Canonical correlation analysis (CCA) is a popular method that has been extensively used in feature learning. In nature, the objective function of CCA is equivalent to minimizing the distance of the paired data, and L₂-norm is used as the distance metric. We know that L₂-norm-based objective function will emphasize the large distance pairs and de-emphasizes the small distance pairs. To alleviate the aforementioned problems of CCA, we propose an approach named CCA based on L₁-norm minimization (CCA-L1) for feature learning. To optimize the objective function, we develop an algorithm that can get a global optimized value. To maintain the distribution and the nonlinear characteristic respectively, we proposed two extensions of CCA-L1. Further, all of the aforementioned three proposed algorithms are extended to deal with multifeature data. The experimental results on an artificial dataset, real-world crop leaf disease dataset, ORL face dataset, and PIE face dataset show that our methods outperform traditional CCA and its variants.

Over the past decades, text recognition technologies have focused immensely on noncursive isolated scripts. A text recognition system for the cursive Pashto script will serve as a great contribution, allowing the traditional, cultural, and educational Pashto literature to be converted into machine-readable form. We propose the use of deep learning architectures based on the transfer learning for the recognition of Pashto ligatures. For recognition analysis and evaluation, the ligature images in the dataset are preprocessed by data augmentation techniques, i.e., negatives, contours, and rotated to increase the variation of each sample and size of the original dataset. Rich feature representations are automatically extracted from the Pashto ligature images using deep convolution layers of the convolution neural network (CNN) architectures using fine-tuned approach. Pretrained CNN architectures: AlexNet, GoogleNet, and VGG (VGG-16 and VGG-19) are used for classification by feeding the extracted features to a fully connected layer and a softmax layer. The proposed deep transfer-based learning has achieved phenomenal recognition rates for Pashto ligatures on benchmark FAST-NU Pashto dataset. An accuracy of 97.24%, 97.46%, and 99.03% is achieved using AlexNext, GoogleNet, and VGGNet architectures, respectively.

Optimum compression ratio (OCR) of neutron and x-ray radiography images with minimum decomposition distortion is the main target of this paper. Artificial bee colony (ABC) and firefly optimal algorithms are applied for achieving this goal. These algorithms are applied in conjunction with proposed image encoding/decoding and decomposition/decompression algorithms. The encoding/decoding algorithms depend on binary image, Huffman, and hierarchical trees. In addition, the decompression algorithms use real dual-tree transform, Hilbert transform, real a trou transform, and Gaussian pyramidal transform. A fitness function (FF) that correlates to compression ratio and image quality is conducted. The OCR with the best radiography image quality is the purpose of the suggested FF. A comparison between the optimized compressed ratio of x-ray and neutron radiography images for both optimization techniques is demonstrated. The proposed algorithms are validated. For the ABC algorithm, optimum values of 80.3610% and 80.3610% are fulfilled for x-ray and neutron images via hierarchical trees with Hilbert transforms. An optimum value of 85.333333% is attained by the Huffman algorithm with discrete cosine transform for both images. For the firefly optimization algorithm (FOA), the hierarchical trees with Hilbert transform achieve an optimum value of 97.0499% for both images. The binary image encoding with pyramidal transform gives optimum values of 68.8055% and 77.4566% for x-ray and neutron images, respectively. The FOA is noted to accomplish more advanced results over the ABC algorithm for compressed radiography images. Additionally, the binary image algorithm with a trou decompression provides an optimum value of 76.77% for neutron radiography images. The achieved optimum values of neutron images are better than of x-ray images due to the absorption of more energy by neutron radiation. Robustness and the significant task of the supposed FF are clarified. This confirms the applicability of image compression optimization algorithms with neutron and x-ray radiography images.

A capsule network encodes entity features into a capsule and maps a spatial relationship from the local feature to the overall feature by dynamic routing. This structure allows the capsule network to fully capture feature information but inevitably leads to a lack of spatial relationship guidance, sensitivity to noise features, and easy susceptibility to falling into local optimization. Therefore, we propose a novel capsule network based on feature and spatial relationship coding (FSc-CapsNet). Feature and spatial relationship extractors are introduced to capture features and spatial relationships, respectively. The feature extractor abstracts feature information from bottom to top, while attenuating interference from noise features, and the spatial relationship extractor provides spatial relationship guidance from top to bottom. Then, instead of dynamic routing, a feature and spatial relationship encoder is proposed to find the optimal combination of features and spatial relationships. The encoder abandons the idea of iterative optimization but adds the optimization process to the backpropagation. The experimental results show that, compared with the capsule network and its multiple derivatives, the proposed FSc-CapsNet achieves significantly better performance on both the Fashion-MNIST and CIFAR-10 datasets. In addition, compared with some mainstream deep learning frameworks, FSc-CapsNet performs quite competitively on Fashion-MNIST.

