Electroencephalogram (EEG)-based emotion recognition holds significant value in fields such as human-computer interaction. However the domain shift problem caused by individual and session differences severely restricts the model’s generalization capability. To address these challenges, we propose a novel method termed AE-MSRN-CADDPL, which integrates subdomain adaptive feature enhancement with a confidence-adjudicated pseudo-labeling strategy for cross-subject and cross-session EEG emotion recognition. In our method, firstly, the Attention-Enhanced Multi-Scale Residual Network (AE-MSRN) is presented as a hierarchical feature representation module. It utilizes attentional multi-scale convolutional blocks to capture rich local patterns, integrated with a multi-head self-attention mechanism to capture global feature dependencies, thereby generating robust domain-adaptive representations from EEG sequences. Secondly, a novel Confidence-Adjudicated Dual-Decision Pseudo-Labeling (CADDPL) strategy is proposed. This strategy collaboratively employs the main classifier and a K-Nearest Neighbors (KNN) classifier to select pseudo-labels for target samples. This process yields a larger and more reliable set of pseudo-labels, effectively mitigating confirmation bias and suppressing the cumulative effects of noisy labels. Finally, a joint optimization mechanism is employed by integrating source domain classification loss, Local Maximum Mean Discrepancy (LMMD)-based subdomain adaptation loss, and a joint triplet contrastive loss. This enhances feature discriminability and aligns conditional distributions across domains. Extensive experiments on the SEED and SEED-IV datasets demonstrate that the proposed method outperforms relevant state-of-the-art methods in cross-subject tasks and achieves competitive performance in cross-session scenarios.

The development of accurate and robust pest detection is a crucial step toward reliable forecasting of agricultural pests in precision agriculture and has gained significant attention in many countries. Compared to conventional object detection, pest detection faces greater challenges in recognizing fine details. This is because pests often have small and intricate identifying features, such as antennae, legs, or specific body markings, which may be easily lost during the feature extraction process. Furthermore, pest detection must deal with more complex environmental information, which includes plant species, different parts of plants, and growth stages of plants. Therefore, we propose an end-to-end pest detection method using feature representation compensation (FRC) and regional pale-shaped self-attention (RPSA). In this method, we designed an FRC module based on CSWin, which can compensate for the loss of feature information caused by the down-sampling process. This module can significantly improve the performance of pest detection models without increasing the computational load. To further improve the performance, a transformer encoder based on RPSA is proposed to capture global information and enhance local information of the feature map, which can organically combine the global and local features of the image and achieve more comprehensive feature representations. Finally, extensive experiments are conducted on IP102 and FPD to validate the effectiveness of our method. The mAP scores achieved by our method surpass that of the state-of-the-art CNN-based methods by 4.5% and 5.7% on the IP102 and FPD, respectively. The source code and pre-trained models are available at https://github.com/LionelLinH/FRC-and-RPSA-for-pest-detection.

To mitigate the shortage of labeled data, Few-Shot Classification (FSC) methods train deep neural networks (DNNs) on a base dataset with sufficient labeled data, and then adapt them to target tasks using a few labeled data. Despite notable progress, a single FSC model remains prone to high variance and low confidence. As a result, ensemble FSC has garnered increasing attention. However, the limited labeled data and the high computational cost associated with DNNs present significant challenges for ensemble FSC methods. This paper presents a novel ensemble method that generates multiple FSC models via combining various training and adaptation algorithms. Due to the reuse of training phases, the proposed method significantly reduces the learning cost while generating base models with greater diversity. To further minimize reliance on labeled data, we provide each model with pseudo-labeled data selected by the majority vote of other models. Compared with self-training style methods, this “one-vs-others” learning strategy effectively reduces pseudo-label noise and confirmation bias. Finally, we conduct extensive experiments on miniImageNet, tieredImageNet and CUB datasets. The experimental results demonstrate that our method outperforms other state-of-the-art FSC methods. Especially, our method achieves the greatest improvement in the performance of base models. The source code and related models are available at https://github.com/tn1999tn/Ensemble-FSC/tree/master.

Semi-supervised sound event detection (SSED) tasks typically leverage a large amount of unlabeled and synthetic data to facilitate model generalization during training, reducing overfitting on a limited set of labeled data. However, the generalization training process often encounters challenges from noisy interference introduced by pseudo-labels or domain knowledge gaps. To alleviate noisy interference in class distribution learning, we propose an efficient semi-supervised class distribution learning method through dynamic prompt tuning, named prompting class distribution optimization (PADO). Specifically, when modeling real labeled data, PADO dynamically incorporates independent learnable prompt tokens to explore prior knowledge about the true distribution. Then, the prior knowledge serves as prompt information, dynamically interacting with the posterior noisy-class distribution information. In this case, PADO achieves class distribution optimization while maintaining model generalization, leading to a significant improvement in the efficiency of class distribution learning. Compared with state-of-the-art methods on the SSED datasets from DCASE 2019, 2020, and 2021 challenges, PADO achieves significant performance improvements. Furthermore, it is readily extendable to other benchmark models.

