Multi-scale Dilated Convolution Research Articles

Underwater small object detection via automatic encoding pyramid with neighborhood information

Abstract With the rapid advancement of Internet of Things (IoT) technology, underwater sensor networks have shown significant potential in marine resource exploration, environmental monitoring, and related fields. However, detecting small objects in underwater images remains challenging due to complex conditions and the diminutive size of targets. The sparse features of small objects, combined with the pyramid structure of pooling and down-sampling, lead to the gradual loss of pixel-level features during propagation, thereby limiting detection performance. Moreover, the pyramid-style feature extraction network introduces information conflicts between features of different scales when integrating multi-density information, further degrading the detection accuracy of small objects in underwater images. To address these issues, we propose a novel feature pyramid network, termed AENI-USOD, which leverages neighborhood information and generates conflict suppression information to enhance the network’s focus on underwater objects. Specifically, we design a Neighborhood Information Fusion Module (NIFM) that employs multi-scale dilated convolutions to fuse features and generate neighborhood information for more precise object detection. To mitigate information conflicts during multi-scale feature fusion, we introduce a Conflict Suppression Module (CSM) that filters out conflicting interference, preventing small objects from being obscured by redundant information. Additionally, we implement the NMS-Circle post-processing method, which utilizes a circular feature attenuation function to reduce reliance on threshold parameters in soft non-maximum suppression. Our method is extensively evaluated on the UDD, RUOD, and URPC2019 datasets, demonstrating superior accuracy compared to state-of-the-art methods. This work provides robust technical support for intelligent object recognition within underwater IoT systems.

FAD-Net: Automated Framework for Steel Surface Defect Detection in Urban Infrastructure Health Monitoring

Steel plays a fundamental role in modern smart city development, where its surface structural integrity is decisive for operational safety and long-term sustainability. While deep learning approaches show promise, their effectiveness remains limited by inadequate receptive field adaptability, suboptimal feature fusion strategies, and insufficient sensitivity to small defects. To overcome these limitations, we propose FAD-Net, a deep learning framework specifically designed for surface defect detection in steel materials within urban infrastructure. The network incorporates three key innovations: The RFCAConv module, which leverages dynamic receptive field construction and coordinate attention mechanisms to enhance feature representation for defects with long-range spatial dependencies and low-contrast characteristics. The MSDFConv module, employing multi-scale dilated convolutions with optimized dilation rates to preserve fine details while expanding the receptive field. An Auxiliary Head that introduces hierarchical supervision to improve the detection of small-scale defects. Experiments on the GC10-DET dataset showed that FAD-Net achieved 5.0% higher mAP@0.5 than baseline models. Cross-dataset validation with NEU and RDD2022 further confirmed its robustness. These results demonstrate FAD-Net’s effectiveness for automated infrastructure health monitoring.

MSFHNet: a hybrid deep learning network for multi-scale spatiotemporal feature extraction of spatial cognitive EEG signals in BCI-VR systems.

The integration of brain-computer interface (BCI) and virtual reality (VR) systems offers transformative potential for spatial cognition training and assessment. By leveraging artificial intelligence (AI) to analyze electroencephalogram (EEG) data, brain activity patterns during spatial tasks can be decoded with high precision. In this context, a hybrid neural network named MSFHNet is proposed, optimized for extracting spatiotemporal features from spatial cognitive EEG signals. The model employs a hierarchical architecture where its temporal module uses multi-scale dilated convolutions to capture dynamic EEG variations, while its spatial module integrates channel-spatial attention mechanisms to model inter-channel dependencies and spatial distributions. Cross-stacked modules further refine discriminative features through deep-level fusion. Evaluations demonstrate the superiority of MSFHNet in the beta2 frequency band, achieving 98.58% classification accuracy and outperforming existing models. This innovation enhances EEG signal representation, advancing AI-powered BCI-VR systems for robust spatial cognitive training.

