A practical face recognition system demands not only high recognition performance but also the capability of detecting spoofing attacks. While emerging approaches to face antispoofing have been proposed in recent years, most of them perform poorly on unseen samples. The generalizability of face antispoofing needs to be significantly improved before it can be adopted by practical application systems. The main reason for the poor generalization of current approaches is the variety of materials among the spoofing devices. As the attacks are produced by putting a spoofing display (e.g., paper, electronic screen, forged mask) in front of a camera, the variety of spoofing materials makes the spoofing attacks quite different. Another reason for the poor generalizability is that limited labeled data are available for training for face antispoofing. We focus on improving the generalizability of convolutional neural network (CNN)-based face antispoofing methods across different kinds of datasets. We propose a deep domain transfer CNN using sparsely unlabeled data from the target domain to learn features that are invariant across domains for face antispoofing. Experiments on five face spoofing datasets show that the proposed method significantly improves the cross-test performance only with a small number of unlabeled samples from the target domain.

Traditional frame interpolation algorithms typically find dense correspondences to synthesize an in-between frame. Finding correspondences is often sensitive to occlusion, disocclusion, and changes in color or luminance. We present a phase-feature-aided multiframe interpolation network that aims to estimate multiple in-between frames in one pass and handle challenging scenarios such as extreme light changes and occlusion. We first model the relation between multiple in-between frames together to enhance the temporal consistency. Two candidate optical flow fields are produced for a given in-between frame, one predicted from our network and the other estimated from those of neighboring frames using a flow fusion map. We also employ an image fusion map to combat occlusion problems in the warping processes, producing two candidate interpolated images that are fed to a shallow network with a residual structure to obtain the final interpolated image. To handle challenging scenarios, we apply a set of Gabor filters to extract phase variations in the feature domain with a multiscale phase subnetwork. Our entire neural network is end-to-end trainable. Our experiments show that this method outperforms the state-of-the-art approaches and achieves marked visual improvement in challenging scenarios.

Most of the existing local binary pattern methods discard local color differences by holding their binary information. And local spatial information is also neglected. To address these problems, a robust color image descriptor local color directional quaternionic pattern (LCDQP) is proposed. In the descriptor, the color distance map (CDM) is generated in the RGB color space to capture the color distribution among three channels. According to the distribution of CDM elements and their neighbors, four edge models are defined to describe the change trend of the original image. Then, based on this, the LCDQP strings are gained according to the distribution of four models in the four directions of 0 deg, 45 deg, 90 deg, and 135 deg. Finally, an effective quaternionic code method is adopted to construct the LCDQP descriptor. The proposed descriptor not only captures the local color features but also reflects the spatial structure information. Experiments on four representative databases demonstrate that the proposed descriptor is superior to other state-of-the-art approaches.

Surveillance of perimeter of critical infrastructures (CIs) has gained a particular attention worldwide due to rising threats of terrorist incidents. However, the generation and dissemination of real-world, challenging datasets concerning CIs for performance assessment of surveillance tasks, particularly tracking, have widely been overlooked in the community. Recent performance evaluation of tracking and surveillance (PETS) workshops have sought to address this issue by introducing ample CI surveillance datasets. We demonstrate the effectiveness of these publicly released real-world PETS CI visual datasets by providing a comprehensive statistically significant performance analysis as well as formative assessment of several state-of-the-art multitarget trackers using well-known and recent performance assessment criteria.

Most existing feature learning methods optimize inflexible handcrafted features and the affinity matrix is constructed by shallow linear embedding methods. Different from these conventional methods, we pretrain a generative neural network by stacking convolutional autoencoders to learn the latent data representation and then construct an affinity graph with them as a prior. Based on the pretrained model and the constructed graph, we add a self-expressive layer to complete the generative model and then fine-tune it with a new loss function, including the reconstruction loss and a deliberately defined locality-preserving loss. The locality-preserving loss designed by the constructed affinity graph serves as prior to preserve the local structure during the fine-tuning stage, which in turn improves the quality of feature representation effectively. Furthermore, the self-expressive layer between the encoder and the decoder is based on the assumption that each latent feature is a linear combination of other latent features, so the weighted combination coefficients of the self-expressive layer are used to construct a new refined affinity graph for representing the data structure. We conduct experiments on four datasets to demonstrate the superiority of the representation ability of our proposed model over the state-of-the-art methods.

