Vehicles have evolved into synthetic engineering systems owing to the developments in mechanical, electronic, and information technologies. Recently, driving assistance systems have gained considerable attention and are installed in vehicles under development by manufacturers. An important technology for the driving assistance systems is intervehicle distance measurement, for which several technologies have been proposed through hardware and software developments. The intervehicle distance measurement generally requires auxiliary sensor systems, complex algorithms, and high costs. An intervehicle distance estimation method is proposed, which calculates the distance using images obtained through a camera installed in the vehicle. For calculating the distance, the proposed method does not need any other physical sensors or information about the roads on which the vehicle moves. For estimating the intervehicle distance, the proposed method takes advantage of the fact that the width between two adjacent lanes on the road is constant. The difference in the distance calculated using the proposed method and the actual distance is <7  %  .

Because of limited labeled samples, semisupervised learning (SSL) methods have attracted much attention for classification of hyperspectral images (HSIs). Graph-based methods that treat data samples as nodes in a graph are very popular classes of SSL in the HSI data analysis. However, constructing a graph that can well capture the essential data structure is critical for these classes of SSL methods. A graph construction method based on low-rank representation (LRR) is proposed. Since LRR only captures the global structure of data, it cannot provide an informative graph for graph-based SSL tasks. To increase the effectiveness of the LRR-based graph, the local structure information is incorporated into the objective function of LRR as an additional penalty term. The proposed low-rank and local linear graph (LRLLG) takes the global and local structure into account, hence it provides a more generative and discriminative graph. Experimental results on two well-known data sets demonstrate that LRLLG outperforms the traditional graph construction methods in label propagation and graph-based SSL methods for HSIs.

A training scheme called region-refocusing (RR) is proposed to improve the accuracy and accelerate the convergence of compact one-stage detection neural networks. Main contributions are as follows: (1) the RR mask is first proposed to incorporate the position information and the significance of objects, whereby the regions containing objects can be learned selectively by the compact student detector, which leads to more reasonable feature expressions; (2) within the RR training framework, the selected objectness features from the large teacher detector are utilized to enrich the supervision information and enhance the loss functions for training the student detector, which eventually contributes to rapid convergence and accurate detection; (3) by virtue of the RR scheme, the mean average precision (mAP) of the compact detector can be significantly improved even if the model is initialized from scratch. Superiority of RR has been verified on several benchmark data sets in comparison with other training schemes; the mAP of the well-known tiny-YOLOv2 can be improved from 57.4% to 63.8% by 6.4 points on the VOC2007 test set when the weights are pretrained on ImageNet. Remarkably, when the pretraining process is omitted, it yields a significant boost of mAP by 22.6 points compared with plain training scheme, which demonstrates the robustness and high efficiency of the RR training scheme. Meanwhile, the compact one-stage detector trained with our framework is competent to be deployed on resource-constrained devices for the competitive precision as well as having a lower requirement for computing power.

Feature extraction is very important for steganalysis of content-adaptive JPEG steganography. The scale co-occurrence matrix feature based on a two-dimensional (2-D) Gabor filter is proposed, and diverse quantization for filter residuals is utilized to improve the detection performance. First, the definition of scale co-occurrence matrix based on a 2-D Gabor filter is given and the rules for feature merge are analyzed. Then, the influence of the scale parameter and quantization step on the detection performance of the scale co-occurrence matrix feature is discussed and verified. Next, the effect of diverse quantization strategy is presented. Last, the detailed extraction process of the proposed steganalysis feature is described. The experimental results show that the proposed steganalysis feature can achieve a performance that is competitive with the state-of-the-art steganalysis features when used for the detection of the latest content-adaptive JPEG steganography algorithms.

Owing to recent advancements, very deep convolutional neural networks (CNNs) have found application in image denoising. However, while deeper models lead to better restoration performance, they are marred by a high number of parameters and increased training difficulty. To address these issues, we propose a CNN-based framework, named dilated residual encode–decode networks (DRED-Net), for image denoising, which learns direct end-to-end mappings from corrupted images to obtain clean images using few parameters. Our proposed network consists of multiple layers of convolution and deconvolution operators; in addition, we use dilated convolutions to boost the performance of our network without increasing the depth of the model or its complexity. Extensive experiments on synthetic noisy images are conducted to evaluate DRED-Net, and the results are compared with those obtained using state-of-the-art denoising methods. Our experimental results show that DRED-Net leads to results comparable with those obtained using other state-of-the-art methods for image denoising tasks.