Mural inpainting via two-stage generative adversarial network

The digital restoration of Dunhuang murals is of extremely high value for the research and dissemination of mural culture and art. However, a large number of murals have been damaged to varying degrees. In this paper, we propose a two-stage coarse-to-fine digital mural restoration framework to solve the large area of irregular shape damage on the mural. The first stage is used to achieve coarse-grained semantic reconstruction, and the second stage is used to achieve fine-grained feature reconstruction. To improve the repair quality, we also designed a new building block (STMA) that integrates the Swin transformer module (SwinT) and the multi-scale dilated convolution attention module (MSDA). Meanwhile, the proposed loss function is to further empower the proposed model to repair damaged murals. Extensive comparative experiments show that the proposed model can effectively restore the missing content of the mural and exceed the comparative methods in both quantitative and qualitative evaluation.

Few-Shot Counting with Multi-Scale Vision Transformers and Attention Mechanisms

Object counting is a fundamental task in computer vision, with critical applications in areas such as crowd monitoring and ecological conservation. Traditional methods typically rely on large-scale annotated datasets, which are costly and time-consuming to obtain. Few-shot object counting has emerged as a promising solution, enabling accurate counting with minimal annotated samples. However, in real-world scenarios, objects often exhibit significant scale variations due to factors such as view distortion, varying shooting distances, and inherent size differences. Existing few-shot methods usually struggle to address this challenge effectively. To address these issues, we propose a Scale-Aware Vision Transformer (SAViT) framework. Specifically, we design a multi-scale dilated convolution module in SAViT, which can adaptively adjust convolution kernel sampling rates to handle objects of varying sizes. Additionally, we incorporate a global channel attention mechanism to strengthen the model’s ability to capture robust feature representations, thereby improving detection accuracy. For practical usability, we integrate the Segment Anything Model (SAM) to create an exemplar box selection module, simplifying the process by allowing users to generate precise exemplar boxes with a single line drawn on the target object. Extensive experiments on the FSC-147 dataset demonstrate the effectiveness of our approach, achieving a Mean Absolute Error (MAE) of 8.92 and a Root Mean Squared Error (RMSE) of 31.26. Compared to the state-of-the-art method, CACViT, our model reduces MAE by 0.21 (2.30% improvement) and RMSE by 17.7 (36.15% improvement). Our approach not only provides an effective solution for few-shot object counting but also provides a new practical paradigm for extending few-shot learning to complex vision tasks requiring multi-scale reasoning. The code of our paper is available at https://github.com/BlouseDong/SAViT .

Small-Target Detection Algorithm Based on STDA-YOLOv8.

Due to the inherent limitations of detection networks and the imbalance in training data, small-target detection has always been a challenging issue in the field of target detection. To address the issues of false positives and missed detections in small-target detection scenarios, a new algorithm based on STDA-YOLOv8 is proposed for small-target detection. A novel network architecture for small-target detection is designed, incorporating a Contextual Augmentation Module (CAM) and a Feature Refinement Module (FRM) to enhance the detection performance for small targets. The CAM introduces multi-scale dilated convolutions, where convolutional kernels with different dilation rates capture contextual information from various receptive fields, enabling more accurate extraction of small-target features. The FRM performs adaptive feature fusion in both channel and spatial dimensions, significantly improving the detection precision for small targets. Addressing the issue of a significant disparity in the number of annotations between small and larger objects in existing classic public datasets, a new data augmentation method called Copy-Reduce-Paste is introduced. Ablation and comparative experiments conducted on the proposed STDA-YOLOv8 model demonstrate that on the VisDrone dataset, its accuracy improved by 5.3% compared to YOLOv8, reaching 93.5%; on the PASCAL VOC dataset, its accuracy increased by 5.7% compared to YOLOv8, achieving 94.2%, outperforming current mainstream target detection models and small-target detection algorithms like QueryDet, effectively enhancing small-target detection capabilities.

Sound signal-based lightweight fault diagnosis for railway turnout system via combining transformer and CNN

