Search

Single-object tracking is regarded as a challenging task in computer vision, especially in complex spatio-temporal contexts. The changes in the environment and object deformation make it difficult to track. In the last 10 years, the application of correlation filters and deep learning enhances the performance of trackers to a large extent. This paper summarizes single-object tracking algorithms based on correlation filters and deep learning. Firstly, we explain the definition of single-object tracking and analyze the components of general object tracking algorithms. Secondly, the single-object tracking algorithms proposed in the past decade are summarized according to different categories. Finally, this paper summarizes the achievements and problems of existing algorithms by analyzing experimental results and discusses the development trends. [ABSTRACT FROM AUTHOR]

An important objective in computer vision is to analyze multiple images and subsequently reconstruct the shape and structure in 3D. Traditional multi-view 3D reconstruction techniques extract and match key features from images with known camera parameters. However, this approach is inefficient and fails to fully exploit the advantages of multi-view information. Advancements in deep learning have revolutionized multi-view 3D reconstruction by enabling end-to-end 3D shape inferencing without the need for sequential feature matching typically found in conventional algorithms. Recent rapid progress in this field necessitates a thorough review of current algorithms and provide insight into method of improving 3D reconstruction performance. This review classifies reconstruction algorithms according to their resultant model, including depth map, voxel, point cloud, mesh, and implicit surface. Additionally, this review encompasses the inclusion of frequently employed network training loss functions for network training, assessment metrics, and the incorporation of 3D datasets. Experimental results are also presented to assess the performance of different algorithms. Finally, the paper concludes with a summary, discussion of challenges, and potential future directions. [ABSTRACT FROM AUTHOR]

Motion blur can easily affect the quality of images. For example, Quick Response (QR) code is hard to be identified with severe motion blur caused by camera shaking or object moving. In this paper, a motion blur QR code identification algorithm based on feature extraction and improved adaptive thresholding is proposed. First, this work designs a feature extraction framework using a deep convolutional network for motion deblurring. The framework consists of a basic end-to-end network for feature extraction, an encoder-decoder structure for increasing training feasibility and several ResBlocks for producing large receptive fields. Then an improved adaptive thresholding method is used to avoid influence caused by uneven illumination. Finally, the proposed algorithm is compared with several recent methods on a dataset including QR code images influenced by both motion blur and uneven illumination. Experimental results demonstrate that the processing time and identification accuracy of the proposed algorithm are improved in executing motion blur QR code identification missions compared with other competing methods. [ABSTRACT FROM AUTHOR]

With the increasing maturity of convolution neural network (CNN) technology, the image super-resolution reconstruction (SR) method based on CNN is booming and has achieved many remarkable results. Undoubtedly, SR has become the mainstream direction of image reconstruction technology. However, most of the existing SR methods improve the reconstruction performance by increasing the depth of networks, which also increases the number of parameters, number of network computations, and difficulty of training network. To solve the performance complexity dilemma in SR, this paper proposes a network called a multi-scale dense recursive fusion network (MSDRFN). The network is composed of three parts: initial feature extraction module, multi-scale dense fusion group module and recursive reconstruction module. In detail, rough features are first extracted through a shallow feature extraction module, and then are inputted into multi-scale dense fusion blocks (MSDFBs) group. Each MSDFB makes full use of image features in convolution kernels of different sizes to obtain different hierarchical features, and further these output features are inputted into the channel attention mechanism to learn their corresponding weights. All MSDFBs outputs will be restored to high resolution images via the recursive reconstruction module. In addition, the network supplements the information loss with residual learning, which is embodied in one long-jump connection and several short-jump connections. The proposed network is mainly trained in the Pytorch deep learning framework. In comparison experiments on benchmark datasets, the proposed method outperformed the most advanced convolutional methods. [ABSTRACT FROM AUTHOR]

Deep Q-Networks algorithm (DQN) was the first reinforcement learning algorithm using deep neural network to successfully surpass human level performance in a number of Atari learning environments. However, divergent and unstable behaviour have been long standing issues in DQNs. The unstable behaviour is often characterised by overestimation in the Q -values, commonly referred to as the overestimation bias. To address the overestimation bias and the divergent behaviour, a number of heuristic extensions have been proposed. Notably, multi-step updates have been shown to drastically reduce unstable behaviour while improving agent's training performance. However, agents are often highly sensitive to the selection of the multi-step update horizon (n), and our empirical experiments show that a poorly chosen static value for n can in many cases lead to worse performance than single-step DQN. Inspired by the success of n -step DQN and the effects that multi-step updates have on overestimation bias, this paper proposes a new algorithm that we call 'Elastic Step DQN' (ES-DQN) to alleviate overestimation bias in DQNs. ES-DQN dynamically varies the step size horizon in multi-step updates based on the similarity between states visited. Our empirical evaluation shows that ES-DQN out-performs n -step with fixed n updates, Double DQN and Average DQN in several OpenAI Gym environments while at the same time alleviating the overestimation bias. [ABSTRACT FROM AUTHOR]