Person re-identification (re-ID) aims at matching two pedestrian images across different cameras. Usually, the main scheme of re-ID based on deep learning includes two phases: feature extraction and metric calculation. We focus on how to extract more discriminative image features for re-ID. To address this problem, we propose a multilevel deep representation fusion (MDRF) model based on the convolutional neural network. Specifically, the MDRF model is designed to extract image features at different network levels through one forward pass. In order to produce the final image representation, these multilevel features are fused by a fusion layer. Then the final image representation is fed into a combined loss of the softmax and the triplet to optimize the model. The proposed method not only utilizes the abstract information of high-level features but also integrates the appearance information of low-level features. Extensive experiments on public datasets including Market-1501, DukeMTMC-reID, and CUHK03 demonstrate the effectiveness of the proposed method for person re-ID.

Recompression-based image forensics has been a largely investigated area in the forensic community over recent years. This mainly involves detecting the number of different compressions in an image by exploiting certain artifacts left behind by the compression stages. The artifacts left in images by Joint Photography Experts Group (JPEG) compression coding operations play an important role in passive forensic analysis of images, which help in assessing the authenticity of the image. In this regard, most of the solutions proposed to date are restricted to double compression detection in JPEG images. However, in practical cases, this restriction may not always work. This is because an image having undergone multiple stages of compression, including triple or higher degrees of compression, will be “missed” by such forgery detection algorithms. To address this issue, we propose a deep convolutional neural network (CNN) model that detects both double and triple compression in JPEG images. The model is trained with the compression features specific to single, double, and triple JPEG compressions. The trained model is capable of determining the authenticity of an image by distinguishing between tampered and untampered image regions. Further, the proposed model localizes the forged regions within an image, distinguishing between double and triple compressed regions. Our experimental results establish the efficiency of the proposed model.

We develop a two-dimensional enhanced hyperchaotic Henon map (2D-EHHM) to overcome the problems of small chaotic range and poor security when the 2-D traditional Henon map is implemented in cryptosystems. The performance evaluations show that the 2D-EHHM has wider chaotic range, higher chaotic complexity, better ergodicity, and hyperchaotic behavior compared with certain existing chaotic maps. Based on the 2D-EHHM, we further design an efficient image encryption algorithm consisting of a multiple block substitution stage (MBSS) and a bidirectional-dynamic diffusion stage (BDDS). In the MBSS, a plain image is divided into several nonoverlapping multiple blocks to carry out permutation operation. In the BDDS, the scrambled image is redivided into nonoverlapping sub-blocks of the same size to be diffused dynamically in the forward and backward directions. Moreover, the SHA 512 function is employed to obtain a 512-bit plain image hash value treated as a raw key. A variable-length secret key (at least 128 bits), which is dynamically selected from the raw key and regarded as a valid key, is utilized to generate the initial values for the chaotic system. Simulation results and security analysis show that the proposed algorithm can resist various cryptanalytic attacks and can be applied to real-time data transmission.

The coal-rock fractures formed by natural geological evolution are complex and the shapes are not fixed, which makes it difficult to manually define the features of coal-rock fractures. Based on the full convolutional neural network U-net, we propose a convolutional neural network with fusion parallel multiscale features (FPMF-U-net). The FPMF-U-net uses two feature extraction networks in parallel to extract multiscale features. The feature fusion layer in this network is responsible for weighted fusion of the shallow high-resolution features and the deep abstract features. Average pooling operator is used in the network to avoid loss of weak boundary fractures caused by max pooling. The FPMF-U-net can automatically extract fracture features and segment coal-rock fractures from images. According to the three-dimensional gradual change feature of the fracture shapes between adjacent images, the segmentation results of the FPMF-U-net are further optimized. Due to the lack of training samples of the fracture, we use a data augmentation technique to increase the number of training samples. At the same time, the transfer learning method is used to improve the convergence speed of the FPMF-U-net. The experimental results show that the proposed FPMF-U-net has a good effect on the segmentation of coal-rock fractures.