Abstract Significant progress has been made in deep learning-based fault diagnosis for railway turnout system. However, the increasing complexity of networks has resulted in higher network parameters and computational demands, making it necessary for algorithms to meet more stringent hardware requirements. Therefore, there is a need for a lightweight deep learning approach that can effectively diagnose faults in railway turnout system while maintaining high performance. In this research, we propose a lightweight model integrating Transformer and Convolutional Neural Networks (CNN) for fault diagnosis in railway turnout systems. The model combines CNN and Transformer to extract both local and global features from monitoring signals and introduces lightweight enhancements. Specifically, a Multi-Scale Depthwise Dilated Convolution (MDDC) block is introduced into the CNN to break down cross-channel convolutions into pointwise and depthwise convolutions, with multiscale dilated convolution kernels applied in the depthwise convolutions. Additionally, an Context-Aware Attention Module (CAAM) is integrated into the Transformer to simplify complex matrix multiplications and multidimensional exponential operations. Experimental results using sound signals collected during actuator operations demonstrate that the optimized algorithm reduces the number of parameters and computational complexity by 75% compared to the baseline model, while also reducing inference time by one-third. The optimized algorithm achieves an accuracy of 99.45%. These findings lay the groundwork for further optimization of fault diagnosis in railway turnout system.

SECrackSeg: A High-Accuracy Crack Segmentation Network Based on Proposed UNet with SAM2 S-Adapter and Edge-Aware Attention.

Crack segmentation is essential for structural health monitoring and infrastructure maintenance, playing a crucial role in early damage detection and safety risk reduction. Traditional methods, including digital image processing techniques have limitations in complex environments. Deep learning-based methods have shown potential, but still face challenges, such as poor generalization with limited samples, insufficient extraction of fine-grained features, feature loss during upsampling, and inadequate capture of crack edge details. This study proposes SECrackSeg, a high-accuracy crack segmentation network that integrates an improved UNet architecture, Segment Anything Model 2 (SAM2), MI-Upsampling, and an Edge-Aware Attention mechanism. The key innovations include: (1) using a SAM2 S-Adapter with a frozen backbone to enhance generalization in low-data scenarios; (2) employing a Multi-Scale Dilated Convolution (MSDC) module to promote multi-scale feature fusion; (3) introducing MI-Upsampling to reduce feature loss during upsampling; and (4) implementing an Edge-Aware Attention mechanism to improve crack edge segmentation precision. Additionally, a custom loss function incorporating weighted binary cross-entropy and weighted IoU loss is utilized to emphasize challenging pixels. This function also applies Multi-Granularity Supervision by optimizing segmentation outputs at three different resolution levels, ensuring better feature consistency and improved model robustness across varying image scales. Experimental results show that SECrackSeg achieves higher precision, recall, F1-score, and mIoU scores on the CFD, Crack500, and DeepCrack datasets compared to state-of-the-art models, demonstrating its excellent performance in fine-grained feature recognition, edge segmentation, and robustness.

DCN-YOLO: A Small-Object Detection Paradigm for Remote Sensing Imagery Leveraging Dilated Convolutional Networks.

With the rapid development of remote sensing technology, optical remote sensing images are increasingly being used in areas such as military reconnaissance, environmental monitoring, and urban planning. Due to the small number of pixels, fuzzy features, and complex background, it is difficult for conventional convolutions to effectively extract features from small objects. To address this problem, we propose to use multi-scale dilated convolutions to increase the receptive field size of the model to adapt to changes in object size, capture multi-scale contextual information of the feature map, and extract richer object features. First, we propose a Dilated Convolutional Residual (DCR) module for high-level feature extraction in the network. Second, the context aggregation (CONTEXT) module uses remote interaction to perform associative computation on images using contextual aggregation, allowing the model to understand the global semantic information of the image. We propose a novel object detection method, DCN-YOLO, which achieves an AP50 of 56.6 on the AI-TOD dataset, effectively improving the detection accuracy and robustness of small objects in remote sensing images. It provides a new technical approach to the detection of small objects in remote sensing.

Lightweight LDA-UNet for accurate and interpretable visualization of retinopathy of prematurity

Retinopathy of prematurity (ROP) is one of the main causes of blindness in children. Timely prevention and treatment can effectively control its progression. Although convolutional neural networks (CNNs) have been widely studied in the field of ROP detection, limitations such as parameter redundancy, spatial information loss, and limited interpretability persist. To address these issues, we proposed LDA-UNet, a lightweight CNN model that integrates depth separable convolution, multi-scale dilated convolution, and residual channel-spatial attention mechanisms. Our model significantly reduces parameters (by 85.4% compared to standard UNet) while alleviating the lack of spatial information loss during the encoding stage. The Dice coefficient is 0.927, which exceeds both the current typical lightweight UNet models and the Transformer-based hybrid structure TransUNet. By combining LDA-UNet with a contour detection algorithm, we introduce an object visualization model that precisely locates lesion areas and delineates contours, enhancing interpretability. Our approach outperforms YOLOv5s in mean Intersection over Union (mIoU) by 17.5% and object detection accuracy by 12.5%, underscoring its efficacy in ROP visualization. This work presents a robust, lightweight solution for ROP diagnosis, facilitating clinical decision-making and improving patient outcomes. Codes are available at https://github.com/honghaolu/ROP_Visualizaiton or https://doi.org/10.5281/zenodo.13922434.