Recently, the Extreme Learning Machine (ELM) algorithm has been applied to various fields due to its rapidity and significant generalization performance. Traditionally, deep learning (DL) and wavelet neural networks (WNN) methods reach a high classification accuracy in machine learning applications. As a result, a new structure based on WNN, deep architecture and ELM is proposed in this paper. The proposed method is based on Extreme Learning Machine Auto-Encoder with DL structure and a composite wavelet activation function used in the hidden nodes. To evaluate the performance of our approach, we used standard benchmark data-sets, namely COIL-20, Pima Indian Diabetes (PID), MNIST and EMNIST. Experimental results show that our method offers satisfactory results and performance compared to other approaches. [ABSTRACT FROM AUTHOR]

As one of the core technologies of the automatic control system for unmanned surface vehicles (USVs), autonomous collision avoidance algorithm is the key to ensure the safe navigation of USVs. In this paper, path planning and dynamic collision avoidance (PPDC) algorithm which obeys COLREGs is proposed for USVs. In order to avoid unnecessary collision avoidance actions, the risk assessment model is developed, which is used to determine the switching time of path planning and dynamic collision avoidance. In order to train the algorithm which complies with the COLREGs, the encounter situation is divided quantitatively, which is regarded as the input state of the system, so that the high-dimensional input is successfully avoided. The state space of the USV is defined by relative parameters to improve the generalization ability of the algorithm, meanwhile, a network structure based on DDPG is designed to achieve the continuous output of thrust and rudder angle. Combined with path planning, collision avoidance, compliance with COLREGs and smooth arrival task, four kinds of reward functions are designed. In order to solve the problem of low training efficiency of experience replay mechanism in DDPG, cumulative priority sampling mechanism is proposed. Through the simulation and verification in a variety of scenarios, it is proved that PPDC algorithm has the function of path planning and dynamic collision avoidance in compliance with COLREGs, which has good real-time performance and security. [ABSTRACT FROM AUTHOR]

The technology for Visual Odometry (VO) that estimates the position and orientation of the moving object through analyzing the image sequences captured by on-board cameras, has been well investigated with the rising interest in autonomous driving. This paper studies monocular VO from the perspective of Deep Learning (DL). Unlike most current learning-based methods, our approach, called DeepAVO, is established on the intuition that features contribute discriminately to different motion patterns. Specifically, we present a novel four-branch network to learn the rotation and translation by leveraging Convolutional Neural Networks (CNNs) to focus on different quadrants of optical flow input. To enhance the ability of feature selection, we further introduce an effective channel-spatial attention mechanism to force each branch to explicitly distill related information for specific Frame to Frame (F2F) motion estimation. Experiments on various datasets involving outdoor driving and indoor walking scenarios show that the proposed DeepAVO outperforms the state-of-the-art monocular methods by a large margin, demonstrating competitive performance to the stereo VO algorithm and verifying promising potential for generalization. [ABSTRACT FROM AUTHOR]

Deep learning is an effective tool to assist doctors with many time-consuming and error-prone medical image analytical tasks. However, deep models are shown to be vulnerable to adversarial attacks, posing significant challenges to clinical applications. Existing works regarding the robustness of deep learning models are scarce, where most of them focus on the attack of medical image classification models. In this paper, a differential evolution attack (DEAttack) method is proposed to generate adversarial examples for medical image segmentation models. Our method does not require extra information such as the network's structures and weights compared with the most widely investigated gradient-based attack methods. Additionally, benefit from the embedded differential evolution algorithm, which can preserve diversities of the optimization space. The proposed method can achieve better results than gradient-based methods, which can successfully attack the segmentation model with only perturbing a small fraction of the image pixels, demonstrating that the medical image segmentation model is more susceptible to adversarial examples. In addition to evaluating model robustness attack with public datasets, our DEAttack method was also tested on the clinical diagnostic dataset, demonstrating its superior performance and elegant process for the robustness evaluation of deep models in medical image segmentation. [ABSTRACT FROM AUTHOR]

