Feature Fusion Module Research Articles

Human pose estimation in complex background videos via Transformer-based multi-scale feature integration

Human posture estimation is still a hot research topic. Previous algorithms based on traditional machine learning have difficulties in feature extraction and low fusion efficiency. To address these problems, we proposed a Transformer-based method. We combined three techniques, namely the Transformer-based feature extraction module, the multi-scale feature fusion module, and the occlusion processing mechanism, to capture the human pose. The Transformer-based feature extraction module uses the self-attention mechanism to extract key features from the input sequence, the multi-scale feature fusion module fuses feature information of different scales to enhance the perception ability of the model, and the occlusion processing mechanism can effectively handle occlusion in the data and effectively remove background interference. Our method has shown excellent performance through verification on the standard dataset Human3.6M and the wild video dataset, achieving accurate pose prediction in both complex actions and challenging samples.

DEAF-Net: Detail-Enhanced Attention Feature Fusion Network for Retinal Vessel Segmentation.

Retinal vessel segmentation is crucial for the diagnosis of ophthalmic and cardiovascular diseases. However, retinal vessels are densely and irregularly distributed, with many capillaries blending into the background, and exhibit low contrast. Moreover, the encoder-decoder-based network for retinal vessel segmentation suffers from irreversible loss of detailed features due to multiple encoding and decoding, leading to incorrect segmentation of the vessels. Meanwhile, the single-dimensional attention mechanisms possess limitations, neglecting the importance of multidimensional features. To solve these issues, in this paper, we propose a detail-enhanced attention feature fusion network (DEAF-Net) for retinal vessel segmentation. First, the detail-enhanced residual block (DERB) module is proposed to strengthen the capacity for detailed representation, ensuring that intricate features are efficiently maintained during the segmentation of delicate vessels. Second, the multidimensional collaborative attention encoder (MCAE) module is proposed to optimize the extraction of multidimensional information. Then, the dynamic decoder (DYD) module is introduced to preserve spatial information during the decoding process and reduce the information loss caused by upsampling operations. Finally, the proposed detail-enhanced feature fusion (DEFF) module composed of DERB, MCAE and DYD modules fuses feature maps from both encoding and decoding and achieves effective aggregation of multi-scale contextual information. The experiments conducted on the datasets of DRIVE, CHASEDB1, and STARE, achieving Sen of 0.8305, 0.8784, and 0.8654, and AUC of 0.9886, 0.9913, and 0.9911 on DRIVE, CHASEDB1, and STARE, respectively, demonstrate the performance of our proposed network, particularly in the segmentation of fine retinal vessels.

YOSMR: A Ship Detection Method for Marine Radar Based on Customized Lightweight Convolutional Networks

In scenarios such as nearshore and inland waterways, the ship spots in a marine radar are easily confused with reefs and shorelines, leading to difficulties in ship identification. In such settings, the conventional ARPA method based on fractal detection and filter tracking performs relatively poorly. To accurately identify radar targets in such scenarios, a novel algorithm, namely YOSMR, based on the deep convolutional network, is proposed. The YOSMR uses the MobileNetV3(Large) network to extract ship imaging data of diverse depths and acquire feature data of various ships. Meanwhile, taking into account the issue of feature suppression for small-scale targets in algorithms composed of deep convolutional networks, the feature fusion module known as PANet has been subject to a lightweight reconstruction leveraging depthwise separable convolutions to enhance the extraction of salient features for small-scale ships while reducing model parameters and computational complexity to mitigate overfitting problems. To enhance the scale invariance of convolutional features, the feature extraction backbone is followed by an SPP module, which employs a design of four max-pooling constructs to preserve the prominent ship features within the feature representations. In the prediction head, the Cluster-NMS method and α-DIoU function are used to optimize non-maximum suppression (NMS) and positioning loss of prediction boxes, improving the accuracy and convergence speed of the algorithm. The experiments showed that the recall, accuracy, and precision of YOSMR reached 0.9308, 0.9204, and 0.9215, respectively. The identification efficacy of this algorithm exceeds that of various YOLO algorithms and other lightweight algorithms. In addition, the parameter size and calculational consumption were controlled to only 12.4 M and 8.63 G, respectively, exhibiting an 80.18% and 86.9% decrease compared to the standard YOLO model. As a result, the YOSMR displays a substantial advantage in terms of convolutional computation. Hence, the algorithm achieves an accurate identification of ships with different trail features and various scenes in marine radar images, especially in different interference and extreme scenarios, showing good robustness and applicability.

