Feature Representation Ability Research Articles

Attention-disentangled re-ID network for unsupervised domain adaptive person re-identification

Unsupervised domain adaptation (UDA) for person re-identification (re-ID) aims to bridge the domain gap by transferring knowledge from the labeled source domain to the unlabeled target domain. Recently, pseudo-label-based approaches have become the dominant solution for addressing this issue. However,most pseudo-label-based methods neglect class boundary samples to reduce the influence of false pseudo labels, sacrificing intra-class semantic diversity. Although hard sample memory bank-based approaches can discover relationships between samples and describe intra-class diversity, they are prone to increasing the risks of generating incorrect pseudo labels. Balancing intra-class variety and pseudo-label accuracy in cross-domain person re-ID poses a challenge. In this study, we propose an attention-disentangled re-ID network (ADDNet) to enhance the discriminative ability of re-ID related feature representations, addressing the contradiction between intra-class diversity and pseudo-label accuracy. Unlike most disentanglement learning-based re-ID approaches that focus on separating explanatory factors, we design a spatial attention-disentangled mechanism to separate re-ID related and unrelated weights, enhancing the discriminative ability of cross-domain feature representation. Additionally, a multiple hard sample memory learning strategy is designed to express intra-class diversity of target samples using re-ID related features extracted from ADDNet. Extensive experiments show that ADDNet achieves 46.7% mAP on the Market-to-MSMT cross-domain re-ID task, outperforming state-of-the-art methods by 6.5 points.

Ship detection method based on attention guidance and multi-sample decision making

Single-stage target detection methods have the characteristics of fast training speed and short detection time. However, its feature pyramid network is difficult to suppress the background and noise information of SAR ship images, and the detection head has prediction errors. To address this problem, this paper proposes a detection model based on attention guidance and multi-sample decision for synthetic aperture radar ship detection. Firstly, an attention guidance network is proposed and added to the highest level of the feature pyramid to suppress background and noise interference, thereby improving the representation ability of features. Secondly, a multi-sample decision network is proposed to participate in the prediction of target position. This network alleviates the impact of prediction errors on detection results by increasing the number of samples output in the regression branch. Finally, a novel maximum likelihood loss function is designed. This loss function constructs a maximum likelihood function using the samples output from the multi-sample decision network, which is used to standardize the training of the decision network and further improve the accuracy of target positioning. Taking the RetinaNet network model as the baseline method, compared with the baseline method and the current advanced target detection methods, this method shows the highest detection accuracy on the ship detection dataset SSDD, with AP reaching 52.8%. Compared with the baseline method, the proposed method improves the AP evaluation index 3.4% ∼ 5.7%, and the training parameter Params only increases by 2.03 M, and the frame rate FPS only decreases 0.5Iter/s.

Power text information extraction based on multi-task learning

In order to improve the analysis and processing speed of power system fault text in actual business scenarios, a power fault text information automatic extraction model based on pre-training and multi-task learning is proposed. The pre-training model is used to learn the context information of power text words, and the first-order and second-order fusion features of words are mined to enhance the representation ability of features. The multi-task learning framework is used to combine the learning of named entity recognition and relationship extraction tasks to achieve mutual complementation and mutual promotion between entity recognition and relationship extraction, thereby improving the performance of power fault text information extraction. Finally, the model is verified by the daily business data of a power grid data center. Compared with other models, the accuracy and recall rate of power fault text entity recognition and relationship extraction are improved.

OGRN: optical-guided residual dense network for SAR image super-resolution reconstruction network

ABSTRACT Although Convolutional Neural Networks have significantly improved the development of SAR image super-resolution (SR) technology in recent years, it is a very challenging problem to reconstruct SAR image with large-scale factors, such as ×4 and ×8 due to limited available information from low-resolution image. The co-registered high-resolution optical image has been successfully applied to enhance the quality of SAR image due to its discriminative characteristics. Compared with single-frame SAR image SR reconstruction technology, optical image-guided SAR image SR reconstruction technology has better performance. This paper proposes an optical-guided residual dense network (OGRN) for SAR image super-resolution reconstruction network. The network is an end-to-end network. Its most important characteristic is that the SR model of SAR images can be obtained through optical image assistance during training, and high-quality SAR images can be generated without optical image assistance during testing. Firstly, a new dense residual backbone network that extracts high-frequency information is constructed. It can extract and propagate high-frequency features of SAR images efficiently to improve the feature representation ability of the network for high-frequency features. Then, the extracted high-frequency features of the SAR image are used to generate an optical image through the designed SAR-to-Optical image translation network. Finally, a reconstruction module is constructed based on the proposed fusion module for optical and SAR images. The fusion module fuses the high-frequency feature information of the extracted optical images with the SAR image features to provide features for SAR image SR reconstruction. Extensive experiments conducted on Sen1-2 and QXS datasets demonstrated that under the guidance of optical image, our OGRN can achieve excellent performance in both quantitative assessment metrics and visual quality.

