Image fusion is the process of combining partially focused images to a completely informative focused image of the same scene. Also, it addresses the loss of depth when using an optical lens. However, many state-of-the-art fusion methods cannot preserve all the significant features of the source images well to obtain an all-in-focus image. In addition, most of these methods are sensitive to noise, usually leading to fusion image distortion and image information loss. We propose a fusion method for multifocus image using two-scale decomposition and global sparse features. Two component decompositions are performed on the source images with gradient-domain-guided image filtering and robust principal component analysis, respectively. The corresponding obtained components are base layers, detail layers, principal components, and sparse components. Joint decision maps are constructed based on the salient features of the obtained layers and sparse components, which guides the fusion of base and detail layers. The fused base and detail layers are integrated to construct the final fused image. The experimental results demonstrate that the proposed method can achieve better fusion performance compared to other reported methods.

Deep convolutional neural networks (DCNN) are extensively used in image classification. However, it is typically impossible to receive ideal effects if directly using it in synthetic aperture radar (SAR) image target recognition. During the data acquisition process, it is difficult to acquire a large number of labeled training data and to ensure the image targets in the center. To extract the width and height features of SAR images and reduce the computational complexity, we constructed an asymmetric parallel convolution module. The module avoids severe over-fitting due to limited training samples and effectively deals with displacement changes with test samples. Meanwhile, the residual learning method is used in the algorithm to avoid the deep network degradation and improve algorithm recognition accuracy (RA). Experimental results show that the RA of the DCNN with residual-learning-based APCRLNet reaches 99.75% under standard operating conditions, which is superior to existing recognition methods. Furthermore, the algorithm also performs well for incomplete training samples and test samples with displacement changes.

Three-dimensional (3-D) face reconstruction is an important task in the field of computer vision. Although 3-D face reconstruction has been developing rapidly in recent years, large pose face reconstruction is still a challenge. That is because much of the information about a face in a large pose will be unknowable. In order to address this issue, we propose a 3-D face reconstruction algorithm (PIFR) based on 3-D morphable model. A model alignment formulation is developed in which the original image and a normalized frontal image are combined to define a weighted loss in a landmark fitting process, with the intuition that the original image provides more expression and pose information, whereas the normalized image provides more identity information. Our method solves the problem of face reconstruction of a single image of a traditional method in a large pose, works on arbitrary pose and expressions, and greatly improves the accuracy of reconstruction. Experiments on the challenging AFW, LFPW, and AFLW database show that our algorithm significantly improves the accuracy of 3-D face reconstruction even under extreme poses (±90 yaw angles).

Establishing cross-spectral feature point correspondences is a challenging problem due to high textural and intensity changes between multisensor images. To overcome such changes, a new feature matching strategy is proposed. The proposed strategy is based on Gaussian mixture model-universal background model (GMM-UBM). GMM-UBM is widely used as a classifier in speech-related applications. GMM-UBM is used to establish feature point correspondences between multisensor images. Experiments were performed using 17 different state-of-the-art feature points and GMM-UBM was validated on two different image datasets. The experimental results show that the proposed GMM-UBM method outperforms traditional Euclidean distance-based feature matching strategies and provides better results on multisensor images.

Automatic detection of fabric defects is an important process in the textile industry, which is required to locate and classify microdefects from a large fabric image. We propose a learning-based system for automatic detection of microfabric defects. A segmentation algorithm based on fractal and gray features is applied to extract microdefect regions. Gabor fractal network is designed to further improve identification ability of this approach. The proposed network achieves superior performance in terms of detection accuracy with a much smaller model size. The best testing accuracy rates on dark line, hole, broken yarn, and dirt are 96.9%, 98.0%, 92.9%, and 98.8%, respectively. Experimental results demonstrate the effectiveness of the proposed scheme in defect detection for microfabrics. The proposed system has great potential for automatic detection of microfabric defects.

