To address the distribution discrepancy between the source and target domains caused by camera-specific style variations, we propose a unsupervised dual-branch cross-domain person re-identification framework based on domain-invariant feature learning. Specifically, during the source domain pre-training phase, considering the distribution shift induced by inter-camera style differences, we treat each camera as an independent style domain. CycleGAN is employed to perform camera-style transfer, which significantly enhances the diversity of training samples and alleviates inter-domain distribution bias. To simultaneously capture fine-grained local details and high-level semantic context, we place the IBN and Non-local modules in Layer2 and Layer3 of the network. Additionally, a fixed exponent GeM pooling strategy is adopted to improve both the discriminability and generalizability of the learned features. During the target domain adaptation stage, in order to suppress the noise introduced by clustering-generated pseudo labels, a dual-branch symmetric architecture is constructed. An Exponential Moving Average model is maintained to generate soft pseudo labels. Using complementary supervision between hard and soft labels, our method effectively mitigates label noise and enhances robustness. Extensive experiments conducted on three widely used datasets (Market1501, DukeMTMC-reID, and MSMT17) demonstrate the effectiveness of the proposed method in both unsupervised domain adaptation and purely unsupervised person re-identification tasks.

To address the challenges faced by low-voltage series arc fault detection, such as difficulty in establishing criteria and high requirement for data consistency, an arc fault identification method based on arc period detection and half-wave symmetry verification is proposed. First, according to the UL1699 standard, an experiment platform was constructed to collect periodic current signals, which were used to build the training dataset. Multiperiod currents were collected to form the test dataset. Second, a fault discrimination strategy combining arc period detection and half-wave symmetry verification was developed. The variation of the currents were analyzed, and both local and global feature descriptors of the current signal were defined. A feature set of training samples was established to train the arc period detector. During the testing, arc period detection was performed on the test currents. Based on the detection results, a half-wave symmetry verification method was introduced to accurately count the number of arc fault half-waves, thereby enabling reliable identification of arc faults in the test currents. Experiment results show that the proposed method achieves an arc fault detection accuracy of $\mathbf{9 0. 5 8 \%}$, validating the effectiveness of the feature description and arc fault detection strategy.

Person re-identification is a challenging research issue in the field of intelligent monitoring. Affected by the complexity and variability of pedestrian images, it has the problems of difficulty in pedestrian feature extraction and low recognition accuracy in the case of a small number of samples. In order to tackle this problem, we propose a hybrid pooling fusion and composite metric method for few-shot person re-identification. Firstly, a hybrid pooling fusion method is proposed. This method focuses on pedestrian samples' local salient features and global statistical features by introducing a maximum pooling layer and an average pooling layer after each feature extraction layer. Moreover, the adaptive weight allocation mechanism is used to fuse the original features, ensuring that the model pays attention to the global and complements the key detailed features to achieve more representative pedestrian feature extraction. Secondly, a composite metric method is proposed, which combines Gaussian kernel metric with relation metric. This method uses the Gaussian kernel to solve the problem of linear non-separability of sample features. Furthermore, it uses relation metrics to focus on the relative relationship between samples, capture the differences in key features. Finally, the two similarity scores are weighted and fused to obtain the joint metric score, and the joint loss is constructed to optimize the recognition performance of the model. Experimental results show that the proposed method performs well in person re-identification with the few-shot environment, effectively overcoming the challenges brought by the complexity and variability of pedestrian images.

Person re-identification (person re-ID) is one of the important contents of joint intelligent analysis based on surveillance video, which plays an important role in maintaining social public safety. The key challenge of person re-ID is to address the problem of large intra-class variations among the same person and small inter-class variations between different persons. To solve this problem, we propose a Person Re-identification Network Based on Multi-feature Fusion to Enhance Pedestrian Features (MFEFNet). This network, through global, attribute, and local branches, leverages the complementary information between different levels of pedestrian features, thereby enhancing the accuracy of person re-ID. Firstly, this network leverages the stability of attribute features to reduce intra-class variations and the sensitivity of local features to increase inter-class differences. Secondly, a self-attention fusion module is proposed to address the issue of small receptive fields caused by residual structures, thereby enhancing the ability to extract global features. Thirdly, an attribute area weight module is proposed to address the issue that different pedestrian attributes focus on different person regions. By localizing regions related to attributes, it reduces information redundancy. Finally, this method achieved 95.63% Rank-1 accuracy and 88.29% mAP on Market-1501 dataset, 90.13% Rank-1 accuracy and 79.85% mAP on DukeMTMC-reID dataset and 77.21% Rank-1 accuracy and 60.34% mAP on Occluded-Market dataset.