A Representation-Learning-Based Graph and Generative Network for Hyperspectral Small Target Detection

Hyperspectral small target detection (HSTD) is a promising pixel-level detection task. However, due to the low contrast and imbalanced number between the target and the background spatially and the high dimensions spectrally, it is a challenging one. To address these issues, this work proposes a representation-learning-based graph and generative network for hyperspectral small target detection. The model builds a fusion network through frequency representation for HSTD, where the novel architecture incorporates irregular topological data and spatial–spectral features to improve its representation ability. Firstly, a Graph Convolutional Network (GCN) module better models the non-local topological relationship between samples to represent the hyperspectral scene’s underlying data structure. The mini-batch-training pattern of the GCN decreases the high computational cost of building an adjacency matrix for high-dimensional data sets. In parallel, the generative model enhances the differentiation reconstruction and the deep feature representation ability with respect to the target spectral signature. Finally, a fusion module compensates for the extracted different types of HS features and integrates their complementary merits for hyperspectral data interpretation while increasing the detection and background suppression capabilities. The performance of the proposed approach is evaluated using the average scores of AUCD,F, AUCF,τ, AUCBS, and AUCSNPR. The corresponding values are 0.99660, 0.00078, 0.99587, and 333.629, respectively. These results demonstrate the accuracy of the model in different evaluation metrics, with AUCD,F achieving the highest score, indicating strong detection performance across varying thresholds. Experiments on different hyperspectral data sets demonstrate the advantages of the proposed architecture.

MobilenetV2-RC: A lightweight network model for retinopathy classification in retinal OCT images

Abstract Retinopathy is an important ophthalmic disease that causes blindness in the elderly population. With the global old peoples increasing, the ophthalmic healthcare system is more and significant to pre-diagnosis. Optical Coherence Tomography (OCT) is considered the gold standard for ophthalmic treatment and diagnosis. OCT technologies and equipment continue to develop towards the intelligence and convenience for requirements of rapid diagnosis in the remote and poverty-stricken areas. We proposed an improved MobilenetV2 lightweight model for retinopathy classification (MobilenetV2-RC), which incorporates Spatial and Channel Reconstruction Convolution (SCConv) and the improved Convolutional Block Attention Module (CBAM) attention mechanism into the framework. Not only can it effectively limit feature redundancy to reduce model parameters, but also enhance the ability of feature representation to improve classification accuracy. The parameters of the proposed model are only 1.96 M with an overall accuracy of 98.96%, which is higher 3.32% than the original MobilenetV2. Compared with ResNet18, InceptionV3, and VGG16_BN, the overall accuracy is increased by 4.6%, 6.3%, and 3.9%, respectively. The test results of UCSD and Duke open-source datasets are more remarkable. Experimental results show that our proposed algorithm has strong reliability and generalization for the accurate classification of retinopathy, and a greater application prospect in the intelligent diagnosis of ophthalmology and mobile detection terminals.

CCM-Net: Color compensation and coordinate attention guided underwater image enhancement with multi-scale feature aggregation

Due to the light scattering and wavelength absorption in water, underwater images exhibit blurred details, low contrast, and color deviation. Existing underwater image enhancement methods are divided into traditional methods and deep learning-based methods. Traditional methods either rely on scene prior and lack robustness, or are not flexible enough resulting in poor enhancement effects. Deep learning methods have achieved good results in the field of underwater image enhancement due to their powerful feature representation ability. However, these methods cannot enhance underwater images with various degradations because they do not consider the inconsistent attenuation of different color channels and spatial regions. In this paper, we propose a novel asymmetric encoder-decoder network for underwater image enhancement, called CCM-Net. Concretely, we first introduce the prior knowledge-based encoder, which includes color compensation (CC) modules and feature extraction modules that consist of depth-wise separable convolution and global-local coordinate attention (GLCA). Then, we design a multi-scale feature aggregation (MFA) module to integrate shallow, middle, and deep features. Finally, we deploy a decoder to reconstruct the underwater images with the extracted features. Extensive experiments on publicly available datasets demonstrate that our CCM-Net effectively improves the visual quality of underwater images and achieves impressive performance.