Region proposal algorithms are beneficial for enhancing the performance of object detection and recognition methods. We present a method for grouping region proposals based on perceptual grouping principles. The grouping principles are simulated to extract image features, and the region proposals are segmented by solving a sequence of parametric maxflow problems. In order to extract complex objects from natural images, the element connectedness cue is introduced in the parametric energy functions. This newly introduced cue is propitious to group objects with diversified patterns. To effectively fuse the grouping principles, a multiclassify-based learning algorithm is proposed to optimize an ensemble of binary segmentation models. The training samples are first divided into groups to pretrain each individual model, and the algorithm adaptively adjusts the sample groups in the iteration procedure to learn an optimal set of models. We conduct the experiments on the PASCAL Visual Object Classes Challenge 2012 segmentation dataset but also in the context of region proposals in optical remote sensing images, and the results show that the proposed method can achieve a favorable performance compared to the existing algorithms.

Among the variety of multimedia formats, color images play a prominent role. A technique for lossless compression of color images is introduced. The technique is composed of first transforming a red, green, and blue image into luminance and chrominance domain (YC_uC_v). Then, the luminance channel Y is compressed with a context-based, adaptive, lossless image coding technique (CALIC). After processing the chrominance channels with a hierarchical prediction technique that was introduced earlier, Burrows–Wheeler inversion coder or JPEG 2000 is used in compression of those C_u and C_v channels. It is demonstrated that, on a wide variety of images, particularly on medical images, the technique achieves substantial compression gains over other well-known compression schemes, such as CALIC, M-CALIC, Better Portable Graphics, JPEG-LS, JPEG 2000, and the previously proposed hierarchical prediction and context-adaptive coding technique LCIC.

A color image encryption algorithm is proposed based on the position index and chaos theory. First, the plain image is decomposed into its constituent red, green, and blue channels, each of which is resized to form one-dimensional arrays. Then, a permutation mechanism based on the position index values is introduced to effectively decrease the correlation between adjacent pixels by dynamically swapping the pixels of the three components. The Tent–Sine system is employed as a pseudorandom sequence generator to provide chaotic sequences in which the initial values and parameters are calculated from the Secure Hash Algorithm (SHA)-256 hash of the plain image. As a result, the algorithm is sensitive to the plain image. In addition, in contrast to the traditional ciphertext feedback mode, each pixel is modified based on the order of the position index sequences. The results of experimental simulations and extensive cryptanalyses indicate that the algorithm satisfactorily resists various typical attacks.

In the last few years, most works on palmprint recognition systems have been focused on developing a practical system that should have high performance in term of recognition accuracy, matching speed, and storage requirement. However, they have certain shortcomings, such as long computational time and sensitiveness to translation, illumination, and rotation. To handle these limitations, we present a simple and effective scheme to produce a meaningful local palmprint representation called patch binarized statistical image features descriptor (PBSIFD) for palmprint identification. The PBSIFD representation significantly exploits the power of the BSIF texture descriptor. In addition, the reduced version of PBSIFD called RPBSIFD is also obtained using whitened linear discriminant analysis. The proposed schemes are successfully applied to four widely used palmprint databases, including PolyU2D, PolyU2D/3D, IITD, and CASIA, and they are compared with recent approaches. It is shown that they outperform the existing methods.

Dynamic linear threshold attenuation linking synaptic computing network (DLSCN) is proposed. The attenuation rate of neural threshold determines the propagation rate of the gamma band. DLSCN employs clustering to segment the external input stimuli of neurons in intensity. Different attenuation rates are set in different intensity stimulus regions to control the propagation rate of the gamma band. Therefore, the ignition timing of neurons can be delayed or advanced so as to obtain the expected distribution of linking synaptic. The intensity distribution of the output image of the network is more consistent with the characteristics of the human visual system. Experiments on four public datasets show that DLSCN achieves good results in enhancing image contrast and boosting image details.