To obtain wider field of view through optical microscopes, we proposed an image mosaicing method based on the convolutional neural network (CNN). The CNN was designed to discriminate the corresponding patches in a pair of images. To train the CNN, a training set that contained image patches and their matching labels was created from an open-access database. The pretrained CNN was then fine-tuned by self-learning from the target image and its transformed images. With the proposed self-learning CNN, the corresponding feature points were detected for the registration in mosaicing. The proposed method was compared with the scale-invariant feature transform detector-based and speed-up robust feature detector-based methods for mosaicing of 30 pairs of microscopic images. The proposed method outperformed the two traditional methods in terms of both visual quality and objective assessment. Results demonstrate that using the self-learning CNN can improve the accuracy of registration in image mosaicing.

Visual applications depend on image quality for algorithmic decision-making, and atmospheric conditions such as smoke and haze produce a challenge to artificial systems that rely on identification of people, objects, and obstacles. Smoke is particularly difficult because of its nonhomogeneous characteristics and irregular image coverage. Nighttime images worsen the problem because of low light and artificial light conditions. Our aim was to develop an iterative process that removes smoke from images in daytime and nighttime scenes. The haze image model was used as our baseline model. First, we developed a detection method to find the smoky regions on the image and used the dark channel haze removal process to estimate the transmission map for each color channel. We ran the algorithm iteratively because a one-time process left residual smoke. Blue smoke produced an unbalanced particle density, so the blue color channel had to be corrected more times than red and green channels. Finally, we optimized the image in postprocessing, and the results produced smoke-free images. We believe our algorithm is the first to successfully remove nighttime smoke.

In medical institutions, pain is one of the important clues for patients to transmit their conditions effectively, which makes the estimation of pain status an exceedingly important task. Of late, many methods have been proposed to address this task. However, most of them estimate the pain from entire face images or videos instead of paying more attention to the regions most relevant to pain. We propose a pain-awareness multistream convolutional neural network (CNN) for pain estimation. Specifically, we separate the regions most relevant to the pain expression, and the multistream CNN is used to learn the corresponding pain-awareness features. These features are combined into pain features with adaptive weights to estimate the intensity of pain. Extensive experiments on the publicly available pain database indicate that our multistream CNN-based method has achieved inspiring results compared to the state-of-the-art technologies.

We propose a real-time change detection system to be used as a vehicle-mounted early-warning system for indicators of improvised explosive devices. Within the context of military route clearance, the system automatically detects suspicious changes in the environment with respect to a previous patrol. For this purpose, historical images of the live scene are retrieved from a database and registered to the live image through 2.5-D view synthesis, using the three-dimensional (3-D) scene geometry acquired from a stereo camera. Changes are then found using local-area statistics in the CIE-Lab color space. A set of spatiotemporal filters is used to reject irrelevant alarms, resulting in a limited set of confident changes to be presented to the operator through an interactive graphical user interface. Next to the algorithmic contributions, we elaborate on the real-time design, featuring graphical processing units for the most time-consuming processing tasks, a pipelined architecture to increase the system throughput, and we split the system into a live and offline processing chain. This way, real-time change detection at 3.5 fps is achieved on images of 1920  ×  1440  pixels. Finally, an extensive system validation featuring realistic experiments shows promising detection capabilities and robustness to, e.g., lateral displacements of up to 6 m.

We address two challenges of applying compressed sensing in a practical application, namely, its poor reconstruction quality and its high computational complexity. Since most signals are not fully sparse in practice, the reconstructed signals from conventional reconstruction methods often suffer from reconstruction artifacts due to the distortion of small coefficients. To improve the reconstruction quality, we introduce referenced compressed sensing (RefCS), a reconstruction method that exploits the spatial and/or temporal redundancy between a pair of signals. We show that using a correlated reference—an arbitrary signal close to the compressed signal—there exists the bound of reconstruction error that depends on the distance between the reference and the signal. By exploiting the correlated reference, RefCS can improve the reconstruction quality by up to 90% in terms of peak signal-to-noise ratio. Moreover, it is possible to reduce the computational complexity of the proposed RefCS using the least squares method. The least squares reconstruction results can be obtained with quality comparable to that of iterative algorithms by employing the correlated reference. Using the least squares method improves the reconstruction time by a factor in the range of 9 to 5.4  ×  10⁴ according to our experiments.

