Surface mount technology (SMT) is a core process in consumer electronics manufacturing, and its soldering quality directly affects product performance, stability, and service life. In post-reflow automated optical inspection (AOI) scenarios, complex and varying backgrounds, imbalanced defect category distributions, and structural interference caused by unmounted regions pose significant challenges to the accuracy of industrial online inspection. To address these issues, this paper proposes a structurally decoupled multi-stage cascaded solder joint defect detection framework for consumer electronics SMT production lines. The framework decomposes the detection task into three stages, namely feature-region segmentation, anomaly localization, and few-shot defect classification, so as to separately handle three subtasks with different statistical characteristics: structural recognition, anomaly screening, and category discrimination. Specifically, a lightweight context-aware semantic segmentation network is first constructed to accurately extract key structural regions. Then, image registration and multi-feature template matching are combined to achieve precise localization of candidate anomalous regions. Finally, a few-shot classification network that integrates physical information and metric learning is designed to improve the discrimination of different defect categories under limited defect samples. Experimental results show that the proposed method achieves a classification accuracy of 96.20%, a defect error rate of 0.14%, and an average inference time of 0.25 s per board. These results demonstrate that the proposed method can effectively reduce false alarms while satisfying industrial real-time requirements, and it exhibits strong stability, interpretability, and engineering deployment potential.

Within the surface mount assembly line, the placement process parameters of the mounter significantly influence assembly quality. This paper decomposes the placement process parameter optimization problem into two distinct subproblems: the displacement prediction problem (DPP) and the process parameter adjustment problem (PPAP). A causal deep displacement prediction model is proposed to address DPP, which is a critical factor in optimizing placement parameters to enhance surface mount technology assembly quality. Then, a deep reinforcement learning algorithm is introduced to solve PPAP. Given the multitude of factors affecting placement accuracy and their complex interactions, a model training methodology incorporating a causal inference-based attention mechanism and simulated interventional operations is further proposed. This methodology drives the network to capture underlying causal relationships rather than mere correlations. Experimental results demonstrate that the proposed prediction model significantly outperforms other state-of-the-art algorithms, while the employed parameter optimization method also surpasses alternative placement methods.

Generalist industrial anomaly detection aims to train a unified model on auxiliary domains and directly apply it to anomalous defect detection in unseen target domains. Existing methods mainly rely on patch-level residual features between query samples and normal support samples to improve model generalization. However, such residual features are essentially derived from local correspondences and, therefore, lack sufficient contextual information. On the other hand, directly introducing contextual information may interfere with the original anomaly representation and consequently cause response diffusion. To address this issue, this work establishes a connection between the residual representation process and the PID controller, and explicitly decomposes anomaly residual representations into proportional, integral, and derivative components. Specifically, the proportional branch adopts the original residual feature to preserve the most direct anomalous deviation signal. The integral branch models the low-frequency context of the query feature to provide stable scene-level constraints for local residuals. The derivative branch captures the local variation trend of the residual through high-frequency residual details, thereby suppressing the overshoot effect caused by contextual injection. On this basis, a controlled PID fusion mechanism is designed, in which residuals serve as the dominant cue, contextual information acts as modulation, and high-frequency residuals provide controlled correction, enabling adaptive regulation of both anomaly response intensity and spatial extent. Extensive experiments on multiple public datasets and a real-world printed circuit board defect dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches in cross-domain generalist anomaly detection tasks.

Automatic optical inspection (AOI) systems are crucial for quality control in surface mount technology (SMT). Their detection accuracy directly affects the yield and reliability of high-density printed circuit boards (PCBs). However, existing methods still face typical industrial challenges, such as high false negative rates and inconsistent responses to similar components when detecting miniature elements. This article addresses the specific requirements for PCB component detection in industrial AOI, particularly in scenarios involving small targets and densely parallel distributions. We propose an adaptive integration and optimization of existing frequency-domain and coordinate-aware methods, introducing a frequency-coordinate collaborative network (FC-YOLO). The model features a dual-branch detection framework, which processes image information in parallel across two dimensions. The coordinate branch uses the coordinate prior information embedding (CPIE) mechanism for high-frequency coordinate encoding, enhancing the model's spatial invariance and fine detail perception, thereby improving the consistency of responses for parallel components. The frequency-domain branch introduces frequency-domain features and designs the Fourier pixel difference convolution (FPDConv) module, employing learnable differential convolution kernels to adaptively suppress frequency-domain noise and amplify key frequency components, thus improving detection performance for small targets. The features from both branches are efficiently fused via the frequency-coordinate feature fusion (FCFF) module, and their synergistic effect enhances the detection accuracy and response consistency of miniature components. Experimental results on the public PCBWACV dataset and the self-constructed EAPCB dataset show that the proposed method achieves a 4.5% improvement in mAP50 and a 4.1% improvement in mAP50-95 compared to YOLO11. After deployment on an enhanced automated optical inspection (EAOI) system with an RTX 4060Ti GPU, the system achieves real-time detection at 206 FPS. It outperforms existing methods in both false positive and false negative rates, showcasing its potential for engineering applications.