CLESSR-VC: Contrastive learning enhanced self-supervised representations for one-shot voice conversion

One-shot voice conversion (VC) has attracted more and more attention due to its broad prospects for practical application. In this task, the representation ability of speech features and the model’s generalization are the focus of attention. This paper proposes a model called CLESSR-VC, which enhances pre-trained self-supervised learning (SSL) representations through contrastive learning for one-shot VC. First, SSL features from the 23rd and 9th layers of the pre-trained WavLM are adopted to extract content embedding and SSL speaker embedding, respectively, to ensure the model’s generalization. Then, the conventional acoustic feature mel-spectrograms and contrastive learning are introduced to enhance the representation ability of speech features. Specifically, contrastive learning combined with the pitch-shift augmentation method is applied to disentangle content information from SSL features accurately. Mel-spectrograms are adopted to extract mel speaker embedding. The AM-Softmax and cross-architecture contrastive learning are applied between SSL and mel speaker embeddings to obtain the fused speaker embedding that helps improve speech quality and speaker similarity. Both objective and subjective evaluation results on the VCTK corpus confirm that the proposed VC model has outstanding performance and few trainable parameters.

FBENet: Feature-Level Boosting Ensemble Network for Hashimoto's Thyroiditis Ultrasound Image Classification.

Distinguishing Hashimoto's thyroiditis (HT) lesions from ordinary thyroid tissues is difficult with ultrasound images. Challenges in achieving high performance of HT ultrasound image classification include the low resolution, blurred features and large area of irrelevant noise. To address these problems, we propose a Feature-level Boosting Ensemble Network (FBENet) for HT ultrasound image classification. Specifically, to capture the features of suspicious HT lesions efficiently, an Ensemble Feature Boosting Module (EFBM) is introduced into the feature-level ensemble to boost the blurred features. Then, the spatial attention mechanism is adopted in backbone models to improve the feature focusing performance and representation ability. Furthermore, feature-level ensemble technique is employed in the training process to achieve more comprehensive feature representation ability. Experimentally, FBENet was trained on 6,503 HT ultrasound images, and tested on 1,626 HT ultrasound images with 82.92% accuracy and 89.24% AUC on average.

A Dynamic Multi‐Output Convolutional Neural Network for Skin Lesion Classification

ABSTRACTSkin cancer is a pressing global health issue, with high incidence and mortality rates. Convolutional neural network (CNN) models have been proven to be effective in improving performance in skin lesion image classification and reducing the medical burden. However, the inherent class imbalance in training data caused by the difficulty of collecting dermoscopy images leads to categorical overfitting, which still limits the effectiveness of these data‐driven models in recognizing few‐shot categories. To address this challenge, we propose a dynamic multi‐output convolutional neural network (DMO‐CNN) model that incorporates exit nodes into the standard CNN structure and includes feature refinement layers (FRLs) and an adaptive output scheduling (AOS) module. This model improves feature representation ability through multi‐scale sub‐feature maps and reduces the inter‐layer dependencies during backpropagation. The FRLs ensure efficient and low‐loss down‐sampling, while the AOS module utilizes a trainable layer selection mechanism to refocus the model's attention on few‐shot lesion categories. Additionally, a novel correction factor loss is introduced to supervise and promote AOS convergence. Our evaluation of the DMO‐CNN model on the HAM10000 dataset demonstrates its effectiveness in multi‐class skin lesion classification and its superior performance in recognizing few‐shot categories. Despite utilizing a very simple VGG structure as the sole backbone structure, DMO‐CNN achieved impressive performance of 0.885 in BACC and 0.983 in weighted AUC. These results are comparable to those of the ensemble model that won the ISIC 2018 challenge, highlighting the strong potential of DMO‐CNN in dealing with few‐shot skin lesion data.