Many RGBT trackers utilize adaptive weighting mechanism to treat dual modalities differently and obtain more robust feature representations for tracking. Although these trackers work well under certain conditions, however, they ignore the information interactions in feature learning, which might limit tracking performance. In this paper, we propose a novel cross-modality message passing model to interactively learn robust deep representations of dual modalities for RGBT tracking. Specifically, we extract features of dual modalities by backbone network and take each channel of these features as a node of a graph. Therefore, all channels of dual modalities can explicitly communicate with each other by the graph learning, and the outputted features are thus more diverse and discriminative. Moreover, we introduce the gate mechanism to control the propagation of information flow to achieve more intelligent fusion. The features generated from the interactive cross-modality message passing model will be passed selectively through the gate layer and concatenated with original features as the final representation. We extend the ATOM tracker into its dual-modality version and combine it with our proposed module for final tracking. Extensive experiments on two RGBT benchmark datasets validate the effectiveness and efficiency of our proposed algorithm. [ABSTRACT FROM AUTHOR]

Lossy image compression is likely to produce blurred images which leads to erroneous image-level understanding. The human visual system (HVS) is focused on the area of interest present in the image. Motivated by this fact, we propose a compression-decompression algorithm focusing on the important content present in different parts of the image. Using the concept of the Guided Grad-CAM (Gradient-weighted Class Activation Mapping) technique to produce heat maps with the help of a heat map generator trained on ResNet-50, a saliency-guided encoding–decoding algorithm is developed. A wider multi-scale saliency guided convolutional neural network (MsSG-CNN) is designed, in which the notion of convolution with different size filters helps to obtain unique but different features. The feature extraction followed by multilevel fusion of features helps deep neural network (DNN) to capture contextual information and obtain high quality, good resolution, visually pleasing images with fine details. The proposed algorithm is tested on the Kodak benchmark dataset, CLIC 2019 challenging dataset, and FDDB facial images dataset. At low bit rates, the MS-SSIM of the proposed algorithm is found to be superior to JPEG, JPEG2000, BPG, WebP, and Minnen's approaches with approximately up to 60%, 24.80%, 11.43%, 23.08% & 75% gains respectively, which is quite a significant improvement, when tested on the Kodak dataset. Similarly, at high bit rates, the improvement in MS-SSIM is approximately up to 41.67%, 37.30%, 23.90% 34.21% & 13.33% when compared with JPEG, JPEG2000, BPG, WebP, and Minnen's approaches respectively. The improvement in PSNR at low and high bit rates is approximately up to 11.32%, 5%, 5.26% and 5.6%, 10.29%, 8.7% as compared to JPEG, Balle's, and Lee's algorithms respectively. The PSNR-HVS has been improved by approximately up to 27.27%, 19.15%, and 28.33%, 28% as compared to JPEG and Toderici's algorithms respectively at low and high bit-rates. A similar type of improvement is obtained with FDDB and CLIC 2019 datasets also, which is discussed in the paper. [ABSTRACT FROM AUTHOR]

• An algorithm that transforms long-tailed problem into hierarchical learning problem. • An effective knowledge transfer paradigm is proposed to help hierarchical learning. • The performance of our model outperforms the state-of-the-art methods. Deep visual recognition methods have achieved excellent performance on artificially constructed image datasets where the data distribution is balanced. However, in real-world scenarios, data distribution is usually extremely imbalanced and exhibit a long-tailed distribution where data in each head class is more than the class in the tail. Many efficient deep learning methods fail to work normally, i.e., they perform well in the head class while poor in the tail class. In this paper, we propose a two-layer Hierarchical-Learning Long-Tailed Recognition (HL-LTR) algorithm which transforms the long-tailed problem into a hierarchical classification problem by constructing a hierarchical superclass tree in which each layer corresponds to a recognition task. In the first layer of the tree, the degree of data imbalance is largely decreased. The recognition task of the second layer is the original long-tailed recognition problem. The training of HL-LTR is top-down. The knowledge learned by the first layer transfers to classes of the second layer and guides the feature learning of the second layer by using attention mechanism module and knowledge distillation method. Compared with directly solving the most difficult long-tailed recognition task, HL-LTR achieves better performance due to its progressive learning method from easy to difficult and effective knowledge transfer strategy. [ABSTRACT FROM AUTHOR]