Sketching has become fashionable with the increasing availability of touch-screens on portable devices. It is typically used for rendering the visual world, automatic sketch style recognition and abstraction, sketch-based image retrieval (SBIR), and sketch-based perceptual grouping. How to automatically generate a sketch from a real image remains an open question. We propose a convolutional neural network-based model, named SG-Net, to generate sketches from natural images. SG-Net is trained to learn the relationship between images and sketches and thus makes full use of edge information to generate a rough sketch. Then, mathematical morphology is further utilized as a postprocess to eliminate the redundant artifacts in the generated sketches. In addition, in order to increase the diversity of generated sketches, we introduce thin plate splines to generate more sketches with different styles. We evaluate the proposed method of sketch generation both quantitatively and qualitatively on the challenging dataset. Our approach achieves superior performance to the established methods. Moreover, we conduct extensive experiments on the SBIR task. The experimental results on the Flickr15k dataset demonstrate that our proposed method leverages the retrieval performance compared with the state-of-the-art methods.

Discriminative correlation filters (DCFs) have shown excellent performance in visual tracking. DCF substitutes the sliding windows sampling strategy in traditional tracking methods with circular shift of the context area. Via projecting the filter learning into the frequency domain, DCF achieves satisfying performance and speed. Appropriate context area size has an influence on the performance of correlation filters. Small context area limits the CF’s ability to handle fast motion and partial occlusion, whereas large context area leads the CF to suffer from boundary effect. To make use of a large area of context and alleviate the accompanying drift risk, we propose a mask-constrained context correlation filter for object tracking. We first analyze the traditional window strategy via Taylor series and design a spatial mask that can be covered by a larger context area. Furthermore, the shape of the mask is adaptive to the target variation. Extensive experimental results in OTB-2015, VOT-2014, and VOT-2016 datasets demonstrate that this mask-constrained operation can improve the CF tracker performance in a large margin.

We introduce an alternative hybrid swarm algorithm for image segmentation that employs multilevel thresholding techniques. For the hybridization, we have combined the whale optimization algorithm (WOA) and the particle swarm optimization (PSO). The proposed method is called WOAPSO, and it operates in a cooperative environment, where the initial population is divided into two subpopulations (the first subpopulation is assigned for WOA and the other is assigned for PSO). Then, the WOA and the PSO operate in parallel during the iterative process to update the solutions and the best solution is selected from the union of the updated subpopulations according to the objective function. Here, two objective functions are used, the Otsu’s method and the fuzzy entropy method. These functions evaluate the quality of the thresholds generated by the WOAPSO considering the variance and the entropy of the classes where the pixels are cataloged. The experimental results and comparisons provide evidence of the ability of the proposed WOAPSO algorithm to reduce the time complexity without affecting the accuracy of the solutions.

We present a liquid-level measurement system based on second-confirm recognition algorithm. Compared with the traditional methods of the liquid-level measurement system, which has low accuracy, slow measurement speed, and limited range of measurements, this system can accurately and quickly measure the liquid level. Moreover, it can automatically complete the measurement of the liquid-level line position. The main component of the system includes a vision detection system and a tracking control system. The vision detection system mainly includes image preprocessing, numeral recognition, and liquid-level line recognition. Among the vision detection system, numeral recognition is the key task. Therefore, we propose a numeral recognition algorithm that employs a second-confirm algorithm to recognize numerals. The algorithm first utilizes template matching to find one numeral from the image. Afterward, we can divide the numeral image out of the rule image. This can greatly optimize the issues of time-consumption. The numerals can be segmented from the numeral image when the numeral image is obtained. After the numerals are captured, the numerals and template image difference is computed to complete numeral recognition. In addition, tracking control system tracks the liquid line in real time through the level value feedback by the vision system, which calculates the position of the liquid level in the image and feeds back the level value to the tracking control system. After experimental verification, the system can meet the expected requirements of liquid-level measurement.