The edge- and structure-preserving smoothing is one of the most vital topics in the field of multimedia signal processing as it can be applicable to a variety of tasks. In this paper, to cater to the wide range of applications, we propose a optimization model based on a generalized nonconvex gradient regularization for edge-preserving image smoothing. The proposed model is highly flexible, thus facilitating varying smoothing behaviors. Furthermore, we extend our model to suit the task of structure-preserving filtering by proposing a nonlocal data term. The nonlocal design incorporates the spatial scale support and thus can be effective for the separation of the scale-dependent textures and structures. We propose an efficient solution based on the additive half-quadratic minimization and Fourier domain optimization so that the edge- and structure-preserving filtering can be facilitated in a unified framework. We have experimented with the proposed filter on a number of low-level vision tasks. Both quantitative and qualitative experiment results indicate the superiority of the proposed filter over the state-of-the-art filters. Moreover, the proposed filter is highly efficient; it is able to process 720P color images in real time (more than 20 fps) on an NVIDIA RTX 3070 graphics processing unit (GPU). Therefore, it can be practical for real applications. (c) 2025 SPIE and IS&T

In recent years, Vision Transformer-based methods have emerged as promising approaches for localizing semantic objects in weakly supervised semantic segmentation tasks. However, existing methods primarily rely on the attention mechanism to establish relations between classes and image patches, often neglecting the intrinsic interrelations among tokens within datasets. To address this gap, we propose the Inter-image Token Relation Learning (ITRL) framework, which advances weakly supervised semantic segmentation by inter-image consistency. Specifically, the Inter-image Class Token Contrast method is introduced to generate comprehensive class representations by contrasting class tokens in a memory bank manner. Additionally, the Inter-image Patch Token Align approach is presented, which enhances the normalized mutual information among patch tokens, thereby strengthening their interdependencies. Extensive experiments validated the proposed framework, showcasing competitive mean Intersection over Union scores on the PASCAL VOC 2012 and MS COCO 2014 datasets.

In the field of Infrared-Visible Image Fusion (IVIF), the preservation of details, edges, and texture is crucial for generating high-quality fused images. However, a major challenge arises due to the inevitable loss of high-frequency information during feature extraction, resulting in fused images that lack significant details. In this paper, we propose a dual-branch auto-encoder by exploiting an invertible high-frequency branch for detailed feature preservation and a transformer-based low-frequency branch for global dependencies modeling. First, the high-frequency branch employs the wavelet transforms and an Invertible Neural Networks (INN)-based encoder to model high-frequency features through an invertible transformation, including a forward process for image fusion and an inverse process for original image reconstruction. Additionally, a high-frequency loss is designed to enhance the high-frequency feature representation for high-quality image fusion. Second, a low-frequency branch based on a transformer encoder and an adaptive fusion module is introduced to capture the global contextual features of the infrared and visible images. Finally, the decoder integrates the low- and high-frequency features from both branches to generate the final fused image. Image fusion, object detection, and semantic segmentation experiments conducted on public datasets such as TNO, MFNet, and (MFD)-F-3, show that our method outperforms the state-of-the-art (SOTA) image fusion methods.

Crowdsourcing can harness grassroots intelligence to effectively solve complex ranking tasks. Two key issues for these tasks are how to recruit high-performance working groups and how to efficiently aggregate their submissions. In past work, many attempts have been made to maximize social benefits or motivate worker participation, without considering the performance and efficiency of task completion. Therefore, we investigate an effective crowdsourcing ranking solution which is concerned with the two problems of optimal working group selection and efficient result aggregation. A performance-aware working group selection method is proposed to solve the high time complexity of combinatorial optimization in working group selection with a fixed number of requester recruits. To improve the performance of aggregated results more efficiently, an efficient aggregation method based on differential evolution algorithm and Top-k pruning is proposed. Finally, the effectiveness of the proposed methods is verified by extensive experiments compared with recent related methods.