Lossy image compression is likely to produce blurred images which leads to erroneous image-level understanding. The human visual system (HVS) is focused on the area of interest present in the image. Motivated by this fact, we propose a compression-decompression algorithm focusing on the important content present in different parts of the image. Using the concept of the Guided Grad-CAM (Gradient-weighted Class Activation Mapping) technique to produce heat maps with the help of a heat map generator trained on ResNet-50, a saliency-guided encoding–decoding algorithm is developed. A wider multi-scale saliency guided convolutional neural network (MsSG-CNN) is designed, in which the notion of convolution with different size filters helps to obtain unique but different features. The feature extraction followed by multilevel fusion of features helps deep neural network (DNN) to capture contextual information and obtain high quality, good resolution, visually pleasing images with fine details. The proposed algorithm is tested on the Kodak benchmark dataset, CLIC 2019 challenging dataset, and FDDB facial images dataset. At low bit rates, the MS-SSIM of the proposed algorithm is found to be superior to JPEG, JPEG2000, BPG, WebP, and Minnen's approaches with approximately up to 60%, 24.80%, 11.43%, 23.08% & 75% gains respectively, which is quite a significant improvement, when tested on the Kodak dataset. Similarly, at high bit rates, the improvement in MS-SSIM is approximately up to 41.67%, 37.30%, 23.90% 34.21% & 13.33% when compared with JPEG, JPEG2000, BPG, WebP, and Minnen's approaches respectively. The improvement in PSNR at low and high bit rates is approximately up to 11.32%, 5%, 5.26% and 5.6%, 10.29%, 8.7% as compared to JPEG, Balle's, and Lee's algorithms respectively. The PSNR-HVS has been improved by approximately up to 27.27%, 19.15%, and 28.33%, 28% as compared to JPEG and Toderici's algorithms respectively at low and high bit-rates. A similar type of improvement is obtained with FDDB and CLIC 2019 datasets also, which is discussed in the paper. [ABSTRACT FROM AUTHOR]

• Developing an under-clustering algorithm to extract the most possible clusters. • Involving heuristics and deep models in the framework for clustering effectiveness. • Making up the deep learning research gap in the field of plant disease annotation. Current annotation for plant disease images depends on manual sorting and handcrafted features by agricultural experts, which is time-consuming and labour-intensive. In this paper, we propose a self-supervised clustering framework for grouping plant disease images based on the vulnerability of Kernel K-means. The main idea is to establish a cross iterative under-clustering algorithm based on Kernel K-means to produce the pseudo-labeled training set and a chaotic cluster to be further classified by a deep learning module. In order to verify the effectiveness of our proposed framework, we conduct extensive experiments on three different plant disease datatsets with five plants and 17 plant diseases. The experimental results show the high superiority of our method to do image-based plant disease classification over balanced and unbalanced datasets by comparing with five state-of-the-art existing works in terms of different metrics. [ABSTRACT FROM AUTHOR]

• Introduce the text detection models based on regression and segmentation according to their implementation methods. • Classify mainstream text recognition models and provide a detailed introduction. • Compare the performance of different text detection and recognition methods. • Discuss the future development trends of OCR. Optical Character Recognition (OCR) poses a crucial challenge within the realm of computer vision research, as it plays a pivotal role in converting vast amounts of unstructured text data into structured formats to support diverse artificial intelligence applications. The OCR process encompasses two core components: text detection and text recognition. Text detection involves identifying and extracting text regions, achieved through either object detection or segmentation techniques, while text recognition focuses on accurately deciphering the content within these identified regions. In recent years, remarkable strides have been made in the domain of text recognition, primarily driven by deep learning-based models. These models eliminate the need for manual feature processing and excel in recognizing text even within complex scenes, surpassing the performance of traditional text recognition methods and subsequently emerging as the dominant approach. The objective of this paper is to present a comprehensive survey of both text detection and text recognition models. Firstly, we systematically categorize and provide an overview of existing off-the-shelf text detection methods. Subsequently, we conduct an in-depth investigation of six distinct text recognition models, taking into account their unique implementations. Additionally, we explore and analyze the principal datasets that currently prevail in the field of text detection and recognition. Furthermore, this research entails a meticulous performance comparison of various text detection algorithms on the CTW1500, TotalText, and ICDAR2015 datasets. Additionally, we evaluate and scrutinize the efficacy of mainstream text recognition algorithms on the IIIT-5K, SVT, ICDAR2013, SVT-P, CUTE80, and ICDAR2015 datasets. Finally, we conclude with a discussion on the future development and research trends concerning text detection and recognition, providing insights that can further drive progress in this crucial area. [ABSTRACT FROM AUTHOR]

In this paper we propose and discuss different Deep Learning-based ensemble algorithms for a problem of low-visibility events prediction due to fog. Specifically, seven different Deep Learning (DL) architectures have been considered, from which multiple individual learners are generated. Hyperparameters of the models, including parameters concerning data preprocessing, models architecture and training procedure, are randomly selected for each model within a pre-defined discrete range. Also, every model is trained with slightly different data sampled randomly, assuring that every models introduce variety in the ensemble. Then, three different information fusion techniques are employed to build the ensemble models. The influence of the filtering process and the elitism level (the percentage of the individual models entering the ensemble) is also assessed. The performance of the proposed methodology have been tested in two real problems of low-visibility events prediction due to orographical and radiation fog, at the north of Spain. Comparison with different Machine Learning, alternative DL algorithms and meteorological-based methods show the good performance of the proposed deep learning ensembles in this problem. [ABSTRACT FROM AUTHOR]