Multi-scale Attention Research Articles

Integrating multi-scale attention network with deep image prior for single image restoration

Integrating multi-scale attention network with deep image prior for single image restoration

Multi-Scale Geometric Feature Extraction and Global Transformer for Real-World Indoor Point Cloud Analysis

Indoor point clouds often present significant challenges due to the complexity and variety of structures and high object similarity. The local geometric structure helps the model learn the shape features of objects at the detail level, while the global context provides overall scene semantics and spatial relationship information between objects. To address these challenges, we propose a novel network architecture, PointMSGT, which includes a multi-scale geometric feature extraction (MSGFE) module and a global Transformer (GT) module. The MSGFE module consists of a geometric feature extraction (GFE) module and a multi-scale attention (MSA) module. The GFE module reconstructs triangles through each point’s two neighbors and extracts detailed local geometric relationships by the triangle’s centroid, normal vector, and plane constant. The MSA module extracts features through multi-scale convolutions and adaptively aggregates features, focusing on both local geometric details and global semantic information at different scale levels, enhancing the understanding of complex scenes. The global Transformer employs a self-attention mechanism to capture long-range dependencies across the entire point cloud. The proposed method demonstrates competitive performance in real-world indoor scenarios, with a mIoU of 68.6% in semantic segmentation on S3DIS and OA of 86.4% in classification on ScanObjectNN.

DeepNeXt: a lightweight polyp segmentation algorithm based on multi-scale attention

DeepNeXt: a lightweight polyp segmentation algorithm based on multi-scale attention

Remaining useful life prediction of rolling bearings based on parallel feature extraction

Purpose This study aims to introduce a method for predicting the remaining useful life (RUL) of bearings based on parallel feature extraction. The proposed model provides prior knowledge and removes redundant handcrafted feature information, additionally, which focuses on the important features at different time scales. Design/methodology/approach Distinct from traditional parallel feature extraction methods, which can lead to information redundancy, a one-dimensional convolutional autoencoder is introduced to process selected indicators to remove redundancy and retain useful feature information. To fully capture the important degradation information within different stages in the feature sequences, a novel multi-scale attention feature fusion module is proposed to extract degradation features at different time scales. Considering the impact of degradation modes on RUL prediction, a dual-task prediction module based on no degradation mode labels is designed to obtain accurate RUL. Findings Comparative experiments and ablation studies on the PHM2012 bearing dataset verified the effectiveness of the proposed method. Furthermore, the rationality of the selected parameters is confirmed through model parameter analysis. Originality/value The novelty of the proposed method is that it not only provides prior knowledge but also further removes redundant information from prior knowledge. In addition, the distribution differences between the original features and their multi-scale convolution results are measured through Kullback–Leibler divergence as the attention scores, which allows the proposed method to focus on important information at different time scales.

AFENet: An Attention-Focused Feature Enhancement Network for the Efficient Semantic Segmentation of Remote Sensing Images

The semantic segmentation of high-resolution remote sensing images (HRRSIs) faces persistent challenges in handling complex architectural structures and shadow occlusions, limiting the effectiveness of existing deep learning approaches. To address these limitations, we propose an attention-focused feature enhancement network (AFENet) with a novel encoder–decoder architecture. The encoder architecture combines ResNet50 with a parallel multistage feature enhancement group (PMFEG), enabling robust feature extraction through optimized channel reduction, scale expansion, and channel reassignment operations. Building upon this foundation, we develop a global multi-scale attention mechanism (GMAM) in the decoder that effectively synthesizes spatial information across multiple scales by learning comprehensive global–local relationships. The architecture is further enhanced by an efficient feature-weighted fusion module (FWFM) that systematically integrates remote spatial features with local semantic information to improve segmentation accuracy. Experimental results across diverse scenarios demonstrate that AFENet achieves superior performance in building structure detection, exhibiting enhanced segmentation connectivity and completeness compared to state-of-the-art methods.