Although discriminative correlation filter-based (DCF) trackers have shown excellent performance in recent years, it still has the problem of model drift. A generic approach alleviating model drift is proposed for DCF tracking. Specific correlation filters are learned in feature channels and the combination of response maps depends on the confidence level of corresponding channels. Model drift is eliminated by allowing contaminated correlation filters to retrogress to the ones that are not influenced by impure samples. As a result, correlation filters are purified by neglecting the impure part of them. Meanwhile, both the confidence level of detection and the variant level of interframe feature space are considered during online update of correlation filters. Our approach is experimented on both OTB-50 and OTB-100 datasets. Especially on OTB-100, our approach outperforms the baseline fast discriminative scale space tracking by 6.9%, 8.1%, and 5.6% in mean distance precision, mean overlap precision, and the area-under-the-curve score, respectively. Moreover, the competitive performance is presented by comparing with other state-of-the-art tracking algorithms.

A hybrid system for pedestrian detection is proposed, in which both visible camera and light detection and ranging (LIDAR) data of the same scene are used. The proposed method is achieved in two main stages. First, the 3-D LIDAR points are clustered and the resulting clusters are projected onto the visible image to get regions of interests (ROIs). A postprocessing method is applied on the resulting ROIs to keep only significant ones. Second, a hybrid feature is defined to classify the ROIs extracted from the first stage. It involves a combination of visual (camera) and range (LIDAR)-based features. The visual feature is a combination of the so-called HOG-Color, which is an extension of the classical histogram of oriented gradients (HOG) feature to the RGB color space and the local self-similarity feature. The range feature we define is called histogram of dominant silhouette orientation descriptor. It is computed from the depth values of the pixels belonging to the ROI. Random forest, support vector machine, and Adaboost classifiers have been tested together with the introduced hybrid feature and the first one is adopted as it outperforms the two other ones. The proposed method has been tested on KITTI dataset and the results are satisfactory when compared to recent state-of-the-art methods.

Sparse representation (SR) has shown great potential in multishot person reidentification. However, conventional SR methods usually encode each probe image individually, which ignores the prior knowledge shared by similar probe images with the same identity. Considering that probe images with the same identity are similar and can be complementary to each other, we propose a joint group sparse representation (JGSR) method to simultaneously encode a set of probe person images with the same identity. Additionally, we consider the fact that prior knowledge is also contained in the set of images with the same identity in the gallery. In JGSR, probe images with the same identity are coded on the same dictionary atoms and shared similar coefficients. Furthermore, the person reidentification performance is also improved by incorporating JGSR with k-means clustering and kernel local Fisher discriminant analysis into a unified framework. Extensive experiments on three publicly available iLIDS-VID, PRID 2011, and SAIVT-SoftBio multishot benchmark datasets were conducted and demonstrated the superior performance of the JGSR method in comparison with current state-of-the-art methods.

Intensity inhomogeneity often occurs in real-life images because of nonuniform illumination, device operating, and technical limitation. We propose a K-means++ clustering-based active contour model for fast image segmentation. The key point is two fitting functions whose value computed by K-means++ clustering algorithm before level set function evolution. At first, we set up a rectangular local window. Two fitting functions represent the center points of brighter and darker subregions in the moving rectangular local windows. The method avoids repeating calculation of the fitting functions during curve evolution compared with the traditional region-based active contour models. Therefore, the proposed model has lower computational costs, and we can obtain correct segmentation results in less time and fewer iterations. The proposed model can efficiently segment images with intensity inhomogeneity. In addition, the experiments have proved that the proposed model has strong robustness to initialization.