Fast and efficient pest detection in resource-constrained agricultural environments remains a critical yet challenging task, as most existing object detection methods prioritize accuracy over efficiency. This paper explores an optimization scheme for detection heads based on sparse convolution and proposes a novel adaptive sparse convolution network that effectively integrates background features. To enhance compact foreground representation, we introduce an adaptive multilayer mask-ratio strategy that dynamically extracts features at different scales. Additionally, an adaptive threshold segmentation method based on Otsu’s algorithm is developed to improve foreground feature segmentation accuracy. To further balance accuracy and efficiency, we propose a lightweight difference-guided feature fusion method, which enhances feature representation while reducing computational complexity. Experimental results on the IP102 public dataset demonstrate that the proposed module consistently improves performance across four base detectors, achieving a 0.6%–1.2% increase in mAP, a 26.7%–33.5% reduction in GFLOPs, and a 16.4%–30.2% improvement in FPS. Furthermore, while maintaining accuracy comparable to state-of-the-art methods, our approach significantly improves detection efficiency and speed, demonstrating strong potential for real-world agricultural applications.

Performing cross-subject and cross-session emotion recognition (ER) using EEG signals is challenging, since the nonstationarity characteristics and individual differences of EEG might lead to EEG feature distribution differences which may reduce the generalization ability of traditional models. Existing studies adopt domain adaptation algorithms to address this issue. However, most of them focus on global alignment without considering much fine-grained information, degenerating the emotion discrimination ability of EEG features. In this study, for cross-subject and cross-session EEG ER, we propose an unsupervised subdomain adaptation framework guided by pseudo label. In the framework, to reduce EEG distribution differences, a PMSan method is proposed to align EEG feature distribution at the subdomain level, where a multi-representation feature extraction (MFE) is introduced to capture diverse high-level representations, and a PLMMD loss is proposed to minimize subdomain discrepancy to learn multiple domain-invariant feature subspaces. In addition, a hierarchical pseudolabel reweighting method for EEG (EEGHPW) is proposed to generate hierarchical weights to reduce pseudo label noise, ensuring the performance of subdomain adaptation. Experiments on SEED and SEED-IV datasets showed that our framework can achieve significant performance improvements in both cross-subject and cross-session scenarios.

Business forms with the dense text boxes have a complicated layout, diverse content, and low quality. It is a challenging task for the existing methods of form understanding to recognize form structure and to meet the requirements in real-time application scenarios. In this paper, we propose a novel multi-task lightweight block pyramid network for form segmentation and classification, named BPFormNet. According to the characteristics of the form images, we exploit the multi-scale pyramidal feature hierarchy of CNN (Convolutional Neural Network) to construct a multi-level, multi-scale block pyramid, which consists of the low-level, mid-level, and high-level convolutional blocks designed for the corresponding feature layer, and builds the semantic feature maps of multi-scale effective fusion at every level. BPFromNet leverages the interdependence between the twin task of segmentation task of form frames and classification task to improve the performance of the classification under the training strategy of small samples. Furthermore, BPFormNet performs comprehensive lightweighting from three levels: multi-level, multi-scale convolutional block combination, multi-size kernel combination, and disassembly of kernels. Experimental results on the collected image dataset of Chinese insurance forms (CIF) show that BPFormNet with the block pyramid has a strong capability of form feature representation. Comparing with some state-of-the-art (SOTA) lightweight models and their combinations, BPFormNet achieves a better performance in both segmentation and classification task than the models of single block, significantly reduces the model complexity while maintaining the model accuracy, and provides the real-time, high-quality results of form structure recognition for the downstream task of text recognition and information extraction.

AbstractPerson re‐identification is aimed at searching for specific target pedestrians from non‐intersecting cameras. However, in real complex scenes, pedestrians are easily obscured, which makes the target pedestrian search task time‐consuming and challenging. To address the problem of pedestrians' susceptibility to occlusion, a person re‐identification via deep compound eye network (CEN) and pose repair module is proposed, which includes (1) A deep CEN based on multi‐camera logical topology is proposed, which adopts graph convolution and a Gated Recurrent Unit to capture the temporal and spatial information of pedestrian walking and finally carries out pedestrian global matching through the Siamese network; (2) An integrated spatial‐temporal information aggregation network is designed to facilitate pose repair. The target pedestrian features under the multi‐level logic topology camera are utilised as auxiliary information to repair the occluded target pedestrian image, so as to reduce the impact of pedestrian mismatch due to pose changes; (3) A joint optimisation mechanism of CEN and pose repair network is introduced, where multi‐camera logical topology inference provides auxiliary information and retrieval order for the pose repair network. The authors conducted experiments on multiple datasets, including Occluded‐DukeMTMC, CUHK‐SYSU, PRW, SLP, and UJS‐reID. The results indicate that the authors’ method achieved significant performance across these datasets. Specifically, on the CUHK‐SYSU dataset, the authors’ model achieved a top‐1 accuracy of 89.1% and a mean Average Precision accuracy of 83.1% in the recognition of occluded individuals.

