Filters are the essential elements in convolutional neural networks (CNNs). Filters generate feature maps and form the main part of the computational and memory requirements of the convolutional networks. In filter pruning methods, a filter with all of its components, including channels and connections, are removed. The removal of a filter can cause a drastic change in the network's performance. Also, the removed filters cannot come back to the network structure. We want to address these problems in this paper. We propose a CNN pruning method based on filter attenuation in which weak filters are not abruptly removed. Instead, weak filters are attenuated and gradually removed. In the proposed attenuation approach, there is a chance for weak filters to return to the network. The filter attenuation method is assessed using the VGG model for the Cifar10 image classification task. Simulation results show that the filter attenuation works well based on different pruning criteria, and better results are obtained in comparison with the conventional pruning methods.

Watermarking is a popular technique used in various applications, such as copyright protection of digital media, including audio, video, and image files. Proper watermarking should satisfy multiple criteria, such as robustness and transparency. While a successful watermarking needs to meet these criteria, there is a tradeoff between the two opposing criteria of robustness and transparency. This paper proposes a method for determining the appropriate locations for embedding watermarks with high strength factors. For this purpose, a deep neural network, known as Mask R-CNN, is used, which is pre-trained on the COCO dataset. This neural network finds a good strength factor for those sub-blocks of the host image selected for embedding. The proposed technique can be used in conjunction with most DWT and DCT based semi-blind watermarking approaches. Experiments show that the proposed method is robust against different attacks and demonstrates good transparency.

Deep neural networks have shown great achievements in solving complex problems. However, there are fundamental challenges which limit their real world applications. Lack of a measurable criterion for estimating uncertainty of the network predictions is one of these challenges. However, we can compute the variance of the network output by applying spatial transformations, distortions or noise injection to network inputs and interpret these variances as uncertainty of the network predictions. In other words, as long as the deformations do not conceptually alter target of interest, we expect the network to produce the same result. Hence, any outputs changes can be a sign of uncertainty in the network predictions. In order to estimate the prediction uncertainty of deep convolutional neural networks we use simple random transformations. By exploiting the network uncertainty, we improve the overall performance of the system. For a real use case, we apply the proposed method to segment left ventricle in MRI cardiac images. Experimental results demonstrate state-of- the-art performance and highlight the potential capabilities of simple ideas in conjunction with deep neural networks.

Estimating scene depth from a single monocular image is a crucial component in computer vision tasks, enabling many further applications such as robot vision, 3-D modeling, and above all, 2-D to 3-D image/video conversion. Since there are an infinite number of possible world scenes, that can produce a unique image, single image depth estimation is a highly challenging task. This paper tackles such an ambiguous problem by using the merits of both global and local information (structures) of a scene. To this end, we formulate single image depth estimation as a regression problem via (on) rich depth related features which describe effective monocular cues. Exploiting the relationship between these image features and depth values is adopted via a learning model which is inspired by modified stacked generalization scheme. The experiments demonstrate competitive results compared with existing data-driven approaches in both quantitative and qualitative analysis with a remarkably simpler approach than previous works.

Colonoscopy video frames might be contaminated by bright spots with unsaturated values known as specular reflection. Detection and removal of such reflections could enhance the quality of colonoscopy images and facilitate diagnosis procedure. In this paper, we propose a novel two-phase method for this purpose, consisting of detection and removal phases. In the detection phase, we employ both HSV and RGB color space information for segmentation of specular reflections. We first train a non-linear SVM for selecting a color space based on statistical image features extracted from each channel of the color spaces. Then, a cost function for detection of specular reflections is introduced. In the removal phase, we propose a two-step inpainting method which consists of appropriate replacement patch selection and removal of the blockiness effects. The proposed method is evaluated by testing on an available colonoscopy image database where accuracy and Dice score of 99.68% and 71.79% are achieved respectively.

Retinal vessel information is helpful in retinal disease screening and diagnosis. Retinal vessel segmentation provides useful information about vessels and can be used by physicians during intraocular surgery and retinal diagnostic operations. Convolutional neural networks (CNNs) are powerful tools for classification and segmentation of medical images. However, complexity of CNNs makes it difficult to implement them in portable devices such as binocular indirect ophthalmoscopes. In this paper a simplification approach is proposed for CNNs based on combination of quantization and pruning. Fully connected layers are quantized and convolutional layers are pruned to have a simple and efficient network structure. Experiments on images of the STARE dataset show that our simplified network is able to segment retinal vessels with acceptable accuracy and low complexity.