Training ESRGAN with multi-scale attention U-Net discriminator

In this paper, we propose MSA-ESRGAN, a novel super-resolution model designed to enhance the perceptual quality of images. The key innovation of our approach lies in the integration of a multi-scale attention U-Net discriminator, which allows for more accurate differentiation between subject and background areas in images. By leveraging this architecture, MSA-ESRGAN surpasses traditional methods and several state-of-the-art super-resolution models in terms of Natural Image Quality Evaluator (NIQE) scores as well as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) across various benchmark datasets, including BSD100, Set5, Set14, Urban100, and OST300. Additionally, subjective evaluations further confirm the enhanced visual quality delivered by MSA-ESRGAN, particularly in terms of preserving texture and overall image realism. To ensure a fair comparison with Real-ESRGAN, we initialized our generator with a pre-trained Real-ESRNET model and followed the same training setup. Our model was trained on the DIV2K dataset using high-resolution image patches and the Adam optimizer, incorporating exponential moving average (EMA) for stability and performance enhancement. Evaluations on multiple benchmark datasets demonstrate that MSA-ESRGAN consistently delivers superior perceptual quality, as evidenced by higher NIQE, PSNR, and SSIM scores compared to other methods. Specifically, our model shows significant improvements in both objective and subjective measures of image quality. Furthermore, an ablation study highlighted the critical role of our multi-scale attention U-Net discriminator in enhancing the model’s performance. The results underscore the effectiveness of MSA-ESRGAN in maintaining image naturalness and perceptual quality, providing a robust benchmark for blind super-resolution tasks.

An weak surface defect inspection approach using efficient multi-scale attention and space-to-depth convolution network

In the field of precision manufacturing, machine vision technology is gradually replacing traditional manual inspection methods as a key technology to improve product quality. In precision manufacturing companies, weak defects on the product surface are unacceptable. However, existing defect detection methods rarely focus on the weak surface defect detection task. To address this challenge, we acquire and build a dataset called USB-DET, which contains weak defect samples. Then, we propose an innovative lightweight deep learning model, SDIA-net, which integrates SPD-Conv, Dysample technique, and attention mechanism-iRMA, to improve the recognition and localization of weak defects effectively. On the USB-DET dataset, SDIA-net achieves 55.1% mAP, which is 3.2% higher than the existing SOTA models. The computational efficiency is 205.1 FPS, which satisfies real-time demands. SDIA-net’s advantages make it well-suited for deployment in resource-limited precision manufacturing environments, providing an effective technical solution for product surface quality control with significant practical application value.

Lightweight multi-scale attention group fusion structure for nuclei segmentation

Lightweight multi-scale attention group fusion structure for nuclei segmentation

One-stage detection for unsupervised domain adaptation with efficient multi-scale attention and confidence-augmented combination

One-stage detection for unsupervised domain adaptation with efficient multi-scale attention and confidence-augmented combination

End-to-End CNN conceptual model for a biometric authentication mechanism for ATM machines

People commonly use automated teller machines, also known as ATMs, in contemporary culture. The use of ATMs for currency withdrawal is becoming increasingly widespread in the modern world. Furthermore, attacks and break-ins have increased, highlighting the need for an ATM that is more secure and equipped with extra safety measures. The existing method of confirming these transactions through credit cards and ATMs carries a high level of risk. Card trapping, cash larceny, shoulder surfing, and ATM skimming are among the examples. The proposed system, which is based on biometrics and deep convolutional neural networks, can effectively resolve the problems associated with conventional ATM cards and PINs. The proposed method relies on face biometrics, which offers a novel technique to address the current challenge. This article presents a novel end-to-end multiscale attention-based light-weighted deep convolutional neural network (MA-LW-DCNN) framework for enhancing the accuracy of recognizing faces. The multiscale attention mechanism in face recognition enhances feature extraction by applying convolution filters at various scales, allowing the network to capture detailed and global facial features. It generates an attention map by normalizing characteristic maps and refining them using activation functions, improving recognition accuracy by focusing on essential face regions. The suggested system is assessed using a synthetic dataset against conventional machine and deep learning-based face recognition techniques. The experiment results for different individuals demonstrate an accuracy rate of around 99.84% in authenticating test samples. The proposed approach produced significantly improved evaluation measures precision, recall, and loss by 0.99, 0.99, and 0.16, respectively.