Under the condition of a low sampling rate, hyperspectral image (HSI) reconstruction faces important challenges in remote sensing. How to efficiently process HSI data is an urgent problem to be solved. We propose a hyperspectral image compressive sensing reconstruction (HSI-CSR) model based on tensor decomposition and a low-rank constraint. This model can efficiently exploit the underlying structure information in the HSI. Specifically, we study how to exploit reasonably the low-rank constraint of the core tensor and nonlocal self-similarity, respectively, to explore the nonlocal spatial–spectral similarity hidden in an HSI. To solve the proposed HSI-CSR model, we design an efficient algorithm based on the alternating direction method of multipliers knowledge. Finally, extensive simulations show that the proposed model achieves superior reconstruction performance, compared with other state-of-the-art methods.

Due to considerable diversities of multimodality remote sensing images in spectral component, the performances of scale-invariant feature transform (or SIFT) may be problematic. In view of this, a model of image registration based on tangent-crossing-point feature is proposed. With the help of tangent-cross-point feature, verifying the correctness of the feature matching is very easy, since it adopts the location information indexed by the correct matching pair of feature-points and transformation information determined by the correct matching pair of curves simultaneously. Experimental results show that this method is more efficient and reliable than the classic SIFT algorithm in some cases.

Deep convolutional neural networks show a good prospect in the fertility detection and classification of specific pathogen-free hatching egg embryos in the production of avian influenza vaccine, and our previous work has mainly investigated three factors of networks to push performance: depth, width, and cardinality. However, an important problem that feeble embryos with weak blood vessels interfering with the classification of resilient fertile ones remains. Inspired by fine-grained classification, we introduce the attention mechanism into our model by proposing a dense pixelwise spatial attention module combined with the existing channel attention through depthwise separable convolutions to further enhance the network class-discriminative ability. In our fused attention module, depthwise convolutions are used for channel-specific features learning, and dilated convolutions with different sampling rates are adopted to capture spatial multiscale context and preserve rich detail, which can maintain high resolution and increase receptive fields simultaneously. The attention mask with strong semantic information generated by aggregating outputs of the spatial pyramid dilated convolution is broadcasted to low-level features via elementwise multiplications, serving as a feature selector to emphasize informative features and suppress less useful ones. A series of experiments conducted on our hatching egg dataset show that our attention network achieves a lower misjudgment rate on weak embryos and a more stable accuracy, which is up to 98.3% and 99.1% on 5-day and 9-day old eggs, respectively.

A robust active contour model driven by an order-statistic filtering and coherence-enhancing diffusion (OSFCED) filter for fast image segmentation is proposed. The main idea of the order-statistic filtering is to construct the edge force function (EFF) to quickly and adaptively attract evolving curves to target boundaries while a coherence-enhancing diffusion (CED) filter aims at filtering noise and enhancing target boundaries so that the segmentation efficiency can be improved. In addition, the computation of the EFF and CED function is completed before iterations. Therefore, the computational cost of the proposed model is low during curve evolution. Furthermore, the addition of optimized distance regularization term and optimized length term makes curve evolution smoother and more stable. Experiments performed show that the proposed model is robust to initial contour, parameter and has higher segmentation efficiency for images with intensity inhomogeneity.

Deep learning has been attributed to incredible advances in computer vision, natural language processing, and general pattern understanding. Recent discoveries have enabled efficient vector representations of both visual and written stimuli. Robustly transferring between the two modalities remains a challenge that could yield benefits for search, retrieval, and storage applications. We introduce a simple, yet highly effective approach for building a connection space where natural language sentences are tightly coupled with visual data. In this connection space, similar concepts lie close, whereas dissimilar concepts lie far apart, irrespective of their modality. We introduce an attention mechanism to align multimodal embeddings that are learned through a multimodal metric loss function. We evaluate the learned common vector space on multiple image–text datasets—Pascal Sentences, NUS-WIDE-10k, XMediaNet, Flowers, and Caltech-UCSD Birds. We extend our method to five different modalities of image, sentence, audio, video, and 3D models to demonstrate cross-modal retrieval on the XMedia dataset. We obtain state-of-the-art retrieval and zero-shot retrieval across all datasets.