The need for CT scan analysis is growing for diagnosis and therapy of abdominal organs. Automatic organ segmentation of abdominal CT scan can help radiologists analyze the scans faster, and diagnose disease and injury more accurately. However, existing methods are not efficient enough to perform the segmentation process for victims of accidents and emergency situations. In this paper, we propose an efficient liver segmentation with our 3D to 2D fully convolution network (3D-2D-FCN). The segmented mask is enhanced using the conditional random field on the organ's border. Consequently, we segment a target liver in less than a minute with Dice score of 93.52%.

We propose a fast and effective method for multi-exposure image fusion. Our method blends multiple exposures under a base-detail decomposition of input images. Construction of blending weights in the proposed method is performed based on an exposedness function using luminance component of the input images. The fused base layer and detail layer are integrated into the final fused image which its detail strength is simply controlled through the integration process. Experimental results demonstrate that the proposed exposure fusion method is much faster than competing methods and can achieve state-of-the-art performance objectively and perceptually.

Quality assessment of stereo images confronts more challenges than its 2D counterparts. Direct use of 2D assessment methods is not sufficient to deal with the challenges of 3D perception. In this paper, an efficient general-purpose no-reference stereo image quality assessment, based on unsupervised feature learning, is presented. The proposed method extracts features without any prior knowledge about the types and levels of distortions. This property enables our method to be adaptable for different applications. The perceived contrast and phase of the binocular combination of original stereo images are utilized to learn individual dictionaries. For each distorted stereo image, two feature vectors are pooled, in a hierarchical manner, over all sparse representation vectors of phase and contrast blocks by their corresponding dictionaries. Performance results of learning a regression model by the features acknowledge the superiority of the proposed method to state-of-the-art algorithms.

Despite the rapid growth of 3D displays in the last few years, insufficient supply of 3D contents has led to considerable effort in devising 2D to 3D conversion algorithms. Inferring associated depth from single 2D image is still a controversial issue in these algorithms. In this paper we propose an algorithm, which unlike previous strategies, aggregates both global and local information from a pool of images with known depth maps. Hence, we propose to extract a set of features from the image patches of globally similar images in a large 3D image repository. These features describe powerful monocular depth perception cues. Using these relevant and robust features and using modified stacked generalization learning scheme, our scheme directly extracts an accurate depth map from given images. Experimental results demonstrate that our method has surpassed state-of-the-art algorithms in both quantitative and qualitative analysis.

Sparse representations over redundant dictionaries have shown to produce high quality results in various signal and image processing tasks. Recent advancements in learning of the sparsifying dictionaries have made image compression based on sparse representation a promising field. In this paper, we present a boosted dictionary learning framework to construct an ensemble of complementary specialized dictionaries for sparse image representation. Boosted dictionaries along with a competitive sparse coding can provide us with more efficient sparse representations. Based on the proposed ensemble model, we then develop a new image compression algorithm using boosted multi-scale dictionaries learned in the wavelet domain. Our algorithm is evaluated for compression of natural images. Experimental results demonstrate that the proposed algorithm has better rate-distortion performance as compared with several competing compression methods including analytic and learned dictionary schemes.

Coronary artery disease is one of the major causes of death throughout the world. An effective method for diagnosing this disease is X-ray angiography. The images are usually of poor quality and low contrast. This is due to non-uniform illumination, appearance of other body organs and artifacts, low SNR values, etc. Accurate segmentation of arteries is a challenging and important task. In this paper we first extract coronary arteries region of interest (ROI) using Hessian filter. Then, we combine these results with the flux flow measurements for accurate identification of vessel pixels. Post processing is performed to eliminate falsely identified vessel pixels. Finally, we segment the coronary arteries by selecting the largest connected component. Qualitative and quantitative evaluations of our method show high effectiveness of the proposed method. In terms of capturing major vessels our method is successful in 96% of cases.