A lightweight network for traffic sign detection via multiple scale context awareness and semantic information guidance

Traffic sign detection, as a critical branch of object detection, plays an essential role in both assisted driving and autonomous driving technologies. In this paper, we propose MASG-Net, a lightweight detection network designed to improve the accuracy and efficiency of traffic sign detection. First, we introduce a channel attention mechanism into MobileNetV3 to create a novel E-block structure and design E-mobilenet, a lightweight backbone network, to replace the backbone in YOLOv4-tiny, significantly enhancing feature extraction while reducing parameters. Second, we propose a multi-scale dilated convolution spatial pyramid pooling (MDSPP) module to expand the receptive field of feature maps, enabling the network to capture multi-scale contextual information effectively. Finally, a semantic information guidance (SIG) module is introduced to leverage deep semantic information to guide shallow feature layers, improving the detection of small traffic signs and enhancing robustness against cluttered backgrounds. Experimental results on the CCTSDB, GTSDB and TT100K datasets demonstrate that MASG-Net achieves superior detection performance, particularly for small and challenging traffic signs, while maintaining high efficiency with an inference speed of 203.6 FPS. These results highlight MASG-Net’s potential for real-time traffic sign detection in practical applications.

Dynamic Graph Attention Meets Multi-Scale Temporal Memory: A Hybrid Framework for Photovoltaic Power Forecasting Under High Renewable Penetration

In the context of the accelerated global energy transition, power fluctuations caused by the integration of a high share of renewable energy have emerged as a critical challenge to the security of power systems. The goal of this research is to improve the accuracy and reliability of short-term photovoltaic (PV) power forecasting by effectively modeling the spatiotemporal coupling characteristics. To achieve this, we propose a hybrid forecasting framework—GLSTM—combining graph attention (GAT) and long short-term memory (LSTM) networks. The model utilizes a dynamic adjacency matrix to capture spatial correlations, along with multi-scale dilated convolution to model temporal dependencies, and optimizes spatiotemporal feature interactions through a gated fusion unit. Experimental results demonstrate that GLSTM achieves RMSE values of 2.3%, 3.5%, and 3.9% for short-term (1 h), medium-term (6 h), and long-term (24 h) forecasting, respectively, and mean absolute error (MAE) values of 3.8%, 6.2%, and 7.0%, outperforming baseline models such as LSTM, ST-GCN, and Transformer by reducing errors by 10–25%. Ablation experiments validate the effectiveness of the dynamic adjacency matrix and the spatiotemporal fusion mechanism, with a 19% reduction in 1 h forecasting error. Robustness tests show that the model remains stable under extreme weather conditions (RMSE 7.5%) and data noise (RMSE 8.2%). Explainability analysis reveals the differentiated contributions of spatiotemporal features. The proposed model offers an efficient solution for high-accuracy renewable energy forecasting, demonstrating its potential to address key challenges in renewable energy integration.

Optimized UNet framework with a joint loss function for underwater image enhancement

As the water economy advances and the concepts of water ecology protection and sustainable development take root in people’s minds, underwater imaging equipment has made remarkable progress. However, due to various factors, underwater images still suffer from low quality. How to enhance the quality of underwater images so that people can understand them quickly has become a crucial issue. Therefore, aiming at the degradation problems such as detail blurring, color imbalance, and noise interference in low-quality underwater images, this paper proposes an optimized UNet framework with a joint loss function (OUNet-JL). Firstly, to alleviate the problem of detail blurring, we construct a multi-residual module (MRM) to enhance the ability to represent detail features by using serially stacked convolutional blocks and residual connections. Secondly, we build a spatial multi-scale feature extraction module fused with channel attention (SMFM) to address the color imbalance issue through multi-scale dilated convolution and channel attention. Thirdly, to improve the signal-to-noise ratio of the enhanced image and solve the problem of blurring distortion, a strengthen-operate-subtract feature reconstruction module (SOSFM) is presented. Fourthly, to guide the network to perform training more efficiently and help it converge rapidly, a joint loss function is designed by integrating four different loss functions. Extensive experiments conducted on the well-known UIEB and UFO-120 datasets have shown the superiority of our OUNet-JL compared with several state-of-the-art algorithms. Moreover, ablation studies have also verified the effectiveness of the proposed modules. Our source code is publicly available at https://github.com/WangXin81/OUNet_JL.