Long context large language models (LLMs) enable million-token understanding and generation, but impose substantial compute and memory costs due to large attention projection parameters and substantial key-value (KV) cache memory. Existing pruning and uniform low rank factorization methods reduce compute and memory use, but often degrade long context accuracy or limit inference efficiency owing to overcompression or undercompression. We propose SEDPA, a singular value decomposition (SVD)–enhanced dual path attention with two variants: SEDPA-S applies energy-aware layerwise SVD to reduce the stored parameter footprint with near-zero quality loss; SEDPA-W augments SEDPA-S with a sliding window to significantly reduce latency at a small accuracy cost. SEDPA supports task dependent tradeoffs between accuracy and latency. On Llama-2-7B-32K and Llama-3-8B-1048K, SEDPA-S reduces parameters by up to 15% while keeping accuracy on the Needle-in-a-Haystack benchmark within 1.1% of the state-of-the-art DuoAttention, and is competitive on LongBench, surpassing DuoAt-tention on 10 tasks. At a small accuracy cost, SEDPA-W reduces latency, KV cache memory, and peak memory by up to 25%, 15%, and 20%, respectively.

Semantic segmentation takes a pivotal role in various applications such as autonomous driving and medical image analysis. When deploying segmentation models in practice, it is critical to test their behaviors in varied and complex scenes in advance. In this paper, we construct an automatic data generation pipeline Gen4Seg to stress-test semantic segmentation models by generating various challenging samples with different attribute changes. Beyond previous evaluation paradigms focusing solely on global weather and style transfer, we investigate variations in both appearance and geometry attributes at the object and image level. These include object color, material, size, and position, as well as image-level variations such as weather and style. To achieve this, we propose to edit visual attributes of existing real images with precise control of structural information, empowered by diffusion models. In this way, the existing segmentation labels can be reused for the edited images, which greatly reduces the labor costs of constructing datasets. Using our pipeline, we construct two new benchmarks, Pascal-EA and COCO-EA. We benchmark a broad variety of semantic segmentation models, spanning from conventional close-set models to recent open-vocabulary large models. We have several key findings: 1) advanced open-vocabulary models do not exhibit greater robustness compared to closed-set methods under geometric variations; 2) traditional data augmentation techniques, such as CutOut and CutMix, are limited in enhancing robustness against appearance variations; 3) our generation pipeline can also be employed as a data augmentation tool and improve both in-distribution and out-of-distribution performances. Our work suggests the potential of generative models as effective tools for automatically analyzing segmentation models, and we hope our findings will assist practitioners and researchers in developing more robust and reliable segmentation models.

Object tracking in multi-drone videos has attracted increasing attention for its ability to provide richer target information, particularly in scenarios involving occlusion or background interference. Most existing trackers adopt a template sharing-based framework, using cross-drone target templates for similarity estimation with intra-drone search regions to predict target responses. However, such framework relies solely on templates to spread cross-drone information, neglecting current temporal and background information. Furthermore, similarity-based multi-drone interaction is highly sensitive to cross-view variations. To address above issues, this paper proposes TSMamba, a novel tracking framework that utilizes consistency and complementary information among cross-drone templates and search features to enhance the responses of challenging samples. Specially, it incorporates a temporal consistency enhance mamba (TCEM) to exploit temporal consistency between target templates and search regions to improve target distinguishability in complex scenarios. Meanwhile, a spatial complementary fusion mamba (SCFM) is integrated to capture spatial complementary from different visible components observed under varying viewpoints to build comprehensive target representations. Additionally, a localization-aware response map correction (LRMC) is introduced to refine target response maps adaptively, improving robustness under dynamic conditions compared to existing predefined methods. To our knowledge, this is the first work explicitly enhance search features using cross-drone information and the first to employ vision mamba for multi-drone feature interaction. Experiments on the MDOT and MDMT benchmarks demonstrate that TSMamba outperforms previous state-of-the-art trackers by 5.0% and 4.0% in accuracy and precision. Moreover, TSMamba exhibits competitive computational efficiency and strong generalization in real-world applications. Codes and experimental data will be released at https://github.com/HanJin0326/TSMamba.