Single image dehazing is a challenging problem because of its severely ill-posed nature. Single image dehazing mainly includes two important tasks, estimating the atmospheric light and calculating the medium transmission map. A single image dehazing algorithm using heterogeneous atmospheric light estimation is presented to enhance the quality of hazy images. First, a max-pooling-based strategy is developed to estimate the heterogeneous atmospheric light, which can accurately describe how the amount of scattering from one or multiple point light sources is spatial-variant. And then, an efficient adaptive filtering approach is introduced to refine the medium transmission map and suppress the annoying halo artifacts produced during the process of dehazing. Finally, based on the heterogeneous atmospheric scattering model, a photorealistic haze-free image can be recovered with the estimated atmospheric light and the refined medium transmission map. The experimental results on a variety of real-world images and synthetic hazy ones demonstrate that the addressed method outperforms state-of-the-art approaches through qualitative and quantitative metrics.

This work is focused on the robustness of the ultimate opening (UO) and its associated function (AFUO) related to illumination changes and marker detection. It is well known that morphological filters commute to anamorphoses. This property ensures the robustness of the morphological filters to the lighting changes. However, the supremum of residues of these filters requires some conditions to commute to anamorphoses. Therefore, initially, a study of the robustness to anamorphoses is carried out. We demonstrate that the UO is invariant to linear anamorphoses and it is robust to power anamorphoses. On the other hand, related to marker detection, a study of the UO and AFUO is carried out. It is detected that small pieces of discs corrupt the UO, complicating use in marker detection. This behavior occurs because opening residues are not idempotent, so a modified UO is introduced to attenuate this inconvenience enabling a better selection of markers. Likewise, an example of a better selection of markers is carried out from images previously filtered and enhanced. Finally, the use of these transformations in image segmentation is illustrated.

Video surveillance systems have become one of the most useful entities in our routine life. Surveillance videos contain plenty of visual information about criminal actions happening in the field-of-view. With the increase of criminal activities, it is mandatory to develop the accurate criminal recognition system. Our paper aims to propose and evaluate action recognition system for the recognition of criminal actions. First, a descriptor is proposed as spatiotemporal human motion acceleration (ST-HMA) over improved dense trajectories (IDT) framework to correctly recognize the criminal actions. Second, a hybrid dataset is developed by the combination of criminal activities, e.g., fight, kick, push, punch, shoot gun, and sword fighting collected from state-of-the-art datasets named as hybrid criminal action (HCA) dataset. The dataset covers the common on-street criminal action poses. We have also evaluated different descriptors over the IDT framework. The achieved accuracies per class are 92.85%, 92.85%, 93.33%, 96.16% for kick, push, punch and fight actions, respectively. Experimental results show that ST-HMA on IDT framework gives better results than HMA descriptor in edge trajectory framework. The proposed framework also achieved high average accuracy rate of 80.89% for ST-HMA descriptor over IDT. Different descriptor applied over IDT also shows good action recognition accuracy for HCA dataset.

Object recognition plays an important role in applications such as surveillance, perimeter, and traffic monitoring. Imaging in thermal infrared (IR) band provides the benefits of operability in no light and harsh weather conditions unlike visible band imaging, which is dependent on illumination. However, infrared imagery poses challenges such as low contrast, lack of sharp boundaries and features, and high variability of object’s signature, making object recognition in infrared band very challenging, demanding robust recognition framework. We have evaluated state-of-the-art feature extractors in a bag of words (BoW) framework for their suitability for IR images and present an optimized bag of features framework for object recognition in IR images. Another bottleneck for target classification research in IR is the limited scope of available databases, which do not capture variations of target categories. To address this issue, a richly diverse IR dataset containing images captured under diverse conditions affecting the infrared signature of the objects, environmental and background factors such as season, time of day, experimental sites, target factors in terms of thermal emission characteristics, pose, distance from the sensor, and technological factors such as camera resolution, sensitivity is presented. The IR dataset consists of 3841 images representing six image categories primarily of urban scenes. We present a hierarchical three-stage classification in an infrared domain, primary classification as {target, background}, secondary classification of target subcategory as {vehicles, pedestrians}, and tertiary classification of vehicle subcategory as {two-wheelers, three-wheelers, light motor vehicles, heavy motor vehicles}. We report the performance of state-of-the-art feature extractors in a BoW framework establishing their limits of performance using evaluation metrics—accuracy, precision, sensitivity, specificity, and F-score.

