Multiple object tracking in dense scenes presents a significant challenge because of mutual occlusion and redundant detections, which can result in feature loss and cumulative error. To address these challenges, joint-detection-and-tracking frameworks based on the bipartite graph model have emerged as a popular paradigm. However, the bipartite graph model is often constrained by König’s minimax theorem, which postulates the need for one-to-one relationships in matching and equates the maximum number of matches with the minimum number of points covered. This restriction easily causes tracking failure for frequently appearing similarity targets. To overcome these limitations, this paper proposes a hypergraph random field model that uses domain hypernodes and trajectory hypernodes to differentiate targets on the basis of domain space features and trajectory fragments, respectively. This approach avoids cumulative error and enables one-to-many relationships for efficient tracking. Additionally, this paper presents an approximate solution that reduces the original complexity from O(n^2) to O(n). The experimental results on MOTChallenge demonstrate competitive performance, with a 1-2

Due to scene complexity, sensor inaccuracies, and processing imprecision, point cloud corruption is inevitable. Over-reliance on input features is the root cause of DNN vulnerabilities. It remains unclear whether this issue exists in 3D tasks involving point clouds and whether reducing dependence on these features can enhance the model’s robustness to corrupted point clouds. This study attempts to answer these questions. Specifically, we quantified the sensitivity of the DNN to point cloud features using Shapley values and found that models trained using traditional methods exhibited high sensitivity values for certain features. Furthermore, under an equal pruning ratio, prioritizing the pruning of highly sensitive features causes more severe damage to model performance than random pruning. We propose ‘Desensitized Adversarial Training’ (DesenAT), generating adversarial samples using feature desensitization and conducting training within a self-distillation framework, which aims to alleviate DNN’s over-reliance on point clouds features by smoothing sensitivity. First, data points with high contribution components are eliminated, and spatial transformation is used to simulate corruption scenes, generate adversarial samples, and conduct adversarial training on the model. Next, to compensate for information loss in adversarial samples, we use the self-distillation method to transfer knowledge from clean samples to adversarial samples, and perform adversarial training in a distillation manner. Extensive experiments on ModelNet-C and PointCloud-C demonstrate show that the propose method can effectively improve the robustness of the model without reducing the performance of clean data sets. This code is publicly available at https://github.com/JerkyT/DesenAT.

It is still a challenge to robustly extract discriminative features from video for multi-object tracking (MOT), due to the problems of frequent occlusions and appearance degradation. Current Transformer-based trackers lack an explicit spatial inductive bias, making target tokens susceptible to background contamination and erroneous context aggregation during occlusion. To address these limitations, this paper reports an occlusion-robust framework, named as STGMA, from the perspective of spatiotemporal token evolution. Specifically, we introduce a Recursive Gaussian Calibration (RGC) module, which can effectively model the spatial uncertainty of tokens, to inject ''temporal inertia'' by leveraging high-confidence historical distributions. By generating Positional Anchoring Tokens, RGC constrains the attention distribution to continuously lock onto the target subject, thereby preventing the attention focus from drifting to adjacent distractors. Concurrently, we design a Global Magnitude-aware Aggregation (GMA) module to represent the positive correlation between feature magnitude and semantic saliency for token-level semantic repair. By leveraging magnitude-aware linear attention to generate Semantic Compensatory Tokens, GMA efficiently amplifies weak residual signals and reconstructs missing semantics through global topological modeling. Extensive experiments on the MOT17, MOT20, and DanceTrack benchmarks demonstrate that our STGMA achieves competitive performance, significantly enhancing trajectory consistency and discriminability in complex interactive scenarios.

In Multiple Object Tracking (MOT), target interactions and occlusions in dense scenes increase the risk of identity changes for similar targets. To overcome these challenges, the use of higher-order semantic relations can accurately capture complex target interactions and help distinguish similar targets in complex environments. However, due to the heterogeneous nature of semantic cues, traditional methods struggle to efficiently integrate such cues into a unified framework, which limits the effective acquisition of higher-order semantic relations and affects the stability and accuracy of tracking. To overcome these issues, this paper transforms the MOT problem into a task for solving node relations in heterogeneous bipartite graphs and proposes a Heterogeneous Graph Network (HGN) model for capturing higher-order semantic relations. Compared to other approaches, the advantage of this model lies in the integration of heterogeneous clues into the model by constructing different semantic edges within the heterogeneous bipartite graph. To further improve the tracking efficiency, a Latent Trajectory Generation (LTG) module is introduced to predict and identify latent target trajectories to speed up the HGN solving process. Experimental results show that this model improves Multiple Object Tracking Accuracy (MOTA) results in the MOTchallenge and DanceTrack by 1-4 %, with minimal impact on processing speed. For ease of reading, this paper summarizes some key Nomenclature in Table 1.