PSD-ELGAN: A pseudo self-distillation based CycleGAN with enhanced local adversarial interaction for single image dehazing

Compared to pixel-level content loss, domain-level style loss in CycleGAN-based dehazing algorithms just imposes relatively soft constraints on the intermediate translated images, resulting in struggling to accurately model haze-free features from real hazy scenes. Furthermore, globally perceptual discriminator may misclassify real hazy images with significant scene depth variations as clean style, thereby resulting in severe haze residue. To address these issues, we propose a pseudo self-distillation based CycleGAN with enhanced local adversarial interaction for image dehazing, termed as PSD-ELGAN. On the one hand, we leverage the characteristic of CycleGAN to generate pseudo image pairs during training. Knowledge distillation is employed in this unsupervised framework to transfer the informative high-quality features from the self-reconstruction network of real clean images to the dehazing generator of paired pseudo hazy images, which effectively improves its haze-free feature representation ability without increasing network parameters. On the other hand, in the output of dehazing generator, four non-uniform image patches severely affected by residual haze are adaptively selected as input samples. The local discriminator could easily distinguish their hazy style, thereby further compelling the dehazing generator to suppress haze residues in such regions, thus enhancing its dehazing performance. Extensive experiments show that our PSD-ELGAN can achieve promising results and better generality across various datasets.

Upstream process condition monitoring for froth flotation based on feature performance evaluation and parameter-mapped GRNN

Timely and reliable estimation of the upstream process condition in flotation process is significant for whole flotation operations. An upstream process condition monitoring model based on feature performance evaluation and parameter-mapped generalized regression neural network (GRNN) is proposed to evaluate visual feature performance and improve the generalization ability of ensemble models for zinc flotation. Firstly, different visual features are extracted, and the froth video representation is obtained. Then, feature performance evaluation is constructed, considering the correlation between different visual features and flotation indicators, and stability of feature representation ability vary. Meanwhile, feature performance evaluation indexes are calculated introducing a scale factor to integrate the Pearson correlation with feature stability coefficient. Finally, a parameter-mapped GRNN, whose base model parameters are mapped into a certain interval, is designed to monitor the upstream process condition based on the feature performance evaluation index-weighted feature vector. Experiments using real-world lead–zinc flotation data were implemented to show the potential of the feature performance evaluation and effectiveness of the proposed model.

Spatial attention inference model for cascaded siamese tracking with dynamic residual update strategy

Target representation is crucial for visual tracking. Most Siamese-based trackers try their best to establish target models by using various deep networks. However, they neglect the exploration of correlation among features, which leads to the inability to learn more representative features. In this paper, we propose a spatial attention inference model for cascaded Siamese tracking with dynamic residual update strategy. First, a spatial attention inference model is constructed. The model fuses interlayer multi-scale features generated by dilation convolution to enhance the spatial representation ability of features. On this basis, we use self-attention to capture interaction between target and context, and use cross-attention to aggregate interdependencies between target and background. The model infers potential feature information by exploiting the correlations among features for building better appearance models. Second, a cascaded localization-aware network is introduced to bridge a gap between classification and regression. We propose an alignment-aware branch to resample and learn object-aware features from the predicted bounding boxes for obtaining localization confidence, which is used to correct the classification confidence by weighted integration. This cascaded strategy alleviates the misalignment problem between classification and regression. Finally, a dynamic residual update strategy is proposed. This strategy utilizes the Context Fusion Network (CFNet) to fuse the templates of historical and current frames to generate the optimal templates. Meanwhile, we use a dynamic threshold function to determine when to update by judging the tracking results. The strategy uses temporal context to fully explore the intrinsic properties of the target, which enhances the adaptability to changes in the target’s appearance. We conducted extensive experiments on seven tracking benchmarks, including OTB100, UAV123, TC128, VOT2016, VOT2018, GOT10k and LaSOT, to validate the effectiveness of our proposed algorithm.

Lightweight Corn Leaf Detection and Counting Using Improved YOLOv8.

The number of maize leaves is an important indicator for assessing plant growth and regulating population structure. However, the traditional leaf counting method mainly relies on manual work, which is both time-consuming and straining, while the existing image processing methods have low accuracy and poor adaptability, making it difficult to meet the standards for practical application. To accurately detect the growth status of maize, an improved lightweight YOLOv8 maize leaf detection and counting method was proposed in this study. Firstly, the backbone of the YOLOv8 network is replaced using the StarNet network and the convolution and attention fusion module (CAFM) is introduced, which combines the local convolution and global attention mechanisms to enhance the ability of feature representation and fusion of information from different channels. Secondly, in the neck network part, the StarBlock module is used to improve the C2f module to capture more complex features while preserving the original feature information through jump connections to improve training stability and performance. Finally, a lightweight shared convolutional detection head (LSCD) is used to reduce repetitive computations and improve computational efficiency. The experimental results show that the precision, recall, and mAP50 of the improved model are 97.9%, 95.5%, and 97.5%, and the numbers of model parameters and model size are 1.8 M and 3.8 MB, which are reduced by 40.86% and 39.68% compared to YOLOv8. This study shows that the model improves the accuracy of maize leaf detection, assists breeders in making scientific decisions, provides a reference for the deployment and application of maize leaf number mobile end detection devices, and provides technical support for the high-quality assessment of maize growth.