Effective knowledge transfer across interrelated tasks remains a core challenge in multi-objective evolutionary multitask optimization, especially when task similarity is heterogeneous across regions of the search space. This paper proposes the Similarity-aware, Multifactorial, Subpopulation-based Evolutionary Algorithm (SMS-MFEA), which extends the MFEA framework with a similarity-driven transfer mechanism at the subpopulation level. By employing K-means clustering to partition populations into subpopulations, SMS-MFEA quantifies inter-task similarity through maximum mean discrepancy to identify transferable knowledge domains. The strategy incorporates an adaptive three-strategy selection mechanism that determines optimal knowledge transfer approaches based on quantified task similarities. This integration enhances matching efficiency, improves shared knowledge utilization, and accelerates population evolution. The strategy ensures effective knowledge sharing for strongly correlated tasks while mitigating negative transfer between dissimilar tasks. Experimental validation demonstrates the method’s efficacy, with the algorithm achieving optimal inverted generational distance values in 12 cases and optimal hypervolume values in 13 instances across 18 evaluations from nine established multi-objective multitasking benchmark functions, thereby confirming its operational efficiency.

Mobile manipulators hold considerable promise for automating warehouse operations, especially in item picking. However, traditional approaches that alternate between mobile base movement and manipulator operation often suffer from inefficiencies. This inefficiency arises from their discontinuous operational patterns and the resulting suboptimal coordination between locomotion and manipulation. To overcome this limitation and maintain continuous operation, a multimodal hierarchical synergistic strategy is developed. This strategy establishes a dynamic coupling between manipulator gripping actions and mobile base movement, enabling simultaneous operation and improving overall task efficiency. Specifically, the proposed strategy employs the ELD-RRT* algorithm for smooth trajectory planning, guiding the mobile base to both grasping and placement points. To enhance path planning efficiency, the algorithm incorporates ELOA-Sampling, which strategically reduces sampling in unproductive areas. Furthermore, a dynamic expansion step-size strategy is implemented to accommodate specific requirements within the grasping area, ensuring that generated paths meet the manipulator’s grasping stability criteria. To further refine performance, a position-based visual servoing controller is integrated to improve grasping accuracy. Simulation and experimental results demonstrate that, compared to conventional static methods, this mobile grasping strategy enhances task efficiency by 2.5–9.5

Context: Function-level vulnerability detection is important for software security and code review, but existing learning-based methods often struggle to jointly capture semantic cues and intra-function dependency relations, especially when evidence is distributed across non-adjacent paths or heterogeneous control/data-flow structures. Objective: This study examines whether vulnerability detection can be improved by jointly modeling local dependency evidence, complementary global structural relations, and their alignment with code semantics under the function-level setting. Methods: We propose DualGraphVulD, a dual-view semantic–structural framework. The local view uses gated message passing for fine-grained dependency interactions, while the global view applies restarted diffusion to complement local aggregation during cross-path dependency propagation. A query-guided cross-view alignment module retrieves structure-relevant evidence through cross-attention. We further use Progressive Enhanced Loss as an imbalance-aware loss scheduling strategy without graph-only mutation, preserving consistency between source text and CPG inputs. The evaluation includes graph-size buckets, dependency-span buckets, adaptive gate analysis, Semantic Only behavior, and PrimeVul under realistic splits. Results: Experiments on existing vulnerability benchmarks and PrimeVul show that DualGraphVulD improves overall detection performance under the evaluated settings. Diagnostic analyses indicate that conventional full-graph message passing degrades more on structurally complex or long-span samples, whereas DualGraphVulD maintains more stable performance by combining local aggregation, restarted global diffusion, and semantic–structural alignment. Conclusion: Dual-view dependency modeling and semantic–structural alignment are useful for structurally heterogeneous functions. This study focuses on function-level detection and does not claim complete coverage of inter-procedural vulnerabilities.