Due to the limited capacity to capture low-probability features of abnormal objects in long-tail distributions, intelligent and efficient abnormal object detection still fails to achieve ideal performance in practical environments. Therefore, a lightweight method for abnormal object detection in transmission line corridors is proposed to address the inter-class distribution differences and intra-class attribute imbalance exhibited by abnormal objects. It integrates a class inherent feature extraction module (CIFEM) and a self-moving channel attention module (SMCAM), called LAOD-LTNet. Firstly, the CIFEM is proposed to capture intrinsic features, enhance the understanding of intra-class variations and inter-class similarities, and reduce the negative impacts on tail classes. Then, the feature representation is enhanced using the proposed SMCAM to balance the compactness and accuracy of the model. Finally, the experimental results on the transmission line corridor abnormal object dataset (TLCAOD) show that the proposed method can effectively maintain the coordination between lightweight architecture and high accuracy, providing innovative ideas for overcoming long tail data on transmission line corridors.

To address the challenges in fully expressing the fault features of low-voltage series arcs and the limitations of existing detection algorithms, this paper proposes a novel method combining periodic background subtraction and linear dividing lines for detecting arc faults. During feature extraction, a periodic background subtraction method is introduced, which calculates the difference between the periodic current signal and the average of the first four current periods nearest to the signal. This approach effectively suppresses interference from normal current variations and environmental noise, enabling a more robust expression of arc fault characteristics. For fault detection, a method based on linear dividing lines is developed to determine the optimal dividing line between fault and non-fault states for both linear and nonlinear load types via adaptive learning of the logistic regression model, achieving effective arc fault detection.

To address the challenge of accurately identifying tree-related high-impedance earth faults (THIEFs), a method based on long-term fluctuations in zero-sequence current is proposed. The analysis of the recorded data from staged tests in a real test network reveals the development patterns of THIEFs and their zero-sequence characteristics during the fault process. It was found that, over an extended duration, the fluctuation characteristics of the zero-sequence current RMS value curves for THIEFs differ significantly from those of the other types of high-impedance earth faults (HIEFs). By applying approximate arc length parameterization, the RMS value curve is standardized. The curvature standard deviation, the average curvature variation rate, and its standard deviation are used as feature parameters. An identification method based on an improved grey wolf optimization probabilistic neural network is constructed. Validation with staged test results demonstrates that the proposed method achieves a success rate of 97.5%, accurately distinguishing THIEFs from other types of HIEF.

When a series arc fault occurs in photovoltaic DC system, the existing research mainly focuses on arc fault current waveform analysis and signal processing, failing to reveal the essential reason of current features change caused by the arc fault. To address this issue, this paper conducts research on the modeling and feature analysis of photovoltaic DC series arc faults. Firstly, based on existing research of photovoltaic cells, an equivalent circuit of the photovoltaic system is established. Secondly, the arc is divided into arc initiation stage and arc burning stage. Based on the dynamic characteristics of plasma, the physical impedance model of arc fault is constructed in stages, and the DC series arc fault equivalent circuit of photovoltaic system is established. Then, the current features of the monitoring point are analyzed from the two stages of arc initiation and arc burning, which provides a theoretical basis for defining the arc fault current features. Finally, a real experimental platform is built, and the features of arc fault current are analyzed and verified by real experimental data, and the influence of different factors on the features is discussed. The conclusion is beneficial to the subsequent arc fault feature extraction and provides a new perspective for arc fault detection, which has important practical engineering significance.