MFPIDet: improved YOLOV7 architecture based on multi-scale feature fusion for prohibited item detection in complex environment

Prohibited item detection is crucial for the safety of public places. Deep learning, one of the mainstream methods in prohibited item detection tasks, has shown superior performance far beyond traditional prohibited item detection methods. However, most neural network architectures in deep learning still lack sufficient local feature representation ability for overlapping and small targets, and ignore the problem of semantic conflicts caused by direct feature fusion. In this paper, we propose MFPIDet, a novel prohibited item detection neural network architecture based on improved YOLOV7 to achieve reliable prohibited item detection in complex environments. Specifically, a multi-scale attention module (MAM) backbone is proposed to filter the redundant information of target regions and further applied to enhance the local feature representation ability of overlapping objects. Here, to reduce the redundant information of target regions, a squeeze-excitation (SE) block is used to filter the background. Then, aiming at enhancing the feature expression ability of overlapping objects, a multi-scale feature extraction module (MFEM) is designed for local feature representation. In addition, to obtain richer context information, We design an adaptive fusion feature pyramid network (AF-FPN) to combine the adaptive context information fusion module (ACIFM) with the feature fusion module (FFM) to improve the neck structure of YOLOV7. The proposed method is validated on the PIDray dataset, and the tested results showed that our method obtained the highest mAP (68.7%), which is improved by 3.5% than YOLOV7 methods. Our approach provides a new design pattern for prohibited item detection in complex environments and shows the development potential of deep learning in related fields.

A lightweight self-supervised learning segmentation model for variable and complex high-resolution remote sensing images

The complexity and variability of high-resolution remote sensing data, such as high intra-class variability and inter-class similarity, pose significant challenges to model segmentation. To address the problem, this paper constructs a lightweight self-supervised learning model for multi-label segmentation in the form of phased learning of multi-scale features. The model adopts axial depthwise separable convolutions to reduce computational complexity and enhance feature representation, utilizes dilated rates to capture large-scale and multi-scale contextual information for long-distance feature extraction, and incorporates convolution kernels of varying sizes to acquire both local and global feature information for the improved ability of learning feature representation. The experimental results show that our model achieves competitive performance and has smaller weight parameters, memory usage, and lower computational complexity compared with existing classical models that rely on large-scale weight parameters. Additionally, our ablation study delves into the encountered design issues to elucidate the rationality of our approach. The source code is avaiable: https://github.com/zhangjy2008327/remote-sensing-images.

Vehicle Classification Algorithm Based on Improved Vision Transformer

Vehicle classification technology is one of the foundations in the field of automatic driving. With the development of deep learning technology, visual transformer structures based on attention mechanisms can represent global information quickly and effectively. However, due to direct image segmentation, local feature details and information will be lost. To solve this problem, we propose an improved vision transformer vehicle classification network (IND-ViT). Specifically, we first design a CNN-In D branch module to extract local features before image segmentation to make up for the loss of detail information in the vision transformer. Then, in order to solve the problem of misdetection caused by the large similarity of some vehicles, we propose a sparse attention module, which can screen out the discernible regions in the image and further improve the detailed feature representation ability of the model. Finally, this paper uses the contrast loss function to further increase the intra-class consistency and inter-class difference of classification features and improve the accuracy of vehicle classification recognition. Experimental results show that the accuracy of the proposed model on the datasets of vehicle classification BIT-Vehicles, CIFAR-10, Oxford Flower-102, and Caltech-101 is higher than that of the original vision transformer model. Respectively, it increased by 1.3%, 1.21%, 7.54%, and 3.60%; at the same time, it also met a certain real-time requirement to achieve a balance of accuracy and real time.