Melanoma is the deadliest form of skin cancer. Diagnosis of melanoma in early stages significantly enhances the survival rate. Recently there has been a rising trend in web-based and mobile applications for early detection of melanoma using images captured by conventional cameras. These images usually contain fewer detailed information in comparison with dermoscopic (microscopic) images. Meanwhile, non-dermoscopic images have the advantage of broad availability. In this paper a set of ten features is proposed which cover different color characteristics of melanoma visible in skin images. The first 5 features are extracted using Fuzzy C-means clustering based on color variations and color spatial distributions of pigmented skin. These features are shown to be discriminative for melanoma lesions. The next 5 features consider colors and intensity of the colors. Hence, a 10 dimensional color feature space is formed. Experimental results show that classification accuracy of suspicious moles, by the proposed set of features, outperforms comparable state-of-the-art methods.

With the increasing use of telemedicine there is a great demand in real-time processing and transmission of medical images. Noise is one of the important factors that degrade the quality of medical images. Impulse noise is a common noise that could be caused by malfunctioning of sensors or by data transmission errors. It is one the most common noises that have extensively been studied in recent years. For real-time noise removal hardware techniques are more suited, since software methods are complex and slow. Usually hardware techniques have low complexity and low accuracy. In this paper a low complexity, high accuracy, de-noising method is proposed. It first categorizes image pixels into a number of groups. Then noisy pixels are restored in different ways in each category. Local analysis of image blocks allows us to restore a noisy pixel by using its neighboring non-noisy pixels. All steps are designed to have low hardware complexity. Simulation results show that in the case of MR images, the proposed method removes impulse noise with acceptable accuracy.

Melanoma is the most aggressive form of skin cancer and is on rise. There exists a research trend for computerized analysis of suspicious skin lesions for malignancy using images captured by digital cameras. Analysis of these images is usually challenging due to existence of disturbing factors such as illumination variations and light reflections from skin surface. One important stage in diagnosis of melanoma is segmentation of lesion region from normal skin. In this paper, a method for accurate extraction of lesion region is proposed that is based on deep learning approaches. The input image, after being preprocessed to reduce noisy artifacts, is applied to a deep convolutional neural network (CNN). The CNN combines local and global contextual information and outputs a label for each pixel, producing a segmentation mask that shows the lesion region. This mask will be further refined by some post processing operations. The experimental results show that our proposed method can outperform the existing state-of-the-art algorithms in terms of segmentation accuracy.

Touchless communication is a new field for commanding electronic devices. This method is highlighted when hygiene is a special issue. Automated hand gesture recognition needs processing of hand images. Many research works have tried to cope with this recognition problem. Complexity and high computational costs are important drawbacks that make real-time execution of these algorithms difficult. In this paper a new hand gesture recognition method is proposed. To show the functionality of our method we show how it can be used for recognition of the number of fingers in segmented images. Also the proposed algorithm can estimate angles of fingers, direction of the hand, and positions of fingers. In this work, we transform an image to intercept-slope coordinate using a proposed Radon transform inspired mapping. Using this mapping, the algorithm becomes invariant to rotation, scale and position. Straight and separated fingers will be extracted and their locations and angles are feasible to be determined as well. Simplicity and robustness against rotation, scaling and position and also having no complex mathematical calculation are advantages of our work.

Inferring scene depth from a single monocular image is an essential component in several computer vision applications such as 3D modeling and robotics. This process is an ill-posed problem. To tackle this challenging problem, previous efforts have been focusing on exploiting only global or local depth aware properties. We propose a model that incorporates both of them to obtain significantly more accurate depth estimates than using either global or local properties alone. Specifically, we formulate single image depth estimation as a K nearest neighbor search problem at both image level and patch level. At each level, a set of rich depth aware features, describing monocular depth cues, is employed in a nearest-neighbor regression model. By comparing the results with and without patch based fusion, the importance of our joint local-global framework becomes clear. The experimental results also demonstrate superior performance compared with existing data-driven approaches in both quantitative and qualitative analyses with a significantly simpler algorithm than others.

Sparse representations over redundant dictionaries have shown to produce high-quality results in various signal and image processing tasks. Recent advancements in learning-based dictionary design have made image compression using data-adaptive learned dictionaries a promising field. In this paper, we present a boosted dictionary learning framework to construct an ensemble of complementary specialized dictionaries for sparse image representation. Boosted dictionaries along with a competitive sparse coding form our ensemble model which can provide us with more efficient sparse representations. The constituent dictionaries of the ensemble are obtained using a coherence regularized dictionary learning model for which two novel dictionary optimization algorithms are proposed. These algorithms improve the generalization properties of the trained dictionary compared with several incoherent dictionary learning methods. Based on the proposed ensemble model, we then develop a new image compression algorithm using boosted multi-scale dictionaries learned in the wavelet domain. Our algorithm is evaluated for the compression of natural images. Experimental results demonstrate that the proposed algorithm has better rate-distortion performance as compared with several competing compression methods, including analytic and learned dictionary schemes.