Research on High-Precision Gas Concentration Inversion for Imaging Fourier Transform Spectroscopy Based on Multi-Scale Feature Attention Model

The accurate monitoring of greenhouse gas (GHG) concentrations is crucial in mitigating global warming. The imaging Fourier transform spectrometer (IFTS) is an effective tool for measuring GHG concentrations, offering high throughput and a wide spectral measurement range. In order to address the issue of spectral inconsistency during the detection process of the target gas, which is influenced by external environmental factors, making it difficult to achieve high-precision gas concentration inversion, this paper proposes a multi-scale feature attention (MDISE) model. The model uses a multi-scale dilated convolution (MD) module to retain both global and local shallow features of the spectra; introduces the one-dimensional Inception (1D Inception) module to further extract multi-scale deep features; and incorporates the channel attention mechanism (SE) module to enhance attention to important spectral wavelengths, suppressing redundant and interfering information. A target gas detection system was built in the laboratory, and the proposed model was tested on gas samples collected by two channels of a short and medium-wavelength infrared imaging Fourier transform spectrometer (SMWIR-IFTS). The experimental results show that the MDISE model reduces the root mean square error (RMSE) in both channels by 79.14%, 76.59%, and 69.80%, and 81.45%, 82.65%, and 74.01%, respectively, compared to the partial least squares regression (PLSR), support vector regression (SVR), and conventional one-dimensional convolutional neural network (1D-CNN) models. Additionally, the MDISE model achieved average coefficient of determination (R2) values of 0.997 and 0.995 for the concentration intervals in both channels. The MDISE model demonstrates excellent performance and significantly improves the accuracy of GHG concentration inversion.

MITD-Net: Multi-scale iterative tamper detection network for image manipulation localization

MITD-Net: Multi-scale iterative tamper detection network for image manipulation localization

A generative adversarial network with multiscale and attention mechanisms for underwater image enhancement

Underwater images collected are often of low clarity and suffer from severe color distortion due to the marine environment and Illumination conditions. This directly impacts tasks such as marine ecological monitoring and underwater target detection, which rely on image processing. Therefore, enhancing Underwater images to improve their quality is necessary. A generative adversarial network with an encoder-decoder structure is proposed to improve the quality of Underwater images. The network consists of a generative network and an adversarial network. The generative network is responsible for enhancing the images, while the adversarial network determines whether the input is an enhanced image or a real high-quality image. In the generative network, we first design a residual convolution module to extract more texture and edge information from underwater images. Next, we design a multi-scale dilated convolution module to capture underwater features at different scales. Then, we design a feature fusion adaptive attention module to reduce the interference of redundant features and enhance the local perception capabilities. Finally, we construct the generative network using these modules along with conventional modules. In the adversarial network, we first design a multi-scale feature extraction module to improve the feature extraction ability. We then use the multi-scale feature extraction module along with conventional convolution modules to design the adversarial network. Additionally, we propose an improved loss function by introducing color loss into the conventional loss function. The improved loss function can better measure the color discrepancy between the enhanced image and the real image. It is useful to reduce color distortion in the enhanced images. In experimental simulations, the images enhanced by the proposed methods have the highest PSNR, SSIM, and UIQM values, indicating that the proposed method has superior Underwater image enhancement capabilities compared to other methods.