According to the problems of the traditional image encryption algorithm, such as cumbersome encryption process, small keyspace, and insufficient security, a color image encryption algorithm based on a four-dimensional hyperchaotic system and using “conversion-scrambling-diffusion” mode is used. The algorithm utilizes improved Chen’s hyperchaotic system. First, according to the color image correlation attribute, the gray code iteration number was calculated, and the gray code inversely transformed the pixel value. Subsequently, the chaotic sequence generated by the hyperchaotic system was used to convert the gray code converted pixel matrix. The global pixel position was scrambled, and then, the scrambled matrix was bit-operated to complete image diffusion. Finally, the ciphertext was obtained through matrix transformation. The algorithm can realize a picture and a key. Good image encryption algorithms should meet the requirements of large keyspace, key sensitivity, effective resistance to violent attacks, entropy attacks, plaintext attacks, and other attacks as well as real-time requirements. The simulation experiments calculate and analyze evaluation indexes, such as key sensitivity, histogram, image information entropy, including one-dimensional entropy and two-dimensional entropy, and adjacent element correlation. The information entropies are more than 7.99, and the keyspace is larger than 10⁸⁶. The results prove that the encryption algorithm has robust security and antiattack ability.

We propose a no-reference (NR) method for estimating the scores of metrics assessing the quality of infrared (IR) sequences compressed with H.264 for low-complexity unmanned aerial vehicle (UAV) applications. The scenario studied is to estimate the quality on an on-ground computer to avoid performing the processing onboard, due to the computational and memory limitations of the onboard hardware. For low complexity and fast feedback, a bitstream-based (BB) approach was chosen. The original IR sequences are captured by UAV, and then BB and pixel-based (PB) features are computed. Thereafter, a feature selection process is applied and the selected features are mapped using support vector regression, to predict the quality scores of full reference metrics. The method is evaluated for the NR prediction of four image and one video quality metrics. A set of five UAV- and three ground-IR sequences are used for evaluation. The proposed NR method consistently achieves robust results for the different objective metrics tested (Spearman rank-order correlation coefficients ranging from 0.91 to 0.99). A comparison with estimations based on features from three NR models from the literature proves to be in favor of the proposed method.

Automatic facial beauty scoring in images is an emerging research topic in face-based biometrics. All existing methods adopt fully supervised schemes. We introduce the use of semisupervised learning schemes for solving the problem of face beauty scoring. The paper has two main contributions. First, instead of using fully supervised techniques, we show that graph-based score propagation methods can enrich model learning without the need of additional labeled face images. Second, we propose a nonlinear flexible manifold embedding for solving the score propagation. This model can be used for transductive and inductive settings. The proposed semisupervised schemes were tested on three recent public datasets for face beauty analysis: SCUT-FBP, M²B, and SCUT-FBP5500. These experiments, as well as many comparisons with supervised schemes, show that the nonlinear semisupervised scheme compares favorably with many supervised schemes. They also show that its performances in terms of error prediction and Pearson correlation are better than those reported for the used datasets.

Underwater imaging and image processing play important roles in oceanic scientific research. However, because the light is absorbed and scattered, the obtained underwater images are seriously degraded. Color distortion, low contrast, and detail (edge information) loss are the major problems of underwater images. We propose a method to solve these problems. First, a local adaptive proportion fusion algorithm is proposed to produce a color-balanced image, which is the first input image. Second, an edge-enhanced image is produced as the second input image. Third, a proportion fusion image is produced as the third input image. Finally, the image formation model-based local triple fusion method is used to merge these three input images and get the final result. Experimental results show that the proposed method could effectively balance color distortion and enhance edges of the degraded images. Subjective and objective evaluations show that our method is superior to many state-of-the-art methods.