X-ray angiography is a standard method for diagnosing coronary artery diseases. In order to show coronary arteries, contrast agent and X-ray imaging are used but the produced images are not always of adequate quality for visual examination. The low quality is caused by different artifacts. The presence of catheter and also the surrounding tissues make the processing of these images more difficult. In this paper, we propose a fully automated method to enhance the angiogram images and detect the arteries. Our proposed method contains three main steps which are Hessian filter enhancement, feature extraction and vessel region detection. To further enhance the visual quality of the image, non-vessel areas are blurred. The enhanced images can be utilized for better diagnosis of coronary diseases such as stenosis. Subjective evaluation of the enhanced images shows the effectiveness and accuracy of the proposed method.

Segmentation of X-ray bone images is of concern in many medical applications such as detection of osteoporosis and bone fractures. Segmentation of such images is a challenging process. Varying brightness throughout the image makes it difficult to separate bones from background and soft tissue. Costume made as well as standard segmentation methods, such as active contour and region growing, have been applied to bone X-ray images. Although each method could perform well for some images, due to variety of bone structures and lighting conditions none of these methods can be considered as complete. In this paper we present a new bone segmentation method in which an image goes through preprocessing steps such as noise cancellation and edge detection. Analysis of intensity fluctuations in all rows of the image results in more accurate segmentation of bone regions. Visual evaluation show that the proposed algorithm segments bones better than conventional and some recent bone segmentation approaches.

Multifocus image fusion is an important research area in image processing and machine vision applications. Due to use of optical lenses, captured images are not usually focused everywhere in the image. Therefore objects near the focal range have evident details while other objects appear blurry. Multifocus image fusion algorithm takes several images with different focal ranges and combines them to produce an image that is focused everywhere. To identify focused regions in each of input images, generally spatial domain and transform domain methods are used. These methods usually suffer from artifacts such as blockiness or ringing. In this paper we propose a new criteria to determine focused pixels in an image. We find the points that have the same intensities in input images and segment the input images based on them. Subsequently we calculate surface area of pixels inside every segment based on intensity variations over rows and columns. Segment with more surface area of input images selected as the focus segment. Experimental results reveal the superiority of our method in comparison to compared algorithms.

Retargeting algorithms are needed to transfer an image from a device to another with different size and resolution. The goal is to preserve the best visual quality for important objects of the original image. In order to reduce image size, pixels should be removed from less important parts of the image. Therefore, we need an energy function to select less important pixels in seam carving. Various energy functions have been proposed in previous works to minimize the distortion in salient objects. In this paper we combine three different importance maps to form a new energy map. We first use both gradient and depth maps to highlight the values in the saliency map, eventually generates the final energy map. Experimental results using the proposed energy map show better visual appearance in comparison to previous algorithms even at high resizing percentage. The visual artifacts that cause shape deformation in salient objects and deteriorates geometrical consistency of the scene are considerably reduced in our proposed algorithm.

RNA interference (RNAi) is considered one of the most powerful genomic tools which allows the study of drug discovery and understanding of the complex cellular processes by high-content screens. This field of study, which was the subject of 2006 Nobel Prize of medicine, has drastically changed the conventional methods of analysis of genes. A large number of images have been produced by the RNAi experiments. Even though a number of capable special purpose methods have been proposed recently for the processing of RNAi images but there is no customized compression scheme for these images. Hence, highly proficient tools are required to compress these images. In this paper, we propose a new efficient lossless compression scheme for the RNAi images. A new predictor specifically designed for these images is proposed. It is shown that pixels can be classified into three categories based on their intensity distributions. Using classification of pixels based on the intensity fluctuations among the neighbors of a pixel a context-based method is designed. Comparisons of the proposed method with the existing state-of-the-art lossless compression standards and well-known general-purpose methods are performed to show the efficiency of the proposed method.