Two-Stage GPR Image Inversion Method Based on Multi-Scale Dilated Convolution and Hybrid Attention Gate

Ground penetrating radar (GPR) image inversion is of great significance for interpreting GPR data. In practical applications, the complexity and nonuniformity of underground structures bring noise and clutter interference, making GPR inversion problems more challenging. To address these issues, this study proposes a two-stage GPR image inversion network called MHInvNet based on multi-scale dilated convolution (MSDC) and hybrid attention gate (HAG). This method first denoises the B-scan through the first network MHInvNet1, then combines the denoised B-scan from MHInvNet1 with the undenoised B-scan as input to the second network MHInvNet2 for inversion to reconstruct the distribution of the permittivity of underground targets. To further enhance network performance, the MSDC and HAG modules are simultaneously introduced to both networks. Experimental results from simulated and actual measurement data show that MHInvNet can accurately invert the position, shape, size, and permittivity of underground targets. A comparison with existing methods demonstrates the superior inversion performance of MHInvNet.

Aluminum Electrolysis Fire-Eye Image Segmentation Based on the Improved U-Net Under Carbon Slag Interference

To solve the problem of low segmentation model accuracy due to the complex shape of carbon slag in the aluminum electrolysis fire-eye image and the blurring of the boundary between the slag and the surrounding electrolyte, this paper proposes a segmentation model of the fire-eye image based on an improved U-Net. The model reduces the depth of the traditional U-Net to four layers and uses the multiscale dilated convolution module (MDCM) in the down-sampling stage. Second, the Convolutional Block Attention Module (CBAM) is embedded in the skip connection part of the network to improve the ability of the model to extract contextual features from images of multiple scales, enhance the guidance of high-level features to low-level features, and make the model pay more attention to the critical regions. To alleviate the negative impact of the imbalance of positive and negative examples in the dataset, the weighted binary cross-entropy loss and the Dice loss are used to replace the traditional cross-entropy loss. The experimental results show that the segmentation accuracy of the improved model on the fire-eye dataset reaches 88.03%, which is 5.61 percentage points higher than U-Net.

Cascaded Feature Fusion Grasping Network for Real-Time Robotic Systems.

Grasping objects of irregular shapes and various sizes remains a key challenge in the field of robotic grasping. This paper proposes a novel RGB-D data-based grasping pose prediction network, termed Cascaded Feature Fusion Grasping Network (CFFGN), designed for high-efficiency, lightweight, and rapid grasping pose estimation. The network employs innovative structural designs, including depth-wise separable convolutions to reduce parameters and enhance computational efficiency; convolutional block attention modules to augment the model's ability to focus on key features; multi-scale dilated convolution to expand the receptive field and capture multi-scale information; and bidirectional feature pyramid modules to achieve effective fusion and information flow of features at different levels. In tests on the Cornell dataset, our network achieved grasping pose prediction at a speed of 66.7 frames per second, with accuracy rates of 98.6% and 96.9% for image-wise and object-wise splits, respectively. The experimental results show that our method achieves high-speed processing while maintaining high accuracy. In real-world robotic grasping experiments, our method also proved to be effective, achieving an average grasping success rate of 95.6% on a robot equipped with parallel grippers.

Construction of tennis pose estimation and action recognition model based on improved ST-GCN

With the rapid growth of computer vision and deep learning technologies, the application of pose estimation and action recognition in sports training has become increasingly widespread. Due to factors such as complex movements, fast speed, and limb occlusion, pose estimation and action recognition in tennis face significant challenges. Therefore, this study first introduces selective dropout and pyramid region of interest pooling layer strategies in fast region convolutional neural networks. Secondly, a pose estimation algorithm based on multi-scale fusion pose residual network 50 is designed, and finally a spatiotemporal graph convolutional network model is constructed by fusing channel attention module and multi-scale dilated convolution module. The data showed that the average detection accuracy of the improved attitude residual network 50 was 70.4%, and the accuracy of object detection for small, medium, and large objects was 57.4%, 69.3%, and 79.2%, respectively. The continuous action recognition accuracy and inter action fluency detection time of the improved spatiotemporal graph convolutional network were 93.8% and 19.2 ms, respectively. When the sample size was 1000, its memory usage was 1378 MB and the running time was 32.7 ms. Experiments have shown that the improved model achieves high accuracy and robustness in tennis action recognition tasks, especially in complex scenes and limb occlusion conditions, where the model shows significant advantages. This study aims to provide an efficient and accurate motion recognition technology for tennis posture analysis and intelligent training.