As a component of a power transmission line, the state of an insulator impacts the reliability and safety of the power grid. Self-shattering is an important factor that may cause insulator anomalies. We present a method for detecting insulator self-shattering using mask regions with a convolutional neural network, namely a mask region convolutional neural network. The method can locate fault insulators while finding the fault image with insulator self-shattering. It can also find the insulators and distinguish between normal and self-shattering even if there are multiple insulators in an image. The insulator self-shattering detection program is written in TensorFlow and the Keras deep learning framework. Experiments are conducted on 810 real-world images. The testing results show that the mean average precision can be up to 1 for single-target images and 0.948 for multitarget images.

The background contains potential favorable information for visual tracking. The recent correlation filter-based tracking methods achieve great success by densely sampling at the target and its surrounding background. To further improve the tracking ability against deformation (DEF) and occlusion (OCC), we propose to utilize auxiliary objects. First, auxiliary object candidates with objectness are generated with unsupervised salient region detection. Second, the graph model between the target with all auxiliary object candidates is established when these objects are tracked individually. Third, all auxiliary objects vote for correcting the predicting result of the target tracker. Experimental results show that auxiliary objects are helpful for tracking the target, especially for improving the robustness on DEF and OCC.

Texture feature captures the structural arrangement details of pixels present in a given image. The local binary pattern (LBP) and its modified versions have been used to extract discriminant information about the structural arrangement. A unique textural representation obtained from LBP and its modified versions are affected by rotating the structural arrangement of the pixels available in a particular region. The weight of the binary values obtained from the LBP and its modified versions are constant. The structural arrangement of the local region is effectively determined from the direction-oriented edge information [i.e., local directional pattern (LDP)]. The LDP is also nonrobust toward the structural arrangement changes in the local region. We give a direction-invariant binary pattern to represent the texture details available in a particular location by encoding edge response along with the LBP of that region. The direction-invariant property is derived from the weight of the edge response available in a particular region. The edge response around each local region of the image is captured using the compass edge detector. It detects the edge details from eight directions. The process of encoding eight direction edge responses along with a binary pattern obtained from the center pixel difference gives a more robust texture descriptor than the other rotation-invariant descriptors used in texture feature extraction. The proposed descriptor is tested over five standard databases, namely Brodatz, color Brodatz, KTH-TIPS2b, FTVL, and free photo. The performance of the proposed rotation-invariant texture feature is more effective than the existing rotation-invariant local texture feature representations involved in the image classification and retrieval experiments.

Because of gigantic specialized headways in communication and media, there is a requirement for web correspondence in various areas such as military, medicinal, and e-administration. To take into account this need for data security, a reversible data-hiding algorithm in an encrypted domain for absolute moment block truncation coding (AMBTC)-compressed images is proposed. To secure the content of the cover image, firstly an AMBTC compressed image is encrypted through homomorphic encryption and then a binary secret message in base₄ numeral framework is embedded at embeddable pixel values of the AMBTC-compressed image without occurrence of any overflow problem. The embedding rate can be additionally improved with no disintegration of the visual nature of the cover image. Highlights of the proposed algorithm are elimination of the overflow problem during data embedding in cover image and yielding a peak signal-to-noise ratio (PSNR) value of infinite decibels. The cover image and stego image are indistinguishable from one another. Visual quality of the stego images obtained through the proposed method is superior to all the other compared methods. The experimental results reveal that, with regard to the embedding rate and PSNR value for all test images, the proposed algorithm performed superior to other ones.

This study examines the lossless compression of an original JPEG bitstream such that spare space is reserved to embed secret data in a file header. After an original image is transformed using discrete cosine transform (DCT), the nonzero DCT coefficients at high frequencies are relatively sparse and their absolute values are small. To improve encoding efficiency, we propose a segment encoding algorithm that defines a high frequency in DCT blocks as the termination point and recodes coefficients separately. The coefficients above this point are encoded in the proposed Rxy code, and the coefficients below it are encoded in JPEG code. Then, to further preserve the embedding capacity’s (EC) robustness to different quality factors, we cluster blocks into large 32  ×  32 blocks and record special block indexes that are not suitable for Rxy encoding in appended data. Hence, space is reserved for inserting secret data and appended data after the comment marker in a file header. A receiver can obtain the termination point through data analysis, extract additional data, and recover the original JPEG image in a lossless manner. Encoding redundancy is substantially reduced such that a large EC can be achieved with no file expansion. Moreover, only coefficients at high frequencies are recoded; hence, concealment is computationally simple. Experimental results demonstrate that the proposed method has a larger EC than state-of-the-art methods.