Resource allocation in edge computing is a research hotspot and difficulty in academia and Industry. The nature like urgency and the priority of tasks are not taken into account, which is adverse to obtain a good solution. Meanwhile, 5G has the characteristics of higher network speed, high reliability, low latency, and low-power massive connections. In this article, we present a novel algorithm to solve the multi-objective resource allocation problem in 5G edge computing (EC) network, the objective is to maximize the operator profit and minimize the total completion time of tasks with priorities from the perspective of service operators under time and workload constraints. The algorithm is based on the beluga whale optimization algorithm, and it utilizes three methods to update the positions of beluga whales by swimming, predating, and migrating. In addition, to enhance the ability to escape from local optima during searching the best beluga whale position, it uses centroid information to improve the process of searching the optimal position, and adds the mutation operation in the process of position updating. Simulated results show that the proposed algorithm is high efficient in terms of reducing total task completion time and improving the revenue for operators, compared with existing strategies. For example, our algorithm reduces the time by 1.87% and improves the profit by 10.47% for 80 tasks in comparison with MOPSO.

In the domain of intelligent inspection, the precise recognition of pointer meter readings is of paramount importance for monitoring equipment conditions. To address the challenges of insufficient robustness and diminished detection accuracy encountered in practical applications of existing methods for recognizing pointer meter readings based on object detection, we propose a novel approach that integrates YOLOv11-OBB rotating object detection with adaptive template matching techniques. Firstly, the YOLOv11 object detection algorithm is employed, incorporating a rotational bounding box (OBB) detection mechanism; This effectively enhances the feature extraction capabilities related to pointer rotation direction and dial center, thereby boosting detection robustness. Subsequently, an enhanced angle resolution algorithm is leveraged to develop a mapping model that establishes a relationship between pointer the deflection angle and the instrument range, facilitating precise reading calculation. Experimental findings demonstrate that the proposed method achieves a mean Average Precision (mAP) of 99.1% in a self-compiled pointer instrument dataset. The average relative error of readings is 0.41568%, with a maximum relative error of less than 1.1468%. Furthermore, the method exhibits robustness and reliability when handling low-quality meter images characterized by blur, darkness, overexposure, and tilt. The proposed approach provides a highly adaptable and reliable solution for pointer meter reading recognition in the intelligent industrial field, with significant practical value.

Artificial intelligence-assisted imaging biosensors have attracted increasing attention due to their flexibility, allowing for the digital image analysis and quantification of biomarkers. While deep learning methods have led to advancements in biomarker identification, the diversity in the density and adherence of targets still poses a serious challenge. In this regard, we propose CellNet, a neural network model specifically designed for detecting dense targets. The model uses a shape-aware radial basis function to learn the kernel representation of objects, improving the target counting accuracy, and exhibits excellent performance in identifying adherent polystyrene microspheres, with a detection accuracy of 98.39 %. Considering these factors, we developed a biotin-streptavidin-based biosensing method using artificial intelligence transcoding (bs-SMART) to detect procalcitonin in serum samples. Given its excellent accuracy and sensitivity (limit of detection = 8.5 pg/mL), the technique provides a reliable platform for the accurate diagnosis of diseases. Furthermore, this study validated the ability of CellNet to recognize irregular and adherent cells. Overall, CellNet not only contributes to advancing computer vision and image processing technology but also presents potential benefits for medical diagnostics, food safety testing, and environmental monitoring.

In recent years, the realism of generated face images has significantly improved, making it increasingly difficult for the public to distinguish between real and fake. Existing detection methods often struggle to capture local artifacts in these generated images, presenting limitations when confronted with highly realistic synthetic faces. To address this, this paper proposes a completely new evaluation metrics, the Local Artifact Metrics (LAM), which quantifies the degree of artifacts in key regions of generated face images, providing a more accurate assessment of overall image authenticity. In addition, to enhance the ability to capture generation artifacts, we introduce gradient maps into the Convolutional Block Attention Module (CBAM), resulting in a modified attention mechanism referred to as the Gradient-Enhanced Attention Mechanism (GEAM). Experimental results demonstrate that the proposed generated faces detection method achieves superior detection performance across multiple fake face datasets.

Text in natural scenes exhibits diverse scales and morphologies, which is particularly pronounced in Chinese text. This cross-scale characteristic limits feature propagation in existing cross-scale fusion methods and tends to overlook the capture of detailed features. This paper proposes FDB, a natural scene Chinese text detection model based on DBNet, incorporating Feature Pyramid Enhancement Module with Attention Mechanism. Through a collaborative mechanism of cross-scale feature fusion and channel attention, the model effectively enhances its cross-scale feature integration capability and improves Chinese text feature capture. To address the lack of Chinese scene adaptability in existing datasets, this study specifically constructs a Chinese Text Dataset for model training. Experimental results on public benchmarks demonstrate that the FDB model achieves enhanced validity in Chinese text detection under complex backgrounds, while achieving an improvement in F1-score.