Recent research on image translation has great progress with the development of generative adversarial networks (GANs) techniques. Generating high-resolution images with unsupervised architecture is one of the most challenging tasks for image translation. To this end, we propose an enhanced super-resolution generative adversarial network for image translation. First, for unlabeled datasets, we employ reconstructed consistency loss and mutual dual GANs, which contains two generators:G_{A  →  B}, G_{B  →  A} and two discriminators: D_B, D_A to develop an unsupervised learning framework. As reconstructed consistency loss is added between generators of G_{A  →  B} and G_{B  →  A}, our designed overall architecture can learn mapping function of different domains even without unpaired samples. In addition, the generator network includes encoder network, decoder network, and residual network with skip connections to generate high-resolution images with realistic details. Meanwhile, a stable normalization is proposed to stabilize the training of our discriminator networks. Finally, experimental results are carried out on six different datasets, demonstrating that our algorithms outperform the state-of-the-art methods in terms of the image quality and image resolution.

Recently, microwave computational imaging systems have had various applications ranging from security screening to biomedical diagnosis. However, existing methods are sensitive to noise and have a heavy computational burden in a three-dimensional (3-D) imaging scene. A computational imaging method approached in frequency domain is proposed, which improves imaging quality under noisy conditions and reduces computation complexity. The signal-to-noise ratio of the echo signal is improved by dechirping pulse compression method, which obtains the range resolution concurrently. According to the information of range resolution, the scene is divided into some range bins. With computational imaging algorithms, the azimuth and elevation resolution are obtained in each range bin by spatially diverse patterns of reprogrammable metasurface. A sparse 3-D image can be obtained by combining the reconstructed subimages. The simulation result shows that the proposed method outperforms the conventional methods with better antinoise ability and lower computational complexity in sparse 3-D scene imaging.

The accuracy and efficiency of single-image super-resolution (SR) using techniques based on convolutional neural networks have recently shown much improvement. However, most of the existing algorithms aim at improving the peak signal-to-noise ratio by minimizing the mean squared error between the ground-truth images and the generated SR images. This leads to a lack of high-frequency information and nonconformance with the perception of human eyes. To reconstruct realistic natural images in SR with large up-sampling factors, we combine the benefits of some recent approaches and propose a method based on autoencoding adversarial networks. The proposed architecture includes a generator, which is a symmetric encode–decode network used to extract feature maps and recover high-resolution images, and a conditional discriminator, which is used to determine whether the generated image is from the real distribution or not. In addition, we extract high-level features from a pretrained network to optimize the perceptual loss and make the output more precise. Compared with several state-of-the-art methods, our proposed method shows outstanding performance in recovering fine texture details. The mean opinion score shows that our method yields results that are more satisfactory to human perception than the other methods under comparison.

A high-fidelity reversible data hiding (RDH) algorithm using a discrete cosine transform (DCT) coefficients energy analysis-based predictor and optimized embedding is proposed. The algorithm utilizes the characteristic of energy distribution of natural image DCT coefficients to predict pixel values, which improves the prediction accuracy over conventional methods. Moreover, the algorithm establishes a pure load-image distortion function, which can be used to find optimized pixel expansion direction, secret message embedding location, and pixel sorting/selection threshold in each round of the multipass embedding process. Thus, the algorithm can better balance both image quality and data hiding capacity. Experimental results show that for high-fidelity RDH applications, comparing with several classical methods, the proposed algorithm always produces a high quality stego image in the same embedding capacity.

Privacy of information is one of the utmost vital issues in the epoch of digitally advanced communication. Privacy of information can be achieved through confidentiality, which is one of the most important pillars in information security. There are varieties of protocols and algorithms either symmetric or asymmetric claiming maximum security depending on key lengths, rounds, S-boxes, etc. Modern systems are based on mathematical and logical functions, and their algorithms are based on the fundamental process of factoring large integers into their primes. With the passage of time, there have been advancements in the field to show that, if the key length is not enough, substitution boxes are not too strong, or rounds of encryption are few, then the designed encryption scheme is vulnerable to threats. We have established a scheme using quantum spinning and rotation that is applied to both symmetric and asymmetric cryptographies with minimum key space and that provides maximum security. Our scheme is purely based on rotation of phases that adds a concept in quantum cryptography. The scope of this paper is to cover the fundamental concepts of spins ½ matrices (Pauli’s matrices) and their general rotation operators as applied to asymmetric systems.