Denoising is an essential task in hyperspectral image preprocessing that can improve the performance of subsequent applications. We propose an iterative hyperspectral images denoising method that results in two algorithms, one global and one nonlocal. Both are based on a higher order singular value decomposition (HOSVD) sparse model and realize a regularization in each iteration. The proposed global algorithm treats the hyperspectral image as a whole entity that is able to jointly consider both spatial and spectral information and then uses an iterative regularization framework to mitigate the noise, while the nonlocal algorithm takes the advantages of a patch-based HOSVD sparse model and is more efficient. The experiments with both synthetic noisy and real hyperspectral images show that the proposed iterative method improves the hyperspectral image quality. The subsequent classification results further validate the effectiveness of the proposed hyperspectral image noise reduction.

Although image denoising methods based on a convolutional neural network (CNN) have achieved the state-of-the-art in Gaussian noise reduction, their performances are still very limited on real photographs even beyond the reach of traditional methods, such as block-matching and 3-D filtering and weighted nuclear norm minimization. We use a denoising benchmark to train a generative adversarial network for noise modeling and produce more data indistinguishable from the original ones, which can be regarded as a data augmentation scheme. Then, we utilize this extended dataset, including real images and synthetic images to train a CNN-based denoiser named DRNet for real photograph denoising. In the design of DRNet, we introduce a noise estimation module to improve the robustness of the single learning framework for handling unknown noise levels and a pair of reversible downsampling and upsampling operators to enlarge the receptive field. Experiments on real-world noisy images are conducted to evaluate our algorithm, and the results show that DRNet is effective for real photographs in comparison with other methods, especially in balancing the noise removal and the structure preservation.

The properties of aluminum alloy highly depend on the distribution, shape, and size of the microstructures. Thus accurate segmentation of these microstructures is crucial in the fields of material science. However, it is often challenging due to large variations in microstructural appearance and insufficiency in hand-labeled data. To address these challenges, we propose a hierarchical parameter transfer learning method for the automatic segmentation of microstructures in aluminum alloy micrograph, which can be seen as the generalization of the typical parameter transfer method. In the proposed method, we use the multilayer structure, multinetwork structure, and retraining technology. It can make full use of the advantages of different networks and transfer network parameters in the order from high transferability to low transferability. Several experiments are presented to verify the effectiveness of the proposed method. Our method achieves 98.88% segmentation accuracy and outperforms four typical segmentation methods.

We propose an invariant description method based on Zernike moments to classify hand vein patterns from raw infrared (IR) images. Orthogonal moments provide linearly independent descriptors and are invariant to affine transformations, such as translation, rotation, and scaling. A mathematical expression is given to derive a set of moment invariants. The obtained features have all the properties of moment invariants with the additional feature of image contrast invariance. For dorsal hand vein pattern acquisition, an IR imaging system is implemented. Also, a public database is used for a palm vein recognition task. A correct rate classification (CRC) above 99.9% is achieved using a set of rotation, scale, and intensity Zernike moment invariants. Additionally, multilayer perceptron and K-nearest neighbors are used as classifiers having as input data the Zernike normalized moments. A discriminative feature evaluation of the image moments allows the reduction of the number of descriptors while maintaining a high classification rate of 99%. The efficiency of the moment descriptors is evaluated in terms of accuracy and reduced computational cost by (a) avoiding the necessity of a preprocessing stage and (b) reducing the feature vector dimension. Experimental results show that Zernike moment invariants are able to achieve hand vein recognition without image preprocessing or image normalization with respect to change of size, rotation, and intensity.