Edge-aware image decomposition is an essential topic in the field of multimedia signal processing. In this paper, we propose a novel non-convex penalty function, which we name the generalized Welsch function. We show that the proposed penalty function is more than a generalization of most existing penalty functions for edge-aware regularization, thus, it better facilitates edge-awareness. We embed the proposed penalty function into a novel optimization model for edge-aware image decomposition. To solve the optimization model with non-convex penalty function, we propose an efficient algorithm based on the additive quadratic minimization and Fourier domain optimization. We have experimented with the proposed method in a variety of tasks, including image smoothing, detail enhancement, HDR tone mapping, and JPEG compression artifact removal. Experiment results show that our method outperforms the state-of-the-art image decomposition methods. Furthermore, our method is highly efficient, it is able to render real-time processing of 720P color images on a modern GPU.

Prompt tuning, which fine-tunes the feature distributions in pre-trained Vision-Language (VL) models by adding learnable tokens or contexts into image and text branches, has emerged as a popular method for enhancing task-specific performance. However, this approach may result in overfitting specific target data distributions, thereby undermining the original generalization capabilities of frozen models such as CLIP. To tackle this issue, a novel framework named Constraint Embedding for Prompt Tuning (CEPT) is proposed for optimizing the learnable prompt tokens. To maintain the feature extraction capabilities of the pre-trained CLIP model while extracting relevant data features for downstream tasks, the Block Consistency Constraint (BCC) approach is proposed. This approach adjusts the feature extraction step by ensuring that block-wise embeddings are aligned, thereby preserving the original generalization performance of the pre-trained VL model. Additionally, to achieve a more harmonious distribution of image-text features in the potential space, the Distribution Constraint (DC) strategy is introduced. This strategy enhances multimodal data feature alignment by evenly dispersing different classes of data features and concentrating the same class of image features within the potential space. Finally, CEPT surpassed the state-of-the-art for base-to-novel generalization, achieving a harmonic mean improvement of over 1.04%. Additionally, for few-shot learning, it demonstrates an average improvement of 1.63% across five few-shot scenarios.

The traditional $L_{0}$ filter shows exquisite smoothing quality, but it suffers from high computational cost. In this paper, we propose an efficient solution to the $L_{0}$ -regularized optimization problem based on deep unsupervised learning. The $L_{0}$ -norm involves a discrete counting scheme, which can not be directly optimized with gradient descent. Therefore, in this paper, we propose to decompose the problem into a series of optimization problems based on a truncated $L_{1}$ -norm with varying parameters. Compared with the truncated $L_{2}$ -norm explored in traditional $L_{0}$ filter, the truncated $L_{1}$ -norm promotes the capabilities in structure- and edge-preserving smoothing, reduces the number of iterations, and facilitates the deep learning-based optimization. Furthermore, we propose a deep learning-based parameterized approach to solve the truncated $L_{1}$ -regularized problems so that we only need to train a single fully convolutional network to support varying smoothing parameters. We are not trying to reproduce the traditional $L_{0}$ filter in this paper. Instead, we show that the proposed deep $L_{0}$ filter provides a better smoothing quality. Experimental results indicate that the proposed filter outperforms the state-of-the-art on various applications, including edge-preserving smoothing, non-photorealistic rendering, texture removal, edge extraction, image composition, and compression artifact removal. Moreover, our filter is efficient, it is able to process 720P color images at interactive rates on a modern GPU.

When an abnormal event occurs, the features of the video will gradually change in the temporal sequences. Since the abnormal event is a continuous process, the abnormal features should have a continuous prominence compared to the normal features. Incorporating this differential information is critical for effective anomaly detection. Nevertheless, prior multi-instance learning (MIL) methods have overlooked this aspect. Therefore, this study proposes the spatial–temporal feature fusion enhancement (STFFE) learning approach to address this issue. STFFE improves the discriminative power of normal and abnormal features by enhancing temporal information through the fusion of top-k video segment features with their corresponding temporal information features. Furthermore, a temporal information constraint is proposed to increase the concentration of abnormal information and accentuate the saliency of abnormal features. Extensive experimentation demonstrates that our approach surpasses state-of-the-art (SOTA) methods on the Shanghai-Tech and XD-Violence datasets. Moreover, our method achieves competitive results on the UCF-Crime dataset, with an AUC of 84.52