In the last two decades, in the domain of video coding and compression, researchers have suggested several techniques for computation and time reduction for motion estimation (ME). We present a motion estimation algorithm for x265 video codec, based on a deterministic initial population in the genetic algorithm (GA). GA is known for its adaptive convergence, which is motivated by the biological process of survival of the fittest. The suggested scheme is targeted for the reduction of search points (SP) in a block matching motion estimation algorithm for integer-pel in B and P frames that are set to have three reference frames. The initial population constituted in our approach is a function of pre-encoded coding units at different spatial–temporal locations of the video frames and predefined hexagonal (HEX) locations. We propose a “deterministically starting” GA (GA_Det), toward deployment in x256 structure. In the framework of x265 code, GA_Det is found to offer reduction in SP at selected classes of videos considered for experimentation. To demonstrate the effectiveness of the proposed work, results have been compared with the block-based fast-full-search algorithm and the HEX search algorithm from the reference software. Traditional GA with a randomly constituted initial population, labeled as GA_Stc, is also implemented and an empirical comparison is carried out with GA_Det. The proposed GA_Det framework provides reduction in motion estimation time while rendering acceptable peak signal-to-noise ratio loss and an increase in a bit rate.

In outdoor low-level vision systems, not only is the resolution of the imaging system important, but rain corrupts the visibility of outdoor scenes and may cause computer vision systems to fail. We present a deep convolutional neural network (CNN) architecture for simultaneously performing single-image super-resolution and rain removal. Instead of learning an end-to-end mapping between the low-resolution rainy images and high-resolution clean images in the original image space, we train our network in the detail space, i.e., the space obtained by high-pass filtering the original image. The proposed CNN has a lightweight structure, yet it outperforms super-resolution and rain removal consecutively by a significantly large margin (>1  dB on average).

The sparse-to-dense approach is considered to be the standard method of capturing the long-range motion of small objects during large displacement optical flow. Despite progress in the matching and interpolation of this approach, little work has focused on improving the handling of outliers after dense sampling descriptor matching. We propose an improved grid-based statistical matching method that can quickly remove outliers without calculating backward flow. First, a multigrid statistical matching method is developed to remove the most outliers of the dense sampling descriptor correspondence field. Second, to improve the accuracy of outliers handling, the misjudgment match in the edge grid is corrected based on the statistical matching constraint. The results of extensive experiments on public optical flow datasets demonstrate the effectiveness of the proposed method.

Image quality is affected by various factors, especially low illumination. We propose an image color enhancement method to improve the brightness and quality of images. First, we apply local spatial homomorphic filtering to images to improve the brightness and avoid edge block effect as well as image size limitation at the same time. Second, gradient domain variance guided image filtering is proposed for restrain noise amplification in image enhancement, which has better denoising and edge preservation. The above two algorithms are applied to the V-channel in the hue, saturation, value space to form a color image enhancement algorithm. This method not only improves the brightness and quality of low-illumination images but also avoids noise amplification after image enhancement. The experimental results prove that our proposed algorithm is superior to other methods in improving brightness and noise removal of low-illumination images.

Faced with massive amounts of image data, the performance of classification algorithms based on traditional platforms with single-node architecture drops dramatically. We propose a classification method based on hybrid optimization and combination technology in a cluster environment that is suitable for use with large-scale scene images. Support vector machine (SVM) algorithms are optimized by the artificial bee colony and particle swarm optimization algorithms to produce weak classifiers; then, a strong classifier is constructed by combining the outputs from the 15 weak classifiers using the AdaBoost algorithm. The MapReduce parallel programming model in the Hadoop platform is used to parallelize the algorithm, and a parallel AdaBoost hybrid optimization (PAH)-SVM algorithm is proposed. Finally, a model is constructed for automatic classification of the large-scale scene images. Multiple sets of comparative experiments show that the average classification accuracy of the proposed algorithm when applied to the scene understanding (Caltech-256 and Pascal VOC 2012) database exceeds 85.0%, and its training time is <10  min when 170,000 images are used. Considering the cost of hardware, the execution time and accuracy of this algorithm are superior to those of mainstream classification algorithms, such as P-SVM and CNN. In addition, the speed of the system based on the proposed algorithm increases linearly, and the constructed Hadoop cluster shows good extensibility. The proposed algorithm is suitable for automatic classification and prediction using large-scale scene images.