Based on previous anti-cat-eye effect imaging techniques sacrificing too much imaging quality to achieve substantial retroreflection reduction, an anti-cat-eye effect imaging technique based on light-field imaging is proposed. Relevant studies have been carried out regarding the mechanism and effectiveness of both antilaser reconnaissance and blinding for this technique. By applying the Fresnel–Kirchhoff diffraction theory and defining the microlens array as a superposition of a series of microlens units, the retroreflection formation of the light-field imaging system is theoretically modeled. Based on the physical model, the influences of defocusing on the intensity distributions of spots on both the light-field detector and observation planes are further studied. The results show that, compared with a conventional system with the defocus invariant and flexible reconstruction properties, the light-field imaging system not only increases the interference and blinding thresholds by nearly one order of magnitude but also reduces both the retroreflection maximum intensity and the echo-detector receiving power by more than one order of magnitude, which sufficiently manifests the superior antilaser reconnaissance and blinding performances of the light-field imaging system.

Today, road detection is still a challenging task in intelligent driving. With the continuous improvement of computer vision, many methods of deep learning are used for road detection because they can achieve image features at a deeper level and discover road areas from raw RGB data. However, the method to detect the road areas accurately needs to be improved. We present a method that can extract the road area features and complete road detection tasks. Our method mainly includes the following points: (1) to introduce the cycle-consistent adversarial network to extract the road area features in a picture and complete image to image conversion and (2) to complete road detection by adding a new model and to improve the accuracy of detection. The results of our method are evaluated by uploading to the Karlsruhe Institute of Technology and Toyota Technological Institute road detection benchmark and named it as “road detection cycle-consistent adversarial networks.” Our method achieves an overall max F-measure of 88.63% and precision of 91.35%. In addition to high precision, our method also has a good robustness. Meanwhile, the accuracy for narrow road areas needs to be optimized in the future.

Object detection is a challenging computer vision problem with numerous practical applications. Due to low accuracy and slow detection speed in object detection, we propose a real-time object detection algorithm based on YOLOv3. First, to solve the problem that features are likely to be lost in the feature extraction process of YOLOv3, a DB-Darknet-53 feature extraction network embedded in inception structure is designed, which effectively reduces the loss of features. Second, the detection network of YOLOv3 and the reuse of deep features in multiscale detection network are improved. Finally, the numbers and sizes of anchor boxes are selected by K-means clustering analysis, and the detection model is obtained by means of multiscale training. The improved algorithm has a mean average precision of 0.835 on the PASCAL VOC data set and a detection speed of 35.8  f  /  s, which is better than YOLOv3.

Efficient maintenance of road networks is an important factor for the smooth flow of traffic and is highly dependent on the timely assessment of road conditions. Road defects such as cracks and potholes cause significant safety and economic problems. Automation of these activities is vital for efficient and cost-effective maintenance of road networks. Image analysis techniques have widely been used and proven to be effective in fields such as medical image processing, face recognition, pattern recognition, and computer vision. We propose a two-stage modular approach for road defect recognition. The first module deals with image segmentation using a variant of random walk algorithm, which employs boost C-means clustering for selection of initial seeds. The second module conducts block-based image classification using an enhanced artificial neural network, which utilizes genetic algorithm for weight optimization. Experimental results show that the proposed approach is effective for the automatic recognition of pavement defects. The accuracy is not <73  %  , even when multiple defect categories are considered.