We present white balancing, quad-tree decomposition, and L₁ minimization-based dehazing technique. Atmospheric light is estimated through quad-tree decomposition while holistic transmission map is estimated through coarse-scale net. Coarse and fine layers are refined through fusion and a luminance model is used to preserve details and edges. The final dehazed image is refined using L₁ minimization. Simulation results when equated with existing techniques, verify the effectiveness of the proposed technique.

Image dehazing is one of the most important image processing methods and it is widely used in daily application. A large number of dehazing algorithms have been proposed in recent years. In order to solve the problem that the dehazing performances of the classic dark channel prior methods for hazy images with the sun are bad, we propose a dehazing method for removing mist interference in images with the sun in the sky. An atmospheric scattering model is proposed to estimate the scattered light of the sun in the hazy circumstance. We also propose a gradient-based transmission map estimation method to estimate the refined transmission map accurately while reducing the computational complexity. The effectiveness of our method is confirmed by extensive experiments over a wide variety of images.

The variational method, which is a popular approach for image denoising, aims to estimate the original image from a noisy or corrupted image. To consider the constraints of image pixel values fully, our study investigates a constrained second-order total generalized variational (TGV) model, which includes non-negative and bounded constraints as a special case. By adopting an equivalent definition of the second-order TGV, we transform the proposed constrained minimization problem into a minimization of the sum of two convex functions, where one is composed of a linear transformation. Subsequently, we employ the relaxed primal-dual proximity algorithm to solve it. The advantage of the obtained algorithm is that it is matrix-inversion free and does not involve any subproblem. Numerical results demonstrate that the performance of the constrained TGV model is slightly better than that of the unconstrained model.

Pose-based action recognition has aroused increasing attention for its broad application prospects and excellent performance. Though the pose-based action recognition methods have been significantly advanced, pose-based action recognition remains a challenging task for various human action categories and subtle changes in human poses. To solve those problems, we propose pose-based multisource networks. First, human pose features are extracted from the raw video, followed by a filtration. Then, using a convolutional neural network (CNN) and long short-term memory (LSTM), the extracted pose sequence is fed into the proposed multisource networks. Subsequently, the CNN-based spatial model processes the relative position in each frame, and the LSTM-based temporal model is built to learn the temporal correlation of pose sequence. Afterward, the temporal model contains three sublevels to fully exploit the subtle information in the temporal domain. Finally, the experimental results verify the effectiveness of the proposed approach on SUB-JHMDB, MPII Cooking Activities, SYSU 3D Human-Object Interaction, and NTU RGB+D.

We present a segmentation process for detecting localized corrosion on rust-removed metallic surface based on a deep learning algorithm. In the proposed process, first the corrosion images are enhanced by a preprocessing technique, then a patch extraction process is used to collect the image set to train the deep learning model, which is used to segment the target image later. Finally, the segmentation result is improved by a postprocessing method. The performance of the segmentation process is verified with classification indicators and the receiver operating characteristic curve. The results show that the proposed method is effective in identifying localized corrosion from the metallic background, which can provide an accurate quantitative analysis tool to study the corrosion behavior.

Poisson image deconvolution remains an ill-posed research problem consisting of a nonquadratic data-fidelity term and an implicit regularization function. Recently, the plug-and-play (PnP) framework has provided a new method to reformulate the regularizer model to incorporate efficient denoisers. We develop a PnP inertial forward–backward approach (PnP_IFB) for Poisson noise reconstruction. The key advantages over the conventional alternating direction method of multipliers (ADMM)-based scheme are circumventing the matrix inversion operation and requiring less parameter tuning. The scheme can achieve a numerical convergence rate that is comparable to other acceleration strategies such as the fast iterative shrinkage-thresholding algorithm, but it does not depress the reconstruction quality. Moreover, this characteristic acceleration remains efficient when dealing with some nonconvex regularizers. Then, we demonstrate its effectiveness against other state-of-the-art approaches with respect to their deconvolution performance using synthetic images, and the experimental results prove the superiority of the method when using appropriate denoisers.