To improve the reliability of 3-D point cloud measurements acquired by instrumentation systems such as LiDAR, we propose skeleton-enhanced geometric information reinforcement (SEGIR), a novel training framework that uniquely leverages the skeleton structure of point clouds to preserve intrinsic geometric information. Unlike existing methods that focus on adversarial training or sample mixing without considering shape topology, SEGIR introduces a dual-domain enhancement strategy: 1) in the graph spectral domain, we propose a new frequency-domain fusion technique that integrates structural signals from the skeleton using a CutMix-style operation and 2) in the spatial domain, we reinforce geometric learning by preserving the neighborhood clusters surrounding skeleton points. We further introduce a consistency regularization method to maintain semantic consistency between clean and perturbed samples. Notably, the skeleton is only used during training, ensuring no additional computational burden during inference. Extensive experiments show that SEGIR achieves state-of-the-art performance on ModelNet-C and PointCloud-C, clearly outperforming existing robust training approaches.

In the realm of multi-object tracking, the SORT model is esteemed for its simplicity and efficiency, yet its tracking efficacy hinges significantly on detector performance. Randomly discarding low-threshold detections can result in critical misses and track fragmentation. Hence, it is often complemented by a hierarchical data association strategy centered on detection thresholds. Nonetheless, relying solely on threshold-based categorization may segregate detections and lead to redundant detections for the same target, thereby increasing the impact of redundancies on tracking and causing trajectory drift. To address this, this paper introduces a novel hierarchical data association framework based on network flow, integrating historical trajectory and domain information to efficiently group detections and redundancies. Additionally, an effective global topological structure graph is proposed to manage redundant detections. Experimental results demonstrate competitive performance on MOTChallenge, with a 1-2\% improvement in MOTA over the benchmark, particularly nearing 2% on MOT20.

With the advancement of video analysis technology, the multi-object tracking (MOT) problem in complex scenes involving pedestrians is gaining increasing importance. This challenge primarily involves two key tasks: pedestrian detection and re-identification. While significant progress has been achieved in pedestrian detection tasks in recent years, enhancing the effectiveness of re-identification tasks remains a persistent challenge. This difficulty arises from the large total number of pedestrian samples in multi-object tracking datasets and the scarcity of individual instance samples. Motivated by recent rapid advancements in meta-learning techniques, we introduce MAML MOT, a meta-learning-based training approach for multi-object tracking. This approach leverages the rapid learning capability of meta-learning to tackle the issue of sample scarcity in pedestrian re-identification tasks, aiming to improve the model's generalization performance and robustness. Experimental results demonstrate that the proposed method achieves high accuracy on mainstream datasets in the MOT Challenge. This offers new perspectives and solutions for research in the field of pedestrian multi-object tracking.

It is difficult to train a trustworthy transformer model on a small image classification dataset. This research proposes a sophisticated structured knowledge distillation algorithm that uses CNNs as Transformer's sophisticated teachers, significantly lowering the number of training datasets needed. To better to develop the potential for CNN tutors, this research configures a public data set for CNN teaching as an enlightenment textbook to guide Transformer's training and avoid falling into local optimization prematurely. The distillation process then employs a “learn-digest-self-distillation” learning strategy to enable the Transformer to assimilate CNN knowledge in a structured manner. Sufficient experiments show that the proposed method is significantly better than the direct training Transformer under the condition of limited data sets. Moreover, in order to show the practical application value, this research contributed a practical data set for the classification of smoking and calling. The corresponding code and dataset will be released at https://gitee.com/wustdch/surpass-teacher if this paper is accepted.

At present, license plate recognition algorithm under restricted conditions is relatively mature and widely used in various license plate recognition system. Due to the influence of factors such as large differences in shooting angles and vehicle motion blur, Chinese license plate recognition is quite challenging. In response to the above problems, this research abandoned the single end-to-end deep learning license plate recognition method, and proposed a step-by-step license plate recognition algorithm that integrated detection and classification, and utilized a level-by-level object detection strategy combined with character classification to predict the characters of the license plate result. On the basis of the above, a multi-anchor character position regression algorithm was proposed to further accurately regress the local area position information of all license plate characters. At the same time, in order to meet the needs of character detection and character classification, as well as the imbalance of the existing license plate datasets, this study contributed a series of supporting license plate datasets. According to the published publications, this study contributed the first large-scale character-level annotated license plate dataset. Extensive experiments show that the method in this study can reach the current state-of-the-art on different datasets. If accepted, the dataset will be publicly available at https://gitee.com/wust30405/lpdataset.