A method for geometric distortion correction and space and time-varying blur reduction is proposed, which can recover the high-quality image from a single-frame image distorted by atmospheric turbulence. First, the U-net like deep-stacked autoencoder neural network model is proposed, which is composed of two deep convolutional autoencoder (CAE) neural networks and a U-net. The first CAE is used for feature extraction, the U-net is used for feature deconvolution, and the second CAE is used for image reconstruction. For the loss reduction of image information, transposed convolution instead of upsampling is selected in U-net networks. Moreover, in order to obtain sufficient feature information for reconstruction, the first CAE and the last CAE are symmetric skip connected. This not only enables the fusion of low-level and high-level information but also ensures the integrity of image information greatly. Then, a method of gradually mature training from simple to complex is proposed to overcome the difficulty of convergence on smaller training sets. It makes the network be converged and mature by increasing the complexity of training data gradually so as to restore the high turbulence-degraded image. Experimental results of actual observation data and simulation data show that the algorithm has a stronger antinoise ability and can recover image details and sharpen image edges more effectively. In particular, for atmospheric turbulence severely degraded image restoration, the peak signal-to-noise ratio index is increased by about 10% on average compared with state-of-the-art methods.

In stereo matching, energy-minimization optimizers such as dynamic programming (DP) could achieve efficient disparity estimation. However, DP-based optimizers still suffer from relatively high energy due to too many artificial penalty parameters and a complex energy function, inevitably leading to matching failure in depth discontinuities, occlusions, and low-textured regions. To reach dense high-accuracy disparity estimation, we propose an effective stereo-matching method using texture discrimination-enforced matching cost computation and smoothness-weighted cost regularization. First, pixel adjustment is introduced to the gradient difference calculation so that initial matching accuracy, especially in low-textured regions, can be improved. For determining the degree of pixel adjustment, a gray-gradient co-occurrence matrix and the fuzzy c-means method are adopted to differentiate between two varieties of reference images based on textural property. Second, the support weight in local methods and the smoothness constraints in global optimizations are leveraged. Color information and the confidence map of the input image are combined to form the support weight. By ameliorating the smoothness constraint function during weighted horizontal DP passes, the adverse effects of excessive penalty can be reduced. Third, a postprocessing method is proposed for refining results in both occluded and low-textured regions. The proposed method considers test cases from the Middlebury v.2 and v.3 datasets and can outperform the state-of-the-art stereo-matching algorithms.

In deep exploration, we often use high-noise and high-resolution astronomical images because of long distance transmission, cosmic noise interference, or other disturbances. Designing an effective method for high-noise astronomical image denoising plays a crucial role in this field. The famous compressed sensing (CS) has shown to be a successful technique for high-resolution image reconstruction. CS is employed to solve the denoising problem in high-noise astronomical images, and a CS reconstruction model and a CS curvelet soft thresholding (NCCST) reconstruction algorithm are proposed simultaneously. In image denoising, the l₁-norm reconstruction model based on CS cannot fully consider the internal structure of the image, thus affecting the reconstruction quality. Combined with the l₁-norm reconstruction model and group sparse total variation method, the CS reconstruction model is first established. In this framework, the NCCST algorithm uses an adaptive curvelet adjustment operator proposed to select the sparse coefficients in the stage of image sparsity transform, and then a descending VisuShrink threshold is adopted to pick out these reconstructed astronomical image coefficients in each iteration. The experimental results demonstrate that the algorithm proposed can reconstruct clear images from high-noise and high-resolution astronomical images. Meanwhile, it can preserve more astronomical image details, edges, and textures. Even at a low-compression sampling ratio, this algorithm can achieve superior reconstruction performance. The proposed algorithm can be applied in the satellite receiving station to quickly and accurately reconstruct high-resolution astronomical images with high quality.

With the recent advances in high computing capabilities and reliable delivery networks, it is realistic to expect emerging immersive media applications such as virtual reality (VR) and 360-deg video. ISO/IEC Motion Picture Experts Group (MPEG) has launched an MPEG-I project that focuses on not just the current but also the next-generation immersive media applications. Light field (LF) is one of the candidates under consideration by MPEG-I. A LF lenslet image contains plentiful repeating patterns since each microlens captures the same scene. On the other hand, intrablock copy (IBC) in screen content coding (SCC) increases the coding efficiency of screen content (SC) by finding the repeating patterns. SCC can be an efficient encoder for LF image compression. However, SCC is not optimized for compressing lenslet images. Therefore, we propose an efficient lenslet image coding (LIC) model using SCC to encode lenslet images based on the lenslet image characteristics. Experimental results show that our LIC model further improves the coding efficiency of SCC on lenslet images by 3.72% on average and up to 7.77% in bitrate reduction. For fast approach, LIC is faster than SCC by 44.12%. With our proposed LIC model, the encoded bitstreams are compliant with the SCC standard with an even faster decoding time.