ME-FCN: A Multi-Scale Feature-Enhanced Fully Convolutional Network for Building Footprint Extraction

The precise extraction of building footprints using remote sensing technology is increasingly critical for urban planning and development amid growing urbanization. However, considering the complexity of building backgrounds, diverse scales, and varied appearances, accurately and efficiently extracting building footprints from various remote sensing images remains a significant challenge. In this paper, we propose a novel network architecture called ME-FCN, specifically designed to perceive and optimize multi-scale features to effectively address the challenge of extracting building footprints from complex remote sensing images. We introduce a Squeeze-and-Excitation U-Block (SEUB), which cascades multi-scale semantic information exploration in shallow feature maps and incorporates channel attention to optimize features. In the network’s deeper layers, we implement an Adaptive Multi-scale feature Enhancement Block (AMEB), which captures large receptive field information through concatenated atrous convolutions. Additionally, we develop a novel Dual Multi-scale Attention (DMSA) mechanism to further enhance the accuracy of cascaded features. DMSA captures multi-scale semantic features across both channel and spatial dimensions, suppresses redundant information, and realizes multi-scale feature interaction and fusion, thereby improving the overall accuracy and efficiency. Comprehensive experiments on three datasets demonstrate that ME-FCN outperforms mainstream segmentation methods.

An integrated deep learning model for intelligent recognition of long-distance natural gas pipeline features

Pipeline feature recognition is crucial for the reliability and safety of long-distance natural gas pipelines. Utilizing manual or machine learning methods to recognize pipeline features is not only inefficient, but also has problems such as high misjudgment rate and poor robustness. To overcome the above problems, this paper proposes a pipeline feature recognition method based on Multi-scale Attention Convolutional Neural Network (MACNN) and Gated_Twins_Transformer. MACNN is used to extract multi-scale information of pipeline features, and then the attention mechanism in it to focus on the more important feature information and suppress the less important feature information. It is then transmitted to the Gated_Twins_Transformer model, which uses the gated mechanism and the twins encoder module to determine the importance of the data length and input dimensions, focusing on the feature information of both with different weights, and the Transformer enhances the extraction of global features. Finally, the measured pipeline bending strain data are used as model input, trained and tested, and compared with other advanced models, the superiority of the proposed model in this paper is verified by comparing the metrics of Accuracy, Precision, Recall and F1-score.

ALFusion: Adaptive fusion for infrared and visible images under complex lighting conditions

In the task of infrared and visible image fusion, source images often exhibit complex and variable characteristics due to scene illumination. To address the challenges posed by complex lighting conditions and enhance the quality of fused images, we develop the ALFusion method, which combines dynamic convolution and Transformer for infrared and visible image fusion. The core idea is to utilize the adaptability of dynamic convolution coupled with the superior long-range modeling capabilities of the Transformer to design a hybrid feature extractor. This extractor dynamically and comprehensively captures features from source images under various lighting conditions. To enhance the effectiveness of mixed features, we integrate an elaborate multi-scale attention enhancement module into the skip connections of the U-Net architecture. In this module, we use convolutional kernels of various sizes to expand the receptive field and incorporate an attention mechanism to enhance and highlight the combined use of dynamic convolution and the Transformer. Considering the performance in advanced visual tasks, an illumination detection network is integrated into the loss function, strategically balancing the pixel fusion ratio and optimizing visual impacts under varied lighting conditions, leveraging information from different source images. The fusion results indicate that our method consistently delivers superior background contrast and enhanced texture and structural features across diverse lighting conditions. Ablation and complementary experiments have been conducted, further affirming the effectiveness of our proposed method and highlighting its potential in advanced visual tasks.

A dimension-enhanced residual multi-scale attention framework for identifying anomalous waveforms of fault recorders

The continuous introduction of technologies such as distributed generation, wind power, and photovoltaic energy poses challenges to identifying abnormal waveforms in power disturbances. Due to the constant increase in abnormal features, existing waveform recognition schemes for power disturbance abnormalities cannot meet the requirements of high accuracy and reliability. In this paper, a Dimension-Enhanced Residual Multi-Scale Attention Framework for identifying power disturbance abnormal waveforms is proposed. This framework first employs the Phase Adaptive Adjustment (PAA) method to address the phase offset problem of original recording data, then uses the Gramian Angle Field method to perform dimensionality expansion on the data processed by PAA, and finally utilizes the Residual Pyramid Squeeze Attention Network (ResPSANet) for identifying power disturbance abnormal waveforms. Experiments demonstrate that the proposed approach improves the performance of power disturbance abnormal waveform recognition by 10% compared to existing schemes.