TSEDNet:Task-specific encoder–decoder network for surface defects of strip steel

Deep learning faces challenges in the surface defect segmentation of strip steel. Firstly, insufficient processing of feature maps leads to the loss of task-specific feature information. Secondly, the segmentation of defects with long-tail distributions is not accurate enough. To address these issues, a pixel-level deep segmentation method called task-specific encoder–decoder network (TSEDNet) is proposed to construct an end-to-end defect segmentation model. TSEDNet includes the encoder-multi-decoder structure based on domain knowledge settings tailored to specific tasks, which can achieve effective feature representation and significantly reduce the impact of imbalanced defect quantities. Additionally, a novel metric learning method is introduced to optimize decoder selection. Furthermore, the feature fusion module based on metric learning is proposed to utilize general features for restoring task-specific details, thereby enhancing pixel-level segmentation accuracy. Through experiments and industrial validation, the defect segmentation network demonstrates superior performance compared to other advanced segmentation methods and proves its applicability in practical scenarios.

Direction-aware multi-branch attention and Gaussian label assignment for remote sensing aggregative object detection

ABSTRACT A large number of objects of different scales appear aggregative distribution in remote sensing images, which brings two challenges to remote sensing object detection. The first challenge arises from the presence of objects with large scale differences within the same scene, which complicates the detector’s ability to balance the detection performance between large and small objects. The second challenge is that the distribution of dense objects on remote sensing images with complex backgrounds can reduce the feature representation capability of the network, leading to false detections and missed detections of objects. To address the two challenges present in remote sensing object detection, we propose a new object detection network DA-GLANet (Direction-Aware Multi-Branch Attention and Gaussian Label Assignment Balance Net). For the first challenge, we designed a multi-scale feature fusion module to fully utilize semantic information at different scales, thereby improving the detection accuracy of multi-scale objects. Subsequently, to tackle the second challenge, we proposed a multi-branch attention module with direction-aware capabilities. This module can enhance the network’s feature representation of objects of interest by generating attention maps that are both direction-aware and position-sensitive. In addition, we propose a label assignment strategy based on Gaussian distribution to address the imbalance of positive and negative labels caused by the complexity of remote sensing images. These methods work synergistically to improve aggregative object detection accuracy. We have conducted extensive experiments on the DOTA dataset, Levir dataset, and WHU buildings change detection dataset to demonstrate the superiority of our method. On the DOTA dataset, our method improves mAP by 0.4%, 0.3%, and 0.2% compared to the top three methods. On the Levir dataset, our method improves mAP by 2.5%, 2.2%, and 1.6% compared to the top three methods.

A multi-scale attributes fusion model for travel mode identification using GPS trajectories

ABSTRACT Travel mode recognition is a key issue in urban planning and transportation research. While traditional travel surveys use manual data collection and have limited coverage, poor timeliness, and insufficient sample capacity, recent advancements in Global Positioning System (GPS) technology allow large-scale data collection and offer novel opportunities to enhance travel mode recognition. However, existing studies often neglect regular differences and changes in motion states across different travel modes and fail to fully integrate multi-scale spatio-temporal features, which limits the accurate classification of travel modes. To fill this gap, this study proposes a multi-scale spatio-temporal attribute fusion (MSAF) model for precise travel mode identification using solely GPS trajectories without altering their sampling rate. The MSAF model segments GPS trajectories into various temporal and spatial scales, extracting local motion states and spatial features at multiple scales. The spatio-temporal feature extraction module is constructed to extract local motion states and capture spatio-temporal dependencies. Additionally, the model incorporates a multi-scale feature fusion module, which effectively combines features of various scales through a series of fusion techniques to obtain a comprehensive representation, enabling automatic and accurate travel mode identification. Experiments on real-world datasets, including the GeoLife Trajectories dataset and the Sussex-Huawei Locomotion-Transportation (SHL) dataset, demonstrate the effectiveness of the MSAF model, achieving a competitive accuracy of 95.16% and 91.70%. This represents an improvement of 2.50% to 7.95% and 0.8% to 6.62% over several state-of-the-art baselines, effectively addressing sample imbalance challenges. Moreover, the experiments demonstrate the significant role of multiscale feature fusion in improving model performance.