Few-Shot Action Recognition (FSAR) aims at recognizing novel action classes with only a few labelled samples. Due to its simplicity and effectiveness, the prototypical network has attracted increasing interest in the field of FSAR. It is key to learn representative prototypes from a few labelled videos with various action lengths and speeds. To address this issue, this paper presents a dual-prototype network that combines class-specific and query-specific attentive learning for FSAR. First, we propose a class-specific attentive learning method that computes the within-class similarity for each class of the support sample. This method not only increases the representativeness of prototypes but also mitigates the impact of noises and outlying samples. Second, the class-specific attention is combined with query-specific attention to establish two parallel sets of prototypes for FSAR. The incorporation of query-specific attention further increases the discrimination among prototypes for different query samples. Furthermore, we propose a temporal-relation model to express the temporal dependency of action videos with different lengths and speeds. The proposed method is validated on four benchmark datasets. Extensive experimental results demonstrate the superiority of our method to the other 11 state-of-the-art FSAR methods.

In the field of video anomaly detection, effective representation for the normal and abnormal video features is a key issue for improving the accuracy of detection models. The feature representation based on global spatio-temporal networks lacks to consider negative impacts for the features between normal segment and abnormal segment, resulting in incorrect detection for some segments in video. To address this issue, a new approach of normal–abnormal negative impacts suppressing via normal feature memory for video anomaly detection is proposed. In this method, feature representation with normal–abnormal negative impacts suppression via normal feature memory pool is modeled, which can be used to suppress normal–abnormal negative impacts in video feature learning. Constraint learning for normal feature memory pool is designed to promote more effective learning for the normal feature memory pool and feature of abnormal video segment. Extensive experiments demonstrate that the proposed method outperforms the relative state-of-the-art methods on the Shanghai-Tech and XD-Violence datasets and achieves competitive results on the UCF-Crime dataset. This indicates that the proposed method can effectively enhances the performance of anomaly detection, thereby validating its effectiveness in practical applications.

Traditional classification tasks suffer from the class-imbalanced problem, where some classes far outnumber others. To address this issue, existing class-imbalanced learning (CIL) methods either preprocess class-imbalanced datasets or adapt traditional classification algorithms to the imbalanced class distribution. Inspired by the idea of transductive learning, we propose a post-processing framework called PPF for CIL. Distinct from existing CIL methods, PPF directly adjusts the predicted labels of test data to fit the imbalanced class distribution. Specifically, we relabel some test data according to their prediction probabilities so that the class proportion of test data is close to that of training data. The underlying assumption is that training and test data, drawn independently from one data space, should obey the same class distribution. Furthermore, we propose a Compact Prototype-based Nearest Neighbor (CPNN) algorithm to assist the original classifier with the adjustment. Instead of training a classifier, CPNN classifies test data according to their distances to a set of prototypes estimated on labeled data. Thus, it is computationally simple and relatively robust to class imbalance. As a general framework, PPF can be easily applied to both traditional classification and CIL algorithms. To validate the effectiveness of the proposed method, we conducted extensive experiments on a variety of class-imbalanced datasets, using SVM and C4.5 as the original classifiers, respectively. Measured by F-measure, G-mean, and AUC, both PPF-SVM and PPF-C4.5 outperform 10 state-of-the-art CIL algorithms. Additionally, PPF further improved their performances when applied to 10 CIL algorithms.

As vision sensor technology continues to evolve, the requirements for detecting targets of interest in the images captured by the sensors are increasing. Considering fast detection and high accuracy, the industry favors geometric key point-based solutions. However, there are a large number of small and fuzzy objects in the real world. Geometric key point detectors do not effectively utilize the contextual features of the region of interest, leading to excessive false positive and false negative results. In this work, a simple, effective, and interpretable tiny object detection method called Regional Cross Self-Attention Object Detection Network (RCSANet) is proposed. It adopts Region Proposal Networks and transformers to capture regional background relations and uses regional background relations to generate key point sequences. The regional cross self-attention mechanism is introduced to curtail computation redundancy and minimize the interference of redundant information to the target region. Additionally, a position coding called dynamic implicit position coding is proposed to cooperate with regional cross self-attentiveness. Dynamic implicit location coding can encode arbitrarily long input sequences. The computational cost of RCSANet is significantly lower than that of state-of-the-art object detection solutions. Moreover, RCSANet improves the performance on the four benchmark datasets, of MSCOCO, Tinyperson, DOTA, and AI-TOD, by about 3.0%AP.