In high-efficiency video coding (HEVC), motion estimation (ME) complexity is high. The ME complexity is greatly reduced by implementing a multiplication free one-bit transform (MF-1BT) in the HEVC encoder. Compared with the other one-bit transforms, the computational load of MF-1BT is less. The implementation of MF-1BT in a HEVC encoder greatly reduces the ME complexity in terms of ME time (MET) and total encoding time (ET). Experimental results show that the MF-1BT along with full search reduces the ME complexity in terms of percentage reduction in MET and ET by 89.03% and 87.53%, respectively, as compared with a conventional full search algorithm with a Bjontegaard delta-rate (BD-rate) of 0.1095% and Bjontegaard delta-peak signal-to-noise ratio (BD-PSNR) of 0.1311% in low delay-P main profile. The performance comparison of test zonal search (TZS) along with MF-1BT is evaluated by comparing it with the TZS algorithm in the fast search mode of the HEVC encoder. Hexagon and diamond search algorithm along with MF-1BT is also proposed in the fast search mode of the HEVC encoder to further reduce the MET and ET compared with the TZS algorithm of the HEVC encoder.

An active contour model (ACM) based on grayscale morphology fitting energy for fast image segmentation in the presence of intensity inhomogeneity is proposed. The core idea of grayscale morphology fitting energy is using the grayscale erosion and dilation operations to fit the image intensities on the two sides of contours. By extracting local intensity information using morphological operators, the proposed model can effectively segment images with intensity inhomogeneity, and the computational cost is low because the grayscale morphology fitting functions do not need to be updated during the process of curve evolution. Experiments on synthetic and real images have shown that the proposed model can achieve accurate segmentation. In addition, it is more robust to the choice of initial contour and has a higher segmentation efficiency compared to traditional local fitting-based ACMs.

A method is proposed to construct an improved measurement matrix—chaotic cyclic convolution measurement matrix (CCCMM). It is constructed by convoluting the fractional order Lorenz chaotic sequences and the cyclic matrix; since cyclic matrix is easy to implement on hardware, accurate reconstructed signal can be obtained by using fractional order chaos with pseudorandomness and convoluting them makes computing results smooth. Meanwhile, CCCMM is proved to have the high probability to satisfy the restricted isometry property. Then, the one-dimensional signals and two-dimensional images are simulated by the CCCMM and other methods, and the results show that the CCCMM is more superior in evaluating recovered signals with parameters and the visual effect of the restored images.

Image denoising based on a convolution neural network (CNN) can be described as the problem of learning a mapping function from a noisy image to a clean image through an end-to-end training. We propose a multiscale parallel feature extraction module (MPFE) for CNN denoising, which integrates residual learning and dense connection. The MPFE uses convolution kernels of different sizes to adaptively extract multiple features in different scales from the input image. We introduce dense connection to connect each MPFE, which can make different features interact with each other and concatenate together, so as to fully exploit the image features. The dense connection can pass the features that carry many image details, which help reduce image distortion. Meanwhile, it can also reduce gradient disappearance and improve convergence speed. The MPFE uses residual learning to resolve the gradient loss caused by high network depth while still ensuring that the network learns the details of the noisy image. Simulation experiments show that our denoising method has the ability of suppressing Gaussian noises with different noise levels, it performs superior performance over the state-of-the-art denoising methods.