Aerial scene classification, which aims to automatically label an aerial image with a specific semantic category, is one of the most challenging tasks in computer vision. With the improvement of spatial resolution of aerial images, the texture details of ground objects become increasingly abundant. Although many classification algorithms have been proposed in recent years, accuracy and completeness of the classification results are still imperative problems to be solved. We propose an algorithm via a two-stage voting fusion strategy. First, the superpixels obtained by simple linear iterative cluster algorithm are treated as the input of a random forest classifier. Then, the region growth strategy is applied to complete the clustering of similar local regions, and the learning of mosaic zero pieces is simultaneously realized. Finally, a Bayesian framework is exploited to fuse the classification results in order to get the final classification results. Experiments on two benchmarks demonstrate the superiority of the proposed method (TSS) compared with other state-of-the-art methods, whether in terms of the integrity of big targets or the accuracy of small objects.

Deep learning-based steganalysis techniques have significantly progressed in recent years. Compared with typical steganalysis methods using rich model features and classifiers, deep learning-based steganalysis methods usually train a deep convolutional neural network (CNN) as the steganography detector. One advantage of a deep CNN is the strong capability for discriminative feature learning. The input feature map of the fully connected layer is the steganalysis feature learned using the deep CNN. We can, therefore, extract the steganalysis feature based on the trained deep CNN. Compared with steganalysis features constructed by hand, the extraction of steganalysis features using a deep CNN can take advantage of the strong feature-learning capability of such a network. Therefore, two types of typical steganalysis frameworks are first compared. Then, a deep CNN for steganalysis feature learning is constructed, and the setting of the image preprocessing layer is discussed. Next, the detection performances of the different learned steganalysis features are compared. Finally, the detailed extraction process of the proposed learned steganalysis feature is described. The experiment results show that the proposed steganalysis method, combining learned and handcrafted steganalysis features, can significantly improve the detection performance for content-adaptive JPEG steganography.

Joint distribution matching based on generative adversarial nets (GANs) is an effective method to alleviate the insufficient diversity of labeled samples in semisupervised learning. In fact, in addition to the existing samples, generated samples and corresponding predictive labels must be taken into account to further increase the diversity of labeled samples and the controllability of generated samples. However, current works have not considered this. Therefore, a semisupervised learning model with adversarial training among joint distributions is proposed. The model consists of a generator, a classifier, and three discriminators incorporated with four joint distributions of samples and labels. The theoretical research indicates that, when the model reaches equilibrium, the classifier happens to be the inference network of the generator. Hence, the controllability of the generator and the generalization ability of the classifier are mutually improved. In semisupervised classification experiments, our model achieved state-of-the-art error rates of 0.59%, 16.45%, and 4.86% on MNIST, CIFAR10, and SVHN datasets, respectively. When only 20 labels are available on MNIST dataset, the error rate dropped from the current best of 4% to 1.09%, which indicates that the model is extremely robust to the number of labels. Meanwhile, the model also shows competitiveness in semisupervised generation.