As concerns about copyright protection increased amongst multimedia owners in recent years, many watermarking algorithms proposed to protect copyright of digital images. These methods are either spatial or frequency domain techniques. It is essential for a watermarking method to have acceptable robustness. That is why many existing methods try to improve their robustness against signal processing modifications. In this paper a block based watermarking scheme is proposed that embeds a binary logo into Contourlet coefficients of image. To increase robustness, embedding is done in two scales and watermark is inserted into DCT coefficients of Contourlet blocks to diffuse the effects of the embedding throughout the coefficients. Experimental results, and comparison with a robust Contourlet domain method, show that the proposed scheme has better robustness against some image processing attacks, while improving fidelity. Furthermore the proposed algorithm has the advantage of having a blind extraction phase.

Population of old generation that live alone is growing in most countries. Surveillance systems help them stay home and reduce the burden on the healthcare system. Automatic visual surveillance systems have advantages over wearable devices. They extract features from video sequences and use them for event classification. But these features are dependent on the position of cameras relative to the person. Therefore they need multi-camera for more accuracy that increases cost and complexity. In this paper we propose using silhouette area combined with inclination angle as robust features that can be measured using only one camera with an arbitrary direction. Through rigorous simulations on a publicly available dataset the error rate of the system is found to be less than 1%.

Three dimensional MRI images which are power tools for diagnosis of many diseases require large storage space. A number of lossless compression schemes exist for this purpose. In this paper we propose a new approach for the compression of these images which exploits the inherent symmetry that exists in the 3D MRI images. A block matching routine is employed to work on the symmetrical characteristics of these images. Another type of block matching is also applied to eliminate the inter-slice temporal correlations. The obtained results outperform the existing standard compression techniques.

Compression of images using transform methods has been of interest for many years. In this paper we propose a new multilayer image compression method which uses wavelet and contourlet transforms. We used structure tensor for identifying texture regions of the image by producing a binary mask. Then we apply wavelet to smooth regions and use contourlet transform for texture area. The proposed method avoids the redundancy of contourlet which has been a bottleneck for low bit rate compression purposes. We showed that images that are compressed and reconstructed by our method at low bit rates have good qualities both visually and in terms of the produced PSNRs.

In this paper a low complexity algorithm is proposed for near lossless compression of images. The reconstructed near lossless image can differ from the original one within a pixelwise error tolerance. This property is used to convert the histogram of the original image, by the proposed algorithm, to a new histogram which is proved to have minimum entropy. Hence, a new image is formed which has minimum entropy and high spatial correlation among its pixels and can efficiently be compressed. Simulation results show the effectiveness of this compression algorithm.

In this paper an algorithm is proposed which performs near-lossless image compression. For each pixel in a row of the image a group of value-states are considered, which have values close to that of the pixel. A trellis is constructed for every row of the image where the nodes of the trellis are the states of the pixels of that row. The goal of the algorithm is to find a path on this trellis that creates a sequence which can be efficiently coded using run length encoding (RLE). For sections of the pixels of the row that suitable RLE cannot be achieved then minimization of the entropy is employed to complete a path on the trellis. The application of the algorithm to a wide range of standard images shows that the scheme, while having low computational complexity, is competitive with other near-lossless image compression methods.

Microarray image technology is a powerful tool for monitoring the expression of thousands of genes simultaneously. Each microarray experiment produces large amount of image data, hence efficient compression routines that exploit microarray image structures are required. In this paper we introduce a lossless image compression method which segments the pixels of the image into three categories of background, foreground, and spot edges. The segmentation is performed by finding a threshold value which minimizes the weighted sum of the standard deviations of the foreground and background pixels. Each segment of the image is compressed using a separate predictor. The results of the implementation of the method show its superiority compared to the well-known microarray compression schemes as well as to the general lossless image compression standards.

In this paper, we present a pipeline architecture specifically designed to process and compress DNA microarray images. Many of the pixilated image generation methods produce one row of the image at a time. This property is fully exploited by the proposed pipeline that takes in one row of the produced image at each clock pulse and performs the necessary image processing steps on it. This will remove the present need for sluggish software routines that are considered a major bottleneck in the microarray technology. Moreover, two different structures are proposed for compressing DNA microarray images. The proposed architecture is proved to be highly modular, scalable, and suited for a standard cell VLSI implementation.