Robust target tracking with non-stationary heavy-tailed measurement noise caused by fluctuating radar cross-section (RCS) remains a significant challenge. This letter proposes a novel robust Kalman filter that jointly estimates the target state and its RCS by fusing kinematic measurements and amplitude information. First, we propose a novel Normal-inverse-Wishart-Chi-square-inverse-Gamma distribution to model the non-stationary heavy-tailed measurement noise that arises from RCS fluctuations. Second, the Rician amplitude likelihood is simplified via a Gaussian approximation. Third, the tracking problem is formulated within a variational Bayesian framework to approximate the analytically intractable joint posterior distribution. Simulation results demonstrate that the proposed method significantly enhances tracking accuracy and robustness compared to conventional robust filters.

Rapid object or camera motion during tracking can lead to motion blur, which ultimately reduces tracking accuracy. Motion blur degrades image features, affecting both feature matching and fusion. To alleviate the above problem, FIST is proposed. STARK is used as the baseline network, with dynamic templates integrated into its architecture. The Fourier transform and a multi-scale dilated attention block are incorporated into the network. This effectively mitigates the adverse impacts of motion blur on tracking performance. In the feature extraction stage, feature maps are transformed into frequency domain for filtering operations. In the feature fusion stage, feature maps are supplemented via frequency filtering and multi-scale dilated attention mechanism. Comprehensive ablation studies and comparative experiments are conducted to validate the effectiveness of the proposed method. Both AUC and Pre were improved by 1.5% in the LaSOT test dataset Motion Blur property compared to the baseline network.

Quality and efficiency objectives in surface mount technology (SMT) production are inherently conflicting, posing a significant challenge for collaborative optimization. Dynamic factors, such as process parameter anomalies and mechanical failures, substantially increase the complexity of SMT production optimization. However, existing research typically addresses efficiency or quality objectives separately. This article integrates automatic optical inspection (AOI) feedback into the surface mounter, introducing closed-loop quality optimization on top of the original efficiency optimization to enhance overall production performance dynamically. A multi-objective mixed integer programming model is first formulated to enable dynamic head workload assignment for small-scale problems. For large-scale instances, offline optimization generates Pareto-optimal solutions that balance SMT quality and efficiency for practical production planning by utilizing an improved MOEA/D algorithm with decision tree-based operator selection. This algorithm integrates component clustering encoding, heuristic decoding, and adaptive operator selection. Online dynamic optimization is subsequently performed in response to AOI inspection feedback. A dynamic head workload adjustment heuristic is utilized to maximize efficiency while ensuring that quality metrics are met. These two parts constitute the machine learning-assisted dynamic multi-objective optimization (ML-DMO) framework. Experimental validation demonstrates that ML-DMO achieves competitive performance against state-of-the-art algorithms in both multi-objective optimization convergence and practical production efficiency, while enabling dynamic workload adjustment in response to real-time quality feedback.

Matrix placement machines improve production efficiency of printed circuit board assembly (PCBA), addressing critical needs for flexible and intelligent electronics manufacturing. However, their complex head structure renders solutions for traditional beam-head placement machines inefficient for matrix placement machines. This article proposes a phased hybrid algorithm with adaptive hyper-nondominated sorting genetic algorithm II (NSGA-II) for PCBA optimization. A bidirectional search mechanism is applied to derive feeder distributions and nozzle configurations, and iteratively tighten the solution space using priority-based search strategies. The softmax, max greatest common divisor, and max matching mechanisms are proposed for placement and pickup sequences, which facilitates construction of solution pools. Initial solutions are extracted from the pool and, subsequently, hyperheuristic mechanisms dynamically adjust genetic operators within NSGA-II to minimize placement, pickup, and recognition times with better convergence speed. Experimental validation with real-world production data demonstrates that the proposed algorithm achieves 6.05%-38.18% performance improvements compared to state-of-the-art solutions.

Synthetic Aperture Radar (SAR) is widely used in military and civilian applications, with intelligent target interpretation of SAR images being a crucial component of SAR applications. Vision-Language Models (VLMs) play an important role in SAR target interpretation. By incorporating natural language understanding, VLMs effectively address the challenges posed by large intraclass variability in target characteristics and the scarcity of high-quality labeled samples, thereby advancing the field from purely visual interpretation toward semantic understanding of targets. Drawing upon our team’s extensive research experience in SAR target interpretation theory, algorithms, and applications, this paper provides a comprehensive review of intelligent SAR target interpretation based on VLMs. We provide an in-depth analysis of existing challenges and tasks, summarize the current state of research, and compile available open-source datasets. Furthermore, we systematically outline the evolution, ranging from task-specific VLMs to contrastive-, conversational-, and generative-based VLMs and foundational models. Finally, we discuss the latest challenges and future outlooks in SAR target interpretation by VLMs.