Tracking methods based on correlation filter have been showing great advantages and huge potential. However, it will lead to tracking drift when the target is disturbed, such as occlusion, illumination variation, deformation, and others. In the pursuit of tracking accuracy and robustness, a feature integration model and an adaptive update method for robust tracking are proposed. First, the fine-tuned convolutional neural networks medium is employed to obtain the deep features. The deep features are integrated with traditional manual features by combining with the independent component analysis with reference method to obtain more discriminative features. Second, in the model update phase, an adaptive updating strategy based on the center shift Euclidean distance of image patch is proposed to reduce the unnecessary calculation in model training so as to speed up the tracking speed. Finally, the tracker proposed is evaluated on online tracking benchmark (OTB-2015) and visual object tracking benchmark (VOT-2016). The experimental results show that the method that integrates deep features with traditional manual features can distinguish background and target better. Thus the proposed VOT algorithm performs better robustness and accuracy against state-of-the-art methods when dealing with challenging environments.

Sparse representation-based classification (SRC) and collaborative representation-based classification (CRC) have garnered significant attention recently. In CRC, it is argued that it is the collaborative representation mechanism but not the ℓ₁-norm sparsity that makes SRC successful for classification tasks. However, recent studies reveal that sparsity does play a critical role in accurate classification, thus it should not be totally overlooked due to relatively high computational cost. Inspired by these findings, we propose a method called sparsity augmented weighted collaborative representation-based classification (SA-WCRC) for image classification. First, the representation coefficients of the test sample are obtained via weighted collaborative representation and sparse representation, respectively. Second, we augment the coefficient obtained by weighted collaborative representation with the sparse representation. Finally, the test sample is classified based on the augmented coefficient and the label matrix of the training samples. Both the augmented coefficient and classification scheme make SA-WCRC efficient for classification. Experiments on three face databases and one scene dataset demonstrate the superiority of SA-WCRC over its counterparts.

Due to the absorption and scattering of light caused by particles suspended in water and the variable attenuation of light with different wavelengths, underwater images are typically characterized by low contrast, blurred features, and severe color bias. To improve the quality of underwater images, an underwater image restoration method is proposed. First, based on similarities between the underwater optical imaging model and the atmosphere model, an improved dark channel prior algorithm is used to enhance the original underwater image and an improved gray world algorithm is used to calibrate the image. Then, an improved wavelet fusion algorithm is adopted to combine the results produced by the two algorithms. The experimental results show that the proposed method can effectively improve feature clarity and color accuracy in underwater images with low quality and illumination, resulting in a better visual effect.

The development of deep convolutional neural networks (CNNs) has facilitated a type of Siamese network-based tracking methods. Such kind of trackers generally include two basic modules, i.e., the appearance model for target representation and the discriminative classifier to determine the location. In the existing methods, much attention has been paid on feature fusion for more robust appearance representation. However, they ignore the intrinsic relationship between features from multiple layers. Furthermore, the problem of data imbalance between positive and negative samples in offline training limits the discriminative ability of the CNN-based classifier. We investigate the joint learning of representation and discriminative classifier under the Siamese network for robust tracking. We integrate the top-down modulation for feature fusion by considering the intrinsic relationship between features from different layers. To solve the data imbalance problem, we propose an advanced hinge loss objective function to mine the hard examples in offline training, which helps to improve the robustness of the similarity measure. Experimental results on the public tracking benchmark datasets show that the proposed method obtains favorable tracking accuracy against the state-of-the-art trackers with a real-time tracking speed.

The detection of pavement cracks is essential for damage assessment and maintenance of pavement. Obtaining complete crack paths using traditional approaches is difficult due to the varied appearance of pavement cracks and complex texture noise. A robust graph network refining algorithm guided by multiscale curvilinear structure filtering (CFGNR) is proposed for pavement crack detection. A multiscale curvilinear structure filter consisting of curved linear templates and a local texture inhibition term is first utilized to enhance crack contours. The enhanced pavement image is then presented as a graph of overcomplete crack paths, and a graph network refining approach derived from path saliency and local contrast constraints is utilized to select the optimal subset of crack paths. Finally, an iterative path growing algorithm is employed to obtain pixel-level cracks. Experimental results on four public pavement datasets show that the proposed algorithm significantly improves the completeness of detected cracks and achieves a superior performance compared to six existing algorithms.

This erratum corrects an error in a figure caption.

This erratum corrects a misspelled author name.