Binarization is the starting step of document analysis and recognition systems. A binarization method is proposed for a degraded historical document image. The binarization methodology is based on the joint use of nonsubsampled contourlet transform (NSCT) for enhancement and k-means clustering for binarization. The input degraded image is decomposed by NSCT for generating coefficients, which are handled through a weighting scheme for highlighting significant features. The resulting reconstructed enhanced image is then binarized by mapping pixels into foreground (text) or background (no text) using k-means clustering. Experiments are conducted on document image binarization competition datasets using blind and unblind evaluation protocol. Unblind evaluation is performed on four specific types of degradations, which are stain, ink bleed-through, nonuniform background, and ink intensity variation. The obtained results show the effectiveness of the proposed scheme in terms of objective and subjective evaluations as well as stability with respect to the other well-known methods.

The quantization of neural networks using low-precision numeric is an efficient approach to reduce the requirements for storage and computation resources, which makes it a reality to run the neural network on the resource-constrained devices. However, the accuracy of the neural network deteriorates after the normal quantization, especially for the object detection task. To overcome this dilemma, we propose a teacher–student based quantization training scheme for the one-stage object detection neural network. First, the quantization configurations of weights and activations are determined according to their statistical probability densities of each layer, thus ensuring a lower representation error of the fixed-point numeric. Second, the supplementary supervision information from a high-performance floating-point teacher detection neural network is used to assist the training of the quantized student detection neural network. Therefore, the expression and fitting ability of the quantized student can be significantly improved by selectively imitating the teacher’s key features that are indicated in the feature selection matrix. The superiority of our method has been verified on several benchmark datasets in comparison with other related methods. On the PASCAL VOC2007 test set, the mean average precision (mAP) of the 4-bit quantized tiny-YOLOv2 trained with our method can reach 61.1%, compared with an mAP of 57.4% of its 32-bit floating-point counterpart trained with the plain method. Further, the 3-bit quantized detection neural networks can still achieve satisfactory mAP even without the pretrained weights, demonstrating the robustness of the proposed method.

Regular inspection of wind turbine blades (WTBs), especially the detection of tiny defects, is necessary to maintain safe operation of wind turbine systems. However, current detections are inefficient and subjective because they are conducted merely by human inspectors. An autonomous visual inspection system is proposed in this paper for WTBs, in which a deep learning framework is developed by combining the convolutional neural network (CNN) and the you only look once (YOLO) model. To achieve practically acceptable detection accuracy for small-sized defects on the WTBs, a YOLO-based small object detection approach (YSODA) using a multiscale feature pyramid is proposed by amalgamating features of more layers. To evaluate the proposed YSODA, a database including 23,807 images labeled for three types of defect—crack, oil pollution, and sand inclusion, is developed. Then, the YSODA is with its architecture modified, and is trained, validated, and tested using the images from the database to provide autonomous and accurate visual inspection. After training and testing, resulting detection accuracy reaches 92.7%, 90.7%, and 90.3% for the three types of defect with the average accuracy being 91.3%. The robustness of the trained YSODA is demonstrated and verified in detecting small-sized defects. It is also compared with that of the traditional CNN-based and machine learning methods by applying to a real WTB system, which proved that the proposed YSODA is superior to existing approaches in terms of detection accuracy and reliability.

In recent years, face recognition has rapidly developed in the field of smartphones and control systems. It has been used to unlock the telephone and face-payment applications. With such rapid development, more and more demands are placed on the security of face recognition. However, face biometric-based recognition technologies are still vulnerable to spoofing attacks. Thus, developing robust and reliable antispoofing attack detection is critical to guarantee the security of facial analysis-based authentication. As deep learning techniques have achieved satisfactory performances in computer vision, they can also be employed to face spoofing detection. We present a multichannel linear local binary pattern optimization algorithm, which combines with Lucas–Kanade optical flow algorithm. The extracted facial features are fused and sent to a deep belief network classifier for classification and learning, and finally tested on the MSU-mobile facial spoofing database. Compared with existing deep leaning-based detection methods, our face spoofing detection algorithm has better scalability and robustness. To evaluate the performance of the proposed algorithm, the experiments are conducted on three crossed standard spoofing databases and excellent performance is also achieved.