The multi-input multi-output dendritic neuron model, as an artificial neural network, has achieved remarkable results in handling nonlinear classification problems. However, it not only requires training a filtering matrix for pruning redundant dendrites but also needs to optimize structural parameters. Therefore, Multi-output dendritic neuron model is designed which retains the original soma layer from the dendritic neuron mode and discards the filtering matrix, enabling holistic multi-objective optimization of the model. To address insufficient population diversity and overfitting tendency in model optimization, a dual-criterion dynamic migration multi-objective evolutionary algorithm is proposed. In this algorithm, the original population is divided into three subpopulations, and computational resources are allocated based on two optimization criteria. Then the penalty mechanism and migration strategy are employed to enhance population diversity and exchange elite individuals across subpopulations. The introduction of the multi-output dendritic neuron model architecture converts the multi-output dendritic model structural search as a large-scale multi-objective optimization problem, and the dual-criterion dynamic migration multi-objective evolutionary algorithm is specifically designed to address Experimental results on 11 public datasets demonstrate that the proposed algorithm achieves improvements of 27.48 % in HV, 26.33 % in IGD, and 6.9 % in classification accuracy.

Adversarial purification with diffusion models has emerged as a promising defense strategy. While existing frequency-based approaches better preserve global semantics, they overlook the distinct temporal dynamics of low-and high-frequency components during the reverse process and lack adaptive correction mechanisms. Meanwhile, Diffusion Transformers (DiTs) offer stronger multi-condition alignment and superior scalability, but remain underexplored for adversarial purification, primarily due to prohibitive inference cost. Motivated by these challenges, we propose FEAP-DiT, a frequency-decoupled efficient adversarial purification framework built upon DiTs. FEAP-DiT designs complementary processing schemes for low-and high-frequency features and unifies them through frequency-aware decoupling and fusion. Specifically, low-frequency features are sparsely updated and largely reused with mild selective correction to preserve global semantics. High-frequency features, where adversarial perturbations concentrate, are updated more frequently to suppress perturbations and approximated at intermediate steps via a Hermite-based sequential predictor due to their short-term temporal continuity. To further mitigate prediction deviations in high-frequency components, FEAP-DiT introduces an efficient control-aware refinement mechanism that fuses structural constraints into the DiT backbone. Control blocks are inserted only at selective layers, and control computation is restricted to high-energy tokens over informative regions. Extensive experiments on CIFAR-10 and ImageNet-1K demonstrate the effectiveness of FEAP-DiT, achieving the largest average gains of 2.43% on CIFAR-10 and 3.40% on ImageNet-1K under white-box attacks. Moreover, FEAP-DiT reduces inference latency compared with existing multi-step diffusion-based adversarial purification baselines, while achieving a 15.61% reduction in computational complexity relative to the representative DiT architecture PixArt-alpha.

In Synthetic Aperture Radar (SAR) applications, the continuous emergence of new target classes poses a significant challenge to Automatic Target Recognition (ATR) systems. Adapting to the distribution of new data can induce drastic alterations in the feature space of deep models, resulting in a decline in their ability to recognize old data, termed catastrophic forgetting. To address this challenge, we propose a novel class-incremental SAR ATR method based on Physical Attributes Embedded Prototypical Network (PAEPN). PAEPN embeds physical attributes derived from electromagnetic scattering and geometric priors into the deep model to achieve stable representations. These physical attributes, determined by the target's shape, structure, and material composition, remain invariant throughout the incremental learning process, thereby enhancing the stability and interpretability of deep models. Specifically, PAEPN first extracts and integrates physical attribute priors to establish feature anchors, guiding the deep model in extracting physically consistent features and preventing drastic changes in the feature space. Second, a spatial attention enhancement strategy is introduced to enable the deep model to reliably focus on the key regions of SAR targets. Finally, feature relations that represent semantic similarity are distilled to further mitigate catastrophic forgetting. During testing, PAEPN employs the cosine distance between the sample feature and class prototypes for recognition. Comprehensive experiments on three datasets demonstrate that PAEPN outperforms existing state-of-the-art methods.