为解决判断指针方向的问题,提出一个旋转目标检测网络 R-YOLOv5.采用旋转矩形的形式定位指针的位置和角度方向,采用环形平滑标签的长边表示法区分矩形的方向,将角度回归预测问题转化为分类问题,设计合适的损失函数和旋转 IoU计算方法,解决角度周期性对训练的影响.针对指针仪表数字刻度识别问题,提出一种空间聚类算法,能够鲁棒地识别刻度的数字.实验结果表明,所提指针仪表读数识别方法在误差允许范围内,识别准确率达 93.9％.

针对现有的红外图像增强方法存在欠增强、过增强以及微小细节丢失等缺陷,提出了低频重分布与边缘增强的红外图像增强算法.用基于改进引导滤波的Retinex将红外图像分解为低频和高频图像.为了充分利用像素级的动态空间,对低频图像进行均匀重分布,以提升图像的亮度和清晰度;用提出的方向梯度算子对高频图像进行边缘提取,再对高频图像进行边缘增强,进一步提升图像的对比度.将经增强处理的低频和高频图像作Retinex反变换,得到增强效果的红外图像.实验结果显示,相对于部分现有方法,本文方法的增强图像的信息熵和Brenner指数更高,而NIQE指数更小,因此本文方法能更有效地提升红外图像的对比度,在增强图像纹理细节的同时更好地保持图像的自然度.

Classification tasks in 3D point clouds often assume that class events are independent and identically distributed (IID), although this assumption destroys the correlation between classes. This study proposes a classification strategy, Joint Graph Entropy Knowledge Distillation (JGEKD), suitable for non-independent and identically distributed 3D point cloud data, which achieves knowledge transfer of class correlations through knowledge distillation by constructing a loss function based on joint graph entropy. First, we employ joint graphs to capture the hidden relationships between classes and implement knowledge distillation to train our model by calculating the entropy of add graph. Subsequently, to handle 3D point clouds invariant to spatial transformations, we construct Siamese structures and develop two frameworks, self-knowledge distillation and teacher-knowledge distillation, to facilitate information transfer between different transformation forms of the same data. In addition, we use the above framework to achieve knowledge transfer between point clouds and their corrupted forms, and increase the robustness against corruption of model. Extensive experiments on ScanObject, ModelNet40, ScanntV2_cls and ModelNet-C demonstrate that the proposed strategy can achieve competitive results.

Many tasks generated by mobile user devices are computation-intensive and latency-sensitive, such as autonomous driving and video analysis. However, due to limited energy and computing capacity, a user device may not be able to complete its task within a given time, leading to a poor user experience. Mobile edge computing (MEC) can address this challenge by offloading tasks to edge servers with stronger computing capacity and more resources for execution, which can save energy of user devices and reduce the task computation time. Different offloading strategies will impact the number of tasks completed, latency, energy overhead and so on. This paper investigates the problem of maximizing the number of completed tasks while minimizing the average completion time, energy overhead and cost in MEC under both time and energy constraints. To solve the problem, we develop the mayfly genetic algorithm (MGA), which jointly optimizes task offloading locations and ratios, central processor unit (CPU) frequencies of user devices and computing capacities allocated to user devices by edge servers. Simulation experiments indicate that MGA outperforms state-of-the-art algorithms in terms of the number of completed tasks.

针对巡检机器人所拍摄的数字仪表图像中示数区域存在透视畸变的问题,提出一种基于八点回归的凸四边形示数区域拟合算法.通过增加多个控制点辅助示数区域顶点的回归,得到较准确的示数区域;提出一种结合控制点信息的双线性插值策略,将示数区域图像进行矫正;针对仪表数字间缺乏上下文信息的特点,在数码管数字识别阶段采用单阶段多尺度目标检测算法;提供一种数字仪表数据模拟生成的方法,解决固定场景下训练样本多样性差的问题.实验结果表明,所提方法在巡检机器人拍摄数据集上效果显著,有效提升了巡检机器人数字仪表识别的速度与精度.

The joint-detection-and-tracking framework shares network features of detection and re-identification (re-ID), which drives multi-object tracking (MOT) to be simple, fast and accurate. Most of the existing algorithms utilize global information of image to optimize re-ID feature, which result in weak representation of features and a large number of ID switches in the association phase. To solve the above problems, we present a loss function for re-ID task based on cosine space and angle space. Specifically, the algorithm maximizes the decision boundary distance between different categories in cosine space and angle space, respectively, so as to improve the compactness of feature distribution in the same category and expand the variability of feature between different categories. For the problem of small number of pedestrian samples with the same ID, a mixed data augmentation method is introduced based on statistical information of object location and motion distribution. Experimental results of MOT Challenge benchmark show that the proposed method obtains different degrees of improvement in MOTA and IDF1 scores compared with the baseline model, and the IDs metric decreases significantly, while it can cope with tracking scenarios of different complexity.