The goals of image steganography are to simultaneously reach maximum embedding capacity, enhanced embedding efficiency, better image quality, and more security. We present a two-step data hiding method that improves image quality, capacity, and embedding efficiency, and also provides stego security. In the first step of the proposed method, the hidden data are encrypted with the key words and shifting method to provide security. In the second stage, we propose adaptive least-significant-bit (LSB)+3 type I and adaptive LSB+3 type II methods to hide encrypted data in the cover image. The image quality of the stego image obtained from the proposed adaptive LSB+3 method is better than the traditional three-bit LSB methods. When maximum data are embedded in the cover image, a peak signal-to-noise ratio (PSNR) greater than 41 dB is reached. Experimental studies show that adaptive LSB+3 type I and adaptive LSB+3 type II methods are higher PSNR values (3.48% and 5.73%) than the standard three-bit LSB substitution methods.

With recent advances in pattern recognition and computer vision, mobile palmprint authentication has become an emerging field to provide better facilities and ubiquitous computing for scientific and commercial communities. To effectively streamline this issue, researchers focus on improving authentication performance by designing deep convolutional neural networks. Despite the high potential of the state-of-the-art methods, the challenges of preprocessing computation cost, lack of training samples for big data application, and discriminative feature optimization remain to be carefully addressed. A deep mobile palmprint verification framework focusing on discriminative feature representation is proposed. To this end, an automatic feature mapping is learned from two well-known deep architectures via an effective weighted loss function. Thereafter, a convolution-based feature fusion block is followed by a surrogate model in the feature-matching phase for palmprint verification. From a practical point of view, our framework is cost-effective and can represent discriminative features with high performance. We demonstrate the effectiveness of our framework and mobile database for palmprint verification task beating the state-of-the-art on standard benchmarks. Moreover, experimental results show that our model outperforms previous ones, especially for the few-shot learning application, achieving equal error rates of 0.0281% and 0.0197% for IIT Delhi Touchless Palmprint Database and Hong Kong PolyU Palmprint databases, respectively. It is notable that all codes are open-source and may be accessed online.

We propose a hyperchaotic image encryption algorithm based on plaintext information and DNA computing. Compared to the traditional permutation-diffusion structure, the proposed method enhances the relationship between pixel position, gray value, and the plaintext image. First, the algorithm’s hyperchaotic system iteratively generates multiple pseudorandom matrices and globally scrambles the image position by operating on the pseudorandom matrices and plaintext information. Then, the pseudorandom matrices and the scrambled image are encoded into DNA sequences for subsequent algebraic operations, and a new vector is constructed to modulate the sequence to obtain robust encryption performance. Finally, the processed DNA sequence is decoded and a XOR operation is applied to the binary pseudorandom sequence to obtain a final cipher image. Based on the experimental results and security analysis, the proposed algorithm not only demonstrates excellent encryption but also resists various typical attacks.

As an effective feature extraction method, locality sensitive discriminant analysis (LSDA) utilizes the neighbor relationship of data to characterize the manifold structure of data and uses label information of data to adapt to classification tasks. However, the performance of LSDA is affected by outliers and the destruction of local structure. Aiming at solving the limitations of LSDA, a locality sensitive discriminant projection (LSDP) algorithm is proposed. LSDP minimizes the distance of intraclass neighbor samples to maintain local structure and minimizes the intraclass non-neighbor samples to increase the compactness of intraclass samples after projection. The problem of outliers is alleviated by increasing the compactness of intraclass samples in subspace. At the same time, we redefine the weights of interclass neighbor samples to maintain the neighbor relationship of different labels samples. Holding the local structure of interclass samples maintains the manifold structure of data. Experiments on face datasets demonstrate the effectiveness of the LSDP algorithm.