Representation-based classification methods (RBCM) remain one of the hottest topics in the community of pattern recognition, and the recently proposed non-negative representation-based classification (NRC) achieved impressive recognition results in various classification tasks. However, NRC ignores the relationship between the coding and classification stages. Moreover, there is no regularization term other than the reconstruction error term in the formulation of NRC, which may result in an unstable solution leading to misclassification. To overcome these drawbacks of NRC, we propose a class-specific residual constraint non-negative representation (CRNR) for pattern classification. CRNR introduces a class-specific residual constraint into the formulation of NRC, which encourages training samples from different classes to competitively represent the test sample. Based on the proposed CRNR, we develop a CRNR-based classifier (CRNRC) for pattern classification. Experimental results on several benchmark datasets demonstrate the superiority of CRNRC over conventional RBCM as well as the recently proposed NRC. Moreover, CRNRC works better than or comparable to some state-of-the-art deep approaches on diverse challenging pattern classification tasks.

Color, which is effective information for computer vision, is vulnerable to variations in illumination. It is necessary to achieve illumination invariance for the description of an image’s color information. We propose two new color spaces that are robust to illumination changes: the color information color space based on the diagonal-offset model and the edge color space, the two channels of which are derived from the spherical color space and the linear transformation between the dark channel and the bright channel. Then, we propose a color moment invariant descriptor according to Hu moment invariants in the two new color spaces. We test the performance of the color descriptors with illumination variations in the context of object recognition and image retrieval. Experiments show that our descriptor has satisfactory stability for light variations.

The infrared image (RI) and visible image (VI) fusion method merges complementary information from the infrared and visible imaging sensors to provide an effective way for understanding the scene. The graph filter bank-based graph wavelet transform possesses the advantages of the classic wavelet filter bank and graph representation of a signal. Therefore, we propose an RI and VI fusion method based on oversampled graph filter banks. Specifically, we consider the source images as signals on the regular graph and decompose them into the multiscale representations with M-channel oversampled graph filter banks. Then, the fusion rule for the low-frequency subband is constructed using the modified local coefficient of variation and the bilateral filter. The fusion maps of detail subbands are formed using the standard deviation-based local properties. Finally, the fusion image is obtained by applying the inverse transform on the fusion subband coefficients. The experimental results on benchmark images show the potential of the proposed method in the image fusion applications.

We address kinship verification, which is a challenging problem in computer vision and pattern discovery. It has several applications, such as organizing photoalbums, recognizing resemblances among humans, and finding missing children. We present a system for facial kinship verification based on several kinds of texture descriptors (local binary patterns, local ternary patterns, local directional patterns, local phase quantization, and binarized statistical image features) with pyramid multilevel (PML) face representation for feature extraction along with our proposed paired feature representation and our proposed robust feature selection to reduce the number of features. The proposed approach consists of the following three main stages: (1) face preprocessing, (2) feature extraction and selection, and (3) kinship verification. Extensive experiments are conducted on five publicly available databases (Cornell, UB KinFace, Family 101, KinFace W-I, and KinFace W-II). Additionally, we provided a wide experiment for each stage to find the best and most suitable settings. We present many comparisons with state-of-the-art methods and through these comparisons, it appears that our experiments show stable and good results.

Automatic emotion recognition for video clips has become a popular area of research in recent years. Previous studies have explored emotion recognition methods through monomodal approaches, such as voice, text, facial expression, and physiological information. We focus on the complementarity of the information and construct an automatic emotion recognition model based on deep learning technology and multimodal fusion strategy. In this model, visual features, audio features, and text features are extracted from the video clips. A decision-level fusion strategy, based on the theory of evidence, is proposed to fuse the multiple classification results. To solve the problem of evidence conflict in evidence theory, we study a compatibility algorithm designed to correct conflicting evidence based on the similarity matrix of the evidence. This approach is shown to improve the accuracy of emotion recognition.