In UAV-based power line inspection, vibration dampers present as small targets in images, with their appearance varying due to diverse drone perspectives. This variability complicates defect detection, primarily due to localization inaccuracies that drive high false positive rates. Traditional rule-based image processing offers interpretable detection by leveraging prior knowledge, whereas deep learning methods excel in performance through data-driven learning but lack interpretability. To integrate these strengths and tackle the small-target localization challenge, we propose the Prior-Knowledge-Guided and Model-Data-Driven Method (PGMDM). PGMDM employs classical line detection to identify regions containing vibration dampers, transforming these regions into spatial attention maps via distance transform. These maps guide a deep object detector to focus feature extraction and detection near power lines, effectively narrowing the search space. The network then outputs vibration damper categories and defect states. Experiments on a custom fine-grained vibration damper defect dataset show PGMDM achieves an mAP@0.5 of 83.3%, surpassing YOLOv8 by 1.5% improvement. The key innovation lies in embedding domain-specific prior knowledge into deep learning pipelines through spatial attention, enhancing robustness and accuracy for small-target defect detection in complex aerial inspection scenarios. This work advances reliable and efficient UAV-based autonomous power line maintenance systems.

To address the issue of strong randomness and the difficulty in accurately describing fault features of photovoltaic power generation system series arc, a photovoltaic DC series arc fault detection method based on two-stage feature comprehensive decision is proposed. Firstly, to solve the difficulty in selecting fault detection window size due to the non-periodicity and high randomness of DC signals, a signal windowing strategy based on autocorrelation function is proposed. Based on the transient characteristics of arc initiation stage and the steady-state characteristics of arc burning stage, the whole arc stage is divided into transient stage and steady-state stage. Then, in the arc initiation stage, a transient feature description method based on adjacent windows difference (AWD) is designed on the basis of signal windowing, effectively capturing the waveform mutation caused by arc, achieving the fault occurrence window positioning and the effective expression of transient feature. In the arc burning stage, a steady-state feature description method based on energy difference (ED) is designed on the basis of signal windowing and fault occurrence window positioning, effectively capturing the energy difference caused by arc, achieving a significant expression of steady-state feature, and overcoming the misjudgment issues caused by transient feature. Finally, SVMs are used to classify the proposed features, and voting decision is combined to obtain the arc fault detection results. Experimental results show that the proposed method is feasible and effective in the feature extraction and detection of arc fault, providing a valuable approach for photovoltaic DC series arc fault detection.

With the rapid development of electronic payment technologies, facial recognition-based payment systems have become increasingly popular and indispensable. However, the majority of facial recognition payment systems are vulnerable to being manipulated by facial deepfake technology, and it would be a serious threat to personal property and privacy. In order to effectively defend deepfake models on the premise of minimizing alterations to the original image, we propose a union-saliency attack model which is a well-trained deepfake model while maintaining plausible detail of the original face images. To this aim, we derive a union mask mechanism to accurately determine facial region as a prior in guiding the subsequent perturbations, with the objective of minimizing the information loss on input images. Additionally, we propose a novel structural similarity loss and a noise generator to minimize detail degradation. Experiments prove that the proposed method can interfere with deepfake models effectively and minimize the distortion of the original image simultaneously.

Insulator defects are the critical concern in intelligent inspection systems for transmission lines, typically manifesting in the forms of self-explosion, corrosion, and fouling. However, weak feature defects such as corrosion and fouling, often exhibit limited spatial extent, complex morphology, and mutual interference. Due to the limitations of the optical image acquisition sensor on the unmanned aerial vehicle (UAV), these factors can lead to unclear feature representation, large intraclass variation, and significant interclass similarity, which increase the risks of misclassification and boundary confusion in defect detection. To realize accurate detection of weak feature defects in insulators, this article proposes an improved DeepLabv3+ network fused with a boundary enhancement module (BEM) and interclass feature attention module (I-CFAM), called BEIC-DeepLabv3+. First, the BEM is proposed to enrich shallow detail information and enhance the model's ability to understand the shape and structure of the target. Then, interclass dependencies are captured using the proposed I-CFAM to extract the semantic information of different defects more effectively. Finally, the encoder-decoder structure is used to effectively combine the shallow details and deep semantic information, enabling comprehensive feature extraction and analysis. The experimental results on the insulator defect dataset (IDD) show that the proposed method can effectively alleviate issues of misclassification and boundary confusion in weak feature defect detection of insulators, providing innovative ideas for the development of defect detection technology.