Appearance attributes of a three-dimensional (3-D) object, including reflection and color, are modeled as a spatially varying bidirectional reflectance distribution function (SVBRDF). However, a considerable number of images of the object are required to accurately define an SVBRDF when it exhibits complicated geometric components, thereby leading to a considerable amount of time to acquire, store, and process them. We propose a method to considerably decrease the number of images for SVBRDF modeling while maintaining the quality of the model that is achieved with several images. We introduce an acquisition score and its corresponding objective function by considering the local geometry and reflectance properties to estimate the optimal view and illumination directions for the image acquisition in which an object is measured with a reduced number of images. The captured images are subsequently processed to obtain an SVBRDF of the object. The experimental results demonstrate that the proposed method measures the appearance of a 3-D object with complicated shapes by using a few hundred images and efficiently produces an SVBRDF of the object with complicated shapes. The results indicate that the proposed method photorealistically renders an object with fewer artifacts than existing methods when any arbitrary view and illumination directions are given.

Current developments in sensors open new possible uses across numerous real-life applications, including optical character recognition (OCR). An OCR system requires incorporation of text processing tools into the sensor functionality. The most critical stage in OCR systems is the segmentation stage. It refers to the challenge of subdividing a text image into characters, which can be individually processed using a classifier. The cursive nature of the Arabic script such as the existence of different shapes for each character according to its location in the word besides the existence of diacritics makes Arabic character segmentation a very challenging task. A robust offline character segmentation algorithm for printed Arabic text with diacritics is developed based on the contour extraction technique. The algorithm works through extracting the up-contour part of a word and then identifies the splitting areas of the word characters. Then a postprocessing stage is used to handle the over-segmentation problems that appear in the initial segmentation stage. The proposed scheme is benchmarked using the APTI dataset and a manually collected dataset consisting of image texts varying in font size, type, and style for more than 38,000 words. The experiments show that the proposed algorithm is able to segment Arabic words with diacritics with an average accuracy of 98.5%.

Because of the limitations of the infrared imaging principle and the properties of infrared imaging systems, infrared images have some drawbacks, including a lack of details, indistinct edges, and a large amount of salt-and-pepper noise. Traditionally, the total variation (TV) regularization method with L1 norm is used for image deblurring in preserving edges and removing salt-and-pepper noise. However, the TV-based solutions usually have some staircase effects. To improve the sparse characteristics of the image while maintaining the image edges and weakening staircase artifacts, we propose a method that uses the Lp quasinorm instead of the L1 norm and for infrared image deblurring with an overlapping group sparse TV method. The Lp quasinorm introduces another degree of freedom, better describes image sparsity characteristics, and improves image restoration. Furthermore, we adopt the accelerated alternating direction method of multipliers and fast Fourier transform theory in the proposed method to improve the efficiency and robustness of our algorithm and use an inner loop nested within the optimization minimization iteration to solve the subproblem. Experiments show that under different conditions for blur and salt-and-pepper noise, the proposed method leads to excellent performance in terms of objective evaluation and subjective visual results.

Skeleton-based action recognition is a significant direction of human action recognition, because the skeleton contains important information for recognizing action. The spatial–temporal graph convolutional networks (ST-GCN) automatically learn both the temporal and spatial features from the skeleton data and achieve remarkable performance for skeleton-based action recognition. However, ST-GCN just learns local information on a certain neighborhood but does not capture the correlation information between all joints (i.e., global information). Therefore, we need to introduce global information into the ST-GCN. We propose a model of dynamic skeletons called attention module-based-ST-GCN, which solves these problems by adding attention module. The attention module can capture some global information, which brings stronger expressive power and generalization capability. Experimental results on two large-scale datasets, Kinetics and NTU-RGB+D, demonstrate that our model achieves significant improvements over previous representative methods.

Parallax disposition is a challenging task in image stitching. Thus far, the most effective method for parallax processing is the seam-driven method. However, its ability to handle complex textures is poor. To apply the seam-driven method to more scenes, we have incorporated the idea of spatially varying warps. Before finding the optimal seam, we try to register the overlapping area in as projective a manner as possible. Meanwhile, we set some constraints for the registered image to preserve the original appearance as much as possible. Among the constraints, we introduce local features to further improve the stitched image. After registration, we use the maximum flow method to obtain the optimal seam and introduce a series of seam evaluation systems. On one hand, these systems evaluate whether the resulting seam is misaligned, and on the other hand, they provide feedback to the previous registration, increasing the weights of the feature points near the dislocation point to fine-tune the seam. Comparisons with existing image stitching algorithms demonstrate the effectiveness and universality of our approach.