Facial expression recognition (FER) in an uncontrolled environment is difficult due to changes in occlusion, illumination, noise, and personal attributes. A deep learning enhanced gender conditional random forest (G_DRF) is proposed for FER in an uncontrolled environment. In order to reduce the influence of occlusion, illumination, low image resolution, etc., our method extracts robust facial features by deep multi-instance learning. Then a G_DRF is devised to address the facial personal attributes’ influence, such as gender variation by conditional RF. A large number of experiments were conducted on the public CK+, BU-3DEF, and LFW face databases. The experimental results showed that the proposed method had better performance and robustness than the state-of-the-art methods. The recognition rates of CK+, BU-3DEF, and LFW were 98.83%, 90%, 60.58%, respectively. Compared with other advanced deep learning methods that require a large number of training samples, the proposed method needs a small number of training samples and achieves better results.

Rain streaks blur and degrade the color fidelity of images. This affects road segmentation accuracy in advanced driver assistance systems. To address this problem, a method combining the gradient property with the rain–fog model is proposed to remove rain streaks in single images. The rain detection model is used to detect rain streaks, which aids training for rain removal and determines if rain streaks exist in images. The rain removal model is based on trained patches with the highest proportion of rain streaks in the high-frequency layer for low-cost computation. In order to recover nonrain images without oversmoothing, the gradient property is used prior to handling the overlapping rain streaks in the background layer. The rain–fog model is employed to remove veiling effects and moderately enhance background scenes. Our results showed that this method outperforms existing methods in regard to visual performance and quantitative aspects.

Color-coded fringe pattern is usually used to speedup phase unwrapping in surface profilometry. However, color-coded fringe pattern is subject to phase errors due to gamma nonlinearity and color cross talk. The phase error is conventionally reduced by isolating the color channels and by compensating the phase error in each channel. In this work, a phase error feedback calibration algorithm is proposed to eliminate the phase errors in a single process. The percentage phase errors are reduced from 1.08  ±  2.32  %   to 0.73  ±  0.46  %   with 95% level of confidence after an average of eight iterations of the proposed algorithm. The 3D reconstruction of the objects is improved in terms of accuracy and surface smoothness after calibration.

Recently, deep neural networks have made remarkable performance in the image super-resolution (SR) field. However, they mainly focus on wider or deeper architectural design, neglecting to capture the inherent property of natural images, hence hindering the representational ability of convolutional neural networks. To address this issue, we propose a deep network based on nonlocal (NL) and second-order (SO) feature fusion for image SR. In particular, we draw the observation that a SO attention mechanism could achieve more powerful feature expression and feature correlation learning. On the other hand, NL module is proved to be an effective prior to explore spatial contextual information. Thus, we introduce an SR network architecture by embedding NL operations and SO feature to capture intrinsic statistical characteristics of images. Furthermore, long skip connection is applied in the network to pass more abundant low-frequency information from low-resolution images and ease the network training. Experimental results on a variety of images demonstrate that our proposed method can achieve more desirable performance over several state-of-the-art methods in terms of quantitative metrics and visual quality.

We propose an image encryption algorithm based on gradient decomposition and an improved logistic map. The original image is first decomposed into three subimages with different gradient sizes using the gradient decomposition. The pixel position of the three subimages is then changed by applying Arnold transformation, and the random sequence generated by the improved logistic map is used to perform the XOR operation on the three subimages. Subsequently, the row pixel values of the three subimages are converted into binary sequences, and then the binary sequences are shuffled with the random sequences generated by the improved logistic map. Finally, the three subimages are obtained as the R, G, and B components of the color image to generate the final encrypted image. For image decryption, the original image can be simply restored by joining the three subimages. The experimental results show that the improved logistic map has a larger key space and better randomness compared with the classical logistic map. Furthermore, the proposed encryption scheme exhibits better robustness against various attacks than several existing encryption algorithms.

Since earprint is an encouraging physical trait that has been recently promoted as a biometric asset, we propose it as an alternative to other popular biometrics thanks to its uniqueness and stability. We propose an approach for ear recognition to smart home access in degraded conditions based on local and frequency domain features. The saliency is estimated with the dual tree complex wavelets in five scales and six rotation angles. Several statistic features are generated from the extracted feature vector and its first and second derivatives. The Harris descriptors are deployed to extract corner points invariant to scale, translation, and rotation. All the extracted features are fused at a feature level. To evaluate our research, we use the USTB-I and EVDDC databases. Different classifiers are utilized in the evaluation like the support vector machine, the K-nearest neighbor, and the random forest. The best recorded accuracy is 93.88% and 92.5%, respectively, with the KNN classifier.