Computer vision plays a pivotal role in modern society, which transforms fields such as healthcare, transportation, entertainment, and manufacturing by enabling machines to interpret and understand visual information, revolutionizing industries, and enhancing daily life [...]

Recent years have witnessed a remarkable proliferation of applications in smart cities. Crowd analysis is a crucial subject, and it incorporates two subtasks in smart city systems, i.e. , crowd counting and crowd localization. Nevertheless, the presence of adverse intrinsic factors, i.e. , scale variation and background noise severely degrades the performance of counting and localization. Although great efforts have been made on separate research on counting and localization, few works are capable of performing both tasks at the same time. To this aim, the scale attentive aggregation network (SA 2 Net) is proposed to solve the problems of scale variation and background noise in crowd counting and localization tasks synchronously. Specifically, the SA 2 Net has two vital modules, namely multiscale feature aggregator (MFA) module and background noise suppressor (BNS) module. The MFA module is designed in a four-pathway structure, and it aggregates the multiscale feature so as to facilitate the correlation between different scales. The BNS module utilizes the contextual information between the input keys matrix and self-attention matrix to suppress the background noise. Furthermore, a global consistency loss combined with the Euclidean loss is utilized to optimize the network in counting and localization tasks. Extensive experimental results prove that the SA 2 Net outperforms the state-of-the-art competitors both subjectively and objectively.

The detection of multi-feature fusion is a crucial approach to address the issue of series arc fault detection. Effective feature selection plays a vital role in enhancing the accuracy of the classifier and reducing system complexity. In this study, a feature selection algorithm based on Fisher-mutual information is proposed to tackle the problem of feature selection in multi-feature fusion detection. This algorithm utilizes the characteristics of arc fault voltage source to construct a feature pool. The Fisher-score algorithm and mutual information algorithm are employed to construct an optimal feature subset. The feature subset undergoes rough selection by retaining key features of the classifier and fine selection by eliminating redundant features. Experimental results and comparisons with related methods demonstrate that the proposed feature selection method significantly enhances the classifier's recognition accuracy, reduces classification and recognition time, diminishes the feature dimension, and outperforms other existing methods.

Unsupervised person re-identification(person Re-ID) based on traditional asymmetric metric learning faces the problems of data distribution differences and feature nonlinearity caused by uncontrolled collection of pedestrian images. To solve these problems, an unsupervised person re-identification method based on asymmetric metric learning with kernel distribution constraints is proposed. Firstly, the data distribution constraint is introduced into asymmetric metric learning, and the mean difference between samples is calculated by using the maximum mean difference term. This metric method effectively overcomes the problem of distribution differences in different sample data features caused by scene changes. Secondly, the Gaussian kernel function is introduced into asymmetric metric learning based on data distribution constraints, and pedestrian samples are mapped to high-dimensional space through nonlinear mapping. This method overcomes the problem of linear inseparability of pedestrian features, while avoiding the high dimensionality of features and reducing computational complexity. Finally, the optimal metric matrix is obtained by solving the generalized eigenvalue problem. Extensive experiments are conducted on five datasets: VIPeR, PRID450S, CHUK01, Market-1501, and DukeMTMC-reID. The experimental results of comparison with other methods show that our method achieves competitive performance.

Object counting, defined as the task of accurately predicting the number of objects in static images or videos, has recently attracted considerable interest. However, the unavoidable presence of background noise prevents counting performance from advancing further. To address this issue, we created a group and graph attention network (GGANet) for dense object counting. GGANet is an encoder-decoder architecture incorporating a group channel attention (GCA) module and a learnable graph attention (LGA) module. The GCA module groups the feature map into several subfeatures, each of which is assigned an attention factor through the identical channel attention. The LGA module views the feature map as a graph structure in which the different channels represent diverse feature vertices, and the responses between channels represent edges. The GCA and LGA modules jointly avoid the interference of irrelevant pixels and suppress the background noise. Experiments are conducted on four crowd-counting datasets, two vehicle-counting datasets, one remote-sensing counting dataset, and one few-shot object-counting dataset. Comparative results prove that the proposed abbr achieves superior counting performance.