PSO-based lightweight neural architecture search for object detection

In recent years, Neural Architecture Search (NAS) has received widespread attention for its remarkable ability to design neural networks. However, existing NAS works have mainly focused on network search, with limited emphasis on downstream applications after discovering efficient neural networks. In this paper, we propose a lightweight search strategy based on the particle swarm optimization algorithm and apply the searched network as backbone for object detection tasks. Specifically, we design a lightweight search space based on Ghostconv modules and improved Mobileblocks, achieving comprehensive exploration within the search space using variable-length encoding strategy. During the search process, to balance network performance and resource consumption, we propose a multi-objective fitness function and incorporated the classification accuracy, parameter size, and FLOPs of candidate individuals into optimization. For particle performance evaluation, we propose a new strategy based on weight sharing and dynamic early stopping, significantly accelerating the search process. Finally, we fine-tune the globally optimal particle decoded as the backbone, adding Ghost PAN feature fusion modules and detection heads to build an object detection model, and we achieve a 17.01% mAP on the VisDrone2019 dataset. Experimental results demonstrate the competitiveness of our algorithm in terms of search time and the balance between accuracy and efficiency, and also confirm the effectiveness of object detection models designed through NAS methods.

Global and local complementary multi-path feature fusion network for the classification of crop remote sensing images

Abstract The accuracy and efficiency of crop distribution information extraction are pivotal in ensuring global food security. In long-time-series optical satellite data, most existing methods focus on extracting spatial features using Convolutional Neural Networks (CNNs), which do not adequately mine and model the spatial-temporal information. The development of the attention mechanism allows for the extraction of global features in remote-sensing images of long temporal sequences. To extract global attentional features with complementary features in crop remote sensing images, we propose a Global and Local Complementary Multi-path Feature Fusion Network (GLMP), which is capable of extracting global features from remote sensing images of long temporal sequence, that enhances the local characteristics of crop images derived from CNNs, thus obtaining more effective multi-scale complementary features. This extraction of features enhances the comprehension of crop images, thereby boosting the performance of associated tasks. Within GLMP, we introduce two pivotal modules: the Hybrid Attention and Convolutional Paths Module (HACM) and the Multi-path Feature Fusion Module (MPFM). These modules synergistically converge multi-path features, yielding more discriminative feature information. Experimental results on the ZueriCrop dataset show that the proposed GLMP technique is effective; it performs promisingly having a total accuracy of 90.2% and an F1 value of 62.5%. Furthermore, the ablation study verifies the substantial improvement in classification accuracy for remote sensing crop images of long-time series in nature, specifically attributed to the HACM and MPFM modules.

Hypergraph clustering based multi-label cross-modal retrieval

Most existing cross-modal retrieval methods face challenges in establishing semantic connections between different modalities due to inherent heterogeneity among them. To establish semantic connections between different modalities and align relevant semantic features across modalities, so as to fully capture important information within the same modality, this paper considers the superiority of hypergraph in representing higher-order relationships, and proposes an image-text retrieval method based on hypergraph clustering. Specifically, we construct hypergraphs to capture feature relationships within image and text modalities, as well as between image and text. This allows us to effectively model complex relationships between features of different modalities and explore the semantic connectivity within and across modalities. To compensate for potential semantic feature loss during the construction of the hypergraph