A new and an efficient image cipher using chaos theory and the king chess piece is proposed. The given image is being treated as a large chessboard. The king scrambles the pixels of the given image through its random movement on the chessboard. Once the image is scrambled, it is further diffused through the XOR operation. Chaotic data to be used in the cipher have been generated through the Tinkerbell chaotic map. Image characteristics have been directly used to embed the plaintext sensitivity in the cipher to avoid the complicated and lengthy hash codes. The starting address of the king has also been made a key. It increases both the key space and plaintext sensitivity of the proposed image cipher. Both the simulation and comprehensive security analyses exhibit the efficiency, robustness, security, and applicability of the proposed algorithm.

We present a local stereo matching method with a robust texture category-based matching cost and adaptive support window to deal with the disparity errors caused by repetitive patterns, occlusions, and nontextured cases. First, we decompose an input reference image into textured regions and nontextured regions. Then, different cost computation strategies are adopted for these two regions. For textured regions, we use the common absolute intensity difference and gradient similarity. For nontextured regions, we propose a matching cost computation method that is a combination of gradient and epipolar distance transform (EDT). In the cost aggregation step, we introduce an adaptive support window based on a modified linearly expanded cross skeleton. To obtain the cross skeleton, a depth edge detection technique and a triple expansion strategy are presented. The experimental results demonstrate that the proposed algorithm achieves outstanding matching performance compared with other existing local algorithms on the Middlebury stereo benchmark, especially in repetitive patterns, occlusions, and nontextured regions.

Most previous target detection methods are based on the physical properties of visible-light polarization images, depending on different targets and backgrounds. However, this process is not only complicated but also vulnerable to environmental noises. A multimodal fusion detection network based on the multimodal deep neural network architecture is proposed in this research. The multimodal fusion detection network integrates the high-level semantic information of visible-light polarization image in crater detection. The network contains the base network, the fusion network, and the detection network. Each of the base networks outputs a corresponding feature figure of polarization image, fused by the fusion network later to output a final fused feature figure, which is input into the detection network to detect the target in the image. To learn target characteristics effectively and improve the accuracy of target detection, we select the base network by comparing between VGG and ResNet networks and adopt the strategy of model parameter pretraining. The experimental results demonstrate that the simulated crater detection performance of the proposed method is superior to the traditional and single-modal-based methods in that the extracted polarization characteristics are beneficial to target detection.

The existing approaches usually perform facial landmark detection and head pose estimation independently and sequentially, ignoring their coupled relations. We introduce a unified framework, named coupled cascade regression (CCR), for simultaneous facial landmark detection and head pose estimation. Based on the cascade regression framework, we propose to learn two separate regressors to update the landmark locations and three-dimensional (3D) face model parameters at each cascade level. To capture the coupled relations of the landmark locations and head pose, we further apply the 3D face projection model to refine the prediction results in each cascade iteration and make them consistent. CCR can leverage both the learning methods and the projection model to simultaneously perform facial landmark detection and pose estimation to enhance the performances of both tasks. We also propose to learn the cascade regressors from the combination of real and synthesized face images to solve the problem of limited variations in head pose for training. Experimental results on Helen, labeled face parts in the wild, 300-W, and Boston University datasets show that our proposed CCR method outperforms other conventional methods both for landmark detection and head pose estimation.

Correlation filters are known to have superior performance in tracking speed. To improve tracking results, we propose a context-aware regression correlation filter with a spatial–temporal regularization for tracking. First, the spatial regularization parameters are computed by the spatial correlation between the target and the surrounding information. Meanwhile, the filter weight distribution is also computed to highlight the target region and suppress the background region. Next, a context-aware model is proposed to adaptively expand the search area of the target with the original regression analysis. Due to the introduction of the context-aware information, the target structure will be changed and the regression model will not be adapted to the Gaussian function. Therefore, an optimized regression objective function is constructed according to the context-aware model. In the tracking process, under rotation, out-of-view, deformation, etc., the tracker still can continue the tracking of subsequent frames according to partial background information. In addition, time variables for online updating of the spatial–temporal model employ the similarity of adjacent frames to achieve more accurate tracking. According to the model form, the alternative direction method of multipliers is used to optimize the model optimization. Extensive experimental results on the OTB-2013 and OTB-2015 dataset object tracking benchmarks prove that the proposed algorithm is superior to other state-of-the-art algorithms in terms of success, accuracy, and robustness.