A Novel Multi-Scale Feature Enhancement U-Shaped Network for Pixel-Level Road Crack Segmentation

Timely and accurate detection of pavement cracks, the most common type of road damage, is essential for ensuring road safety. Automatic image segmentation of cracks can accurately locate their pixel positions. This paper proposes a Multi-Scale Feature Enhanced U-shaped Network (MFE-UNet) for pavement crack detection. This network model uses a Residual Detail-Enhanced Block (RDEB) instead of a conventional convolution in the encoder–decoder process. The block combines Efficient Multi-Scale Attention to enhance its feature extraction performance. The Multi-Scale Gating Feature Fusion (MGFF) is incorporated into the skip connections, enhancing the fusion of multi-scale features to capture finer crack details while maintaining rich semantic information. Furthermore, we created a pavement crack image dataset named China_MCrack, consisting of 1500 images collected from road surfaces using smartphone-mounted motorbikes. The proposed network was trained and tested on the China_MCrack, DeepCrack, and Crack-Forest datasets, with additional generalization experiments on the BochumCrackDataset. The results were compared with those of the U-Net model, ResUNet, and Attention U-Net. The experimental results show that the proposed MFE-UNet model achieves accuracies of 82.95%, 91.71%, and 69.02% on three datasets, namely, China_MCrack, DeepCrack, and Crack-Forest datasets, respectively, and the F1_score is improved by 1–4% compared with other networks. Experimental results demonstrate that the proposed method is effective in detecting cracks at the pixel level.

Research on multi-scale fusion image enhancement and improved YOLOv5s lightweight ROV underwater target detection method

Underwater target detection technology is very important for Marine resources detection and underwater environment perception. The problems of low-quality underwater images, a large amount of calculation, and low accuracy of target detection models still need to be solved. Firstly, a new underwater image enhancement algorithm based on multi-scale fusion is proposed. The algorithm combines the improved white balance algorithm and the improved CLAHE (Contrast Limited Adaptive Histogram Equalization) and realizes the multi-scale fusion strategy by introducing different weight maps, aiming to improve the visual effect and quality of underwater images. Secondly, to further improve the accuracy and efficiency of underwater target detection, an improved YOLOv5s lightweight underwater target detection algorithm is proposed. The algorithm combines the Ghost convolution module, EMA(Efficient Multi-Scale Attention) mechanism, and CARAFE (Content-Aware ReAssembly of FEatures) up-sampling mode. The comparison on the URPC(Underwater Robot Perception Challenge) public dataset shows that the proposed algorithm has achieved significant improvement in terms of model size, parameter amount, calculation amount, and mAP@0.5. Thirdly, to verify the effectiveness of the image enhancement algorithm and underwater target detection algorithm proposed in this paper, the ROV(Remote Operated Vehicle) is used to conduct experiments in the experimental pool. By designing underwater target detection experiments before and after image enhancement, the results show that the enhanced image significantly improves the detection ability of underwater targets. Finally, underwater target detection software is developed to provide strong support for the application of underwater-related fields.

RSG-YOLOV8: Detection of rice seed germination rate based on enhanced YOLOv8 and multi-scale attention feature fusion.

The lack of obvious difference between germinated seeds and non-germinated seeds will cause the low accuracy of detecting rice seed germination rate, remains a challenging issue in the field. In view of this, a new model named Rice Seed Germination-YOLOV8 (RSG-YOLOV8) is proposed in this paper. This model initially incorporates CSPDenseNet to streamline computational processes while preserving accuracy. Furthermore, the BRA, a dynamic and sparse attention mechanism is integrated to highlight critical features while minimizing redundancy. The third advancement is the employment of a structured feature fusion network, based on GFPN, aiming to reconfigure the original Neck component of YOLOv8, thus enabling efficient feature fusion across varying levels. An additional detection head is introduced, improving detection performance through the integration of variable anchor box scales and the optimization of regression losses. This paper also explores the influence of various attention mechanisms, feature fusion techniques, and detection head architectures on the precision of rice seed germination rate detection. Experimental results indicate that RSG-YOLOV8 achieves a mAP50 of 0.981, marking a 4% enhancement over the mAP50 of YOLOv8 and setting a new benchmark on the RiceSeedGermination dataset for the detection of rice seed germination rate.

