2-D quaternion DFTs.

This paper is devoted to analysis and further improvement of full-reference metrics of image visual quality. The effectiveness of a metric is characterized by the rank correlation factors between the obtained array of mean opinion scores (MOS) and the corresponding array of given metric values. This allows to determine the correspondence of a considered metric to a human visual system (HVS). Results obtained on the database TID2013 show that Spearman correlation for the best existing metrics (PSNRHMA, FSIM, SFF, etc.) does not exceed 0.85. In this paper, extended verification tools that allow to detect the shortcomings of the metrics taking into account combined distortions is proposed. An example for further improvement of the PSNRHMA metric is presented.

This work presents a background subtraction method based on Independent Components Analysis (ICA) implemented on a Field-Programmable Gate Array (FPGA) System-on-a-Chip (SoC) with an embedded processor. A previous work showed background subtraction utilizing ICA achieves better results than Mixture Of Gaussians (MOG) when the background is dynamic. However, that approach was developed for a computer and its proposed mean of using ICA is too complex, requiring a high-end computer for implementation. With the purpose of extending this approach to current embedded vision systems, a method is developed for an FPGA-SoC with embedded processor. This recent technology complements the parallelism of FPGAs with the general purpose computing of processors, maintaining a small footprint and a low power consumption. In this work, background is continuously updated by estimating the mean with Expectation-Maximization (EM), foreground is extracted via FastICA, and movement is determined by a threshold based on standard deviation. The herein presented approach to these algorithms allows exploiting the architecture of FPGA-SoCs, however, there are alternatives which could replace these algorithms. The developed method was tested on two image sequences and the accuracy of movement detection was measured by the precision and the recall.

Line segmentation performs a significant stage in the OCR systems; it has a direct effect on the character segmentation stage which affects the recognition rate. In this paper a robust algorithm is proposed for line segmentation for Arabic printed text system with and without diacritics based on finding the global maximum peak and the baseline detection. The algorithm is tested for different font sizes and types and results have been obtained from testing 5 types of fonts with total of 43,055 lines with 99.9 % accuracy for text without diacritics and 99.5% accuracy for text with diacritics.

We introduce a content-adaptive approach to image denoising where the filter design is based on mean opinion scores (MOSs) from preliminary experiments with volunteers who evaluated the quality of denoised image fragments. This allows to tune the filter parameters so to improve the perceptual quality of the output image, implicitly accounting for the peculiarities of the human visual system (HVS). A modification of the BM3D image denoising filter (Dabov et al., IEEE TIP, 2007), namely BM3DHVS, is proposed based on this framework. We show that it yields a higher visual quality than the conventional BM3D. Further, we have also analyzed the MOSs against popular full-reference visual quality metrics such as SSIM (Wang et al., IEEE TIP, 2004), its extension FSIM (Zhang et al., IEEE TIP, 2011), and the noreference IL-NIQE (Zhang et al., IEEE TIP, 2015) over each image fragment. Both the Spearman and the Kendall rank order correlation show that these metrics do not correspond well to the human perception. This calls for new visual quality metrics tailored for the benchmarking and optimization of image denoising methods.

The goal in photography is generally the construction of a model of scene appearance. Unfortunately, statistical variations introduced by photon shot and other noise introduce errors in the raw value reported for each pixel sample. Rather than simply accepting those values as the best raw representation, the current work treats them as initial estimates of the correct values, and computes an error model for each pixel's value. The value error models for all pixels in an image are then used to drive a type of texture synthesis which refines the pixel value estimates, potentially increasing both accuracy and precision of each value. Each refined raw pixel value is synthesized from the value estimates of a plurality of pixels with overlapping error bounds and similar context within the same image. The error modeling and texture synthesis algorithms are implemented in and evaluated using KREMY (KentuckY Raw Error Modeler, pronounced "creamy"), a free software tool created for this purpose.