Due to the high-dimensional data space generated by hyperspectral sensors together with the real-time requirements of several remote sensing applications, it is important to accelerate hyperspectral data analysis. For this purpose, we aim to improve the performance of an existing classification algorithm and reduce its execution time. The proposed algorithm is based on sparse representation and using extended multiattribute profiles as spectral–spatial features, and sparse unmixing by variable splitting and augmented Lagrangian as the optimization method. The speeding up is mainly achieved by exploiting the interdependencies among iterative calls and providing an appropriate memorization technique to reduce the extra cost by factorizing the algebraic computations. The experimental results on two HSI data sets prove that the optimized algorithm is really faster than the original one while retaining the same classification accuracy. This study shows how useful it is to adapt the implementation of the generic module in order to become more appropriate to the application and to minimize the extra costs as much as possible.

We propose a multitask convolutional neural network (CNN) for general no-reference image quality assessment (NR-IQA). We decompose the task of rating image quality into two subtasks, namely distortion identification and distortion-level estimation, and then combine the results of the two subtasks to obtain a final image quality score. Unlike conventional multitask convolutional networks, wherein only the early layers are shared and the subsequent layers are different for each subtask, our model shares almost all the layers by integrating a dictionary into the CNN. Moreover, it is trained in an end-to-end manner, and all the parameters, including the weights of the convolutional layers and the codewords of the dictionary, are simultaneously learned from the loss function. We test our method on widely used image quality databases and show that its performance is comparable with those of state-of-the-art general-purpose NR-IQA algorithms.

Texture synthesis is proposed to construct a large digital image from a small sample by taking advantage of its structural content. However, many approaches in texture synthesis are based on the presence of broken features at the overlap of adjacent patches. Due to inaccurate similarity measures, several optimization schemes for patch mergence may fail if these schemes cannot find satisfactory candidates from input samples. Self-similarity of candidate patches was proposed, and an algorithm was developed to perform distance error matching and alignment via the sum of self-similarity. First, when the matching window moves in the texture samples along the scan line, the sum of self-similarity replaces the sum-of-squared difference to decrease broken features or texels of outputs. Next, the synthesis speed was accelerated by eliminating redundancy calculation during matching window sliding. Finally, the proposed method enlarges the search range of the suture from one patch to all patches in the horizontal direction, and the broken features of overlaps are eliminated in the outputs. To demonstrate the performance of the proposed method, the obtained synthesis results were compared with other conventional synthesis methods. In all cases, experimental results demonstrate that the self-similarity matching-based method can decrease the number of feature discontinuities and accelerate synthesis time.

An image feature descriptor named “sparsely encoded distinctive visual features (SEDVF)” is proposed for object recognition. SEDVF is built with the integration of local and global visual features. Visual information on edge orientation, magnitude, color, and pixel intensity is sparsely encoded by a bit plane slicing technique. Distinctive features are obtained using winner-takes-all principle. Edge gradient multiorientation detector method (EGMOD) is proposed to obtain the gradient orientations. EGMOD extracts four-directional (horizontal, vertical, and both diagonal) edge information with the proposed multioriented Scharr operator in YIQ color space. Magnitude features are extracted by incorporating chromatic information along horizontal and vertical directions in the RGB color space. SEDVF can be used as a color feature descriptor that has good discriminative power of visual features. The proposed descriptor is extensively tested for performance evaluation using K-nearest neighbor classifier on three standard datasets, including Columbia object image library, Amsterdam library of object images, and PVOC 2007, respectively. Experimental results reveal outperformance of SEDVF as compared to the state-of-the-art object recognition methods.

In the past few years, a convolutional neural network (CNN) based deep learning model has been broadly applied in image processing and computer vision. And different from other multiscale decomposition methods in infrared and visible image fusion field, a hybrid l₀-l₁ layer decomposition model, which combines the superiority of l₀ sparsity term and l₁ sparsity term, is carried out to decompose the image into the base layer and the detail layer. Thus, a CNN model and visual saliency-based methods are utilized to fuse the detail layer and the base layer, respectively. Finally, the experiments show that this combination of CNN and saliency detection fusion rule has outperform some of the existing methods in infrared and visible image fusion both subjective and objective evaluations.

This publisher’s note corrects a publication error with the paper.