A Lightweight Cotton Field Weed Detection Model Enhanced with EfficientNet and Attention Mechanisms

Cotton is a crucial crop in the global textile industry, with major production regions including China, India, and the United States. While smart agricultural mechanization technologies, such as automated irrigation and precision pesticide systems, have improved crop management, weeds remain a significant challenge. These weeds not only compete with cotton for nutrients but can also serve as hosts for diseases, affecting both cotton yield and quality. Existing weed detection models perform poorly in the complex environment of cotton fields, where the visual features of weeds and crops are similar and often overlap, resulting in low detection accuracy. Furthermore, real-time deployment on edge devices is difficult. To address these issues, this study proposes an improved lightweight weed detection model, YOLO-WL, based on the YOLOv8 architecture. The model leverages EfficientNet to reconstruct the backbone, reducing model complexity and enhancing detection speed. To compensate for any performance loss due to backbone simplification, CA (cross-attention) is introduced into the backbone, improving feature sensitivity. Finally, AFPN (Adaptive Feature Pyramid Network) and EMA (efficient multi-scale attention) mechanisms are integrated into the neck to further strengthen feature extraction and improve weed detection accuracy. At the same time, the model maintains a lightweight design suitable for deployment on edge devices. Experiments on the CottonWeedDet12 dataset show that the YOLO-WL model achieved an mAP of 92.30%, reduced the detection time per image by 75% to 1.9 ms, and decreased the number of parameters by 30.3%. After TensorRT optimization, the video inference time was reduced from 23.134 ms to 2.443 ms per frame, enabling real-time detection in practical agricultural environments.

Automatic Segmentation and Evaluation of Mitral Regurgitation Using Doppler Echocardiographic Images.

Mitral Regurgitation (MR) is a common heart valve disease. Severe MR can lead to pulmonary hypertension, cardiac arrhythmia, and even death. Therefore, early diagnosis and assessment of MR severity are crucial. In this study, we propose a deep learning-based method for segmenting MR regions, aiming to improve the efficiency of MR severity classification and diagnosis. We enhanced the Efficient Multi-Scale Attention (EMA) module to capture multi-scale features more effectively, thereby improving its segmentation performance on MR regions, which vary widely in size. A total of 367 color Doppler echocardiography images were acquired, with 293 images used for model training and 74 images for testing. To fully validate the capability of the improved EMA module, we use ResUNet as the backbone, partially integrating the enhanced EMA module into the decoder's upsampling process. The proposed model is then compared with classic models like Deeplabv3+ and PSPNet, as well as UNet, ResUNet, ResUNet with the original EMA module added, and UNet with the improved EMA module added. The experimental results demonstrate that the model proposed in this study achieved the best performance for the segmentation of the MR region on the test dataset: Jaccard (84.37%), MPA (92.39%), Recall (90.91%), and Precision (91.9%). In addition, the classification of MR severity based on the segmentation mask generated by our proposed model also achieved acceptable performance: Accuracy (95.27%), Precision (88.52%), Recall (91.13%), and F1-score (90.30%). The model proposed in this study achieved accurate segmentation of MR regions, and based on its segmentation mask, automatic and accurate assessment of MR severity can be realized, potentially assisting radiologists and cardiologists in making decisions about MR.

Enhanced HS Code Classification for Import and Export Goods via Multiscale Attention and ERNIE-BiLSTM

Accurate classification of import and export goods’ Harmonized System (HS) codes is essential for ensuring tax security. Applying text classification technologies for HS code classification can significantly enhance the prevention and control of customs tax risks. However, the goods text is a semi-structured one that involves multi-domain Chinese professional vocabulary, which poses challenges for current classification models. These models often suffer from inadequate text representation and imprecise feature extraction. To address these challenges, we propose a novel classification model ERNIE-BiLSTM-Channel attention–Spatial attention (EBLCS). This model integrates ERNIE (Enhanced Representation through Knowledge Integration) with a Bidirectional Long Short-Term Memory Network (BiLSTM) and employs multi-scale attention mechanisms. The ERNIE-BiLSTM model provides a more comprehensive and accurate representation of the goods text, effectively capturing the global features of the text. By introducing channel attention and spatial attention mechanisms, greater weights are assigned to important words and word embedding dimensions, significantly enhancing the model’s ability to perceive key information. The experimental results on a customs dataset demonstrate that the EBLCS model consistently outperforms various baseline models across all evaluation metrics, effectively enhancing the performance of HS code classification.