Improved 3D Object Detection Based on PointPillars

Despite the recent advancements in 3D object detection, the conventional 3D point cloud object detection algorithms have been found to exhibit limited accuracy for the detection of small objects. To address the challenge of poor detection of small-scale objects, this paper adopts the PointPillars algorithm as the baseline model and proposes a two-stage 3D target detection approach. As a cutting-edge solution, point cloud processing is performed using Transformer models. Additionally, a redefined attention mechanism is introduced to further enhance the detection capabilities of the algorithm. In the first stage, the algorithm uses PointPillars as the baseline model. The central concept of this algorithm is to transform the point cloud space into equal-sized columns. During the feature extraction stage, when the features from all cylinders are transformed into pseudo-images, the proposed algorithm incorporates attention mechanisms adapted from the Squeeze-and-Excitation (SE) method to emphasize and suppress feature information. Furthermore, the 2D convolution of the traditional backbone network is replaced by dynamic convolution. Concurrently, the addition of the attention mechanism further improves the feature representation ability of the network. In the second phase, the candidate frames generated in the first phase are refined using a Transformer-based approach. The proposed algorithm applies channel weighting in the decoder to enhance channel information, leading to improved detection accuracy and reduced false detections. The encoder constructs the initial point features from the candidate frames for encoding. Meanwhile, the decoder applies channel weighting to enhance the channel information, thereby improving the detection accuracy and reducing false detections. In the KITTI dataset, the experimental results verify the effectiveness of this method in small objects detection. Experimental results show that the proposed method significantly improves the detection capability of small objects compared with the baseline PointPillars. In concrete terms, in the moderate difficulty detection category, cars, pedestrians, and cyclists average precision (AP) values increased by 5.30%, 8.1%, and 10.6%, respectively. Moreover, the proposed method surpasses existing mainstream approaches in the cyclist category.

Adversarial alignment and graph fusion via information bottleneck for multimodal emotion recognition in conversations

With the rapid development of social media and human–computer interaction, multimodal emotion recognition in conversations (MERC) tasks have begun to receive widespread research attention. The MERC task is to extract and fuse complementary semantic information from different modalities to classify the speaker’s emotion. However, the existing feature fusion methods usually directly map the features of other modalities into the same feature space for information fusion, which cannot eliminate the heterogeneity between different modalities and make the subsequent emotion class boundary learning more difficult. In addition, existing graph contrastive learning methods obtain consistent feature representations by maximizing mutual information between multiple views, which may lead to overfitting of the model. To tackle the above problem, we propose a novel Adversarial Alignment and Graph Fusion via Information Bottleneck for Multimodal Emotion Recognition in Conversations (AGF-IB) method. Firstly, we input video, audio, and text features into a multi-layer perceptron (MLP) to map them into separate feature spaces. Secondly, we build a generator and a discriminator for the three modal features, respectively, through adversarial representation to achieve information interaction between modalities and eliminate the heterogeneity among modalities. Thirdly, we introduce graph contrastive representation learning to capture intra-modal and inter-modal complementary semantic information and learn intra-class and inter-class boundary information of emotion categories. Furthermore, instead of maximizing the mutual information (MI) between multiple views, we use information bottleneck theory to minimize the MI between views. Specifically, we construct a graph structure for the three modal features respectively and perform contrastive representation learning on nodes with different emotions in the same modality and nodes with the same emotion in different modalities, to improve the feature representation ability of nodes. Finally, we use MLP to complete the emotional classification of the speaker. Extensive experiments show that AGF-IB can improve emotion recognition accuracy on IEMOCAP and MELD datasets. Furthermore, since AGF-IB is a general multimodal fusion and contrastive learning method, it can be applied to other multimodal tasks in a plug-and-play manner, e.g., humor detection.

Enhancing local representation learning through global–local integration with functional connectivity for EEG-based emotion recognition

Emotion recognition based on electroencephalogram (EEG) signals is crucial in understanding human affective states. Current research has limitations in extracting local features. The representation capabilities of local features are limited, making it difficult to comprehensively capture emotional information. In this study, a novel approach is proposed to enhance local representation learning through global–local integration with functional connectivity for EEG-based emotion recognition. By leveraging the functional connectivity of brain regions, EEG signals are divided into global embeddings that represent comprehensive brain connectivity patterns throughout the entire process and local embeddings that reflect dynamic interactions within specific brain functional networks at particular moments. Firstly, a convolutional feature extraction branch based on the residual network is designed to extract local features from the global embedding. To further improve the representation ability and accuracy of local features, a multidimensional collaborative attention (MCA) module is introduced. Secondly, the local features and patch embedded local embeddings are integrated into the feature coupling module (FCM), which utilizes hierarchical connections and enhanced cross-attention to couple region-level features, thereby enhancing local representation learning. Experimental results on three public datasets show that compared with other methods, this method improves accuracy by 4.92% on the DEAP, by 1.11% on the SEED, and by 7.76% on the SEED-IV, demonstrating its superior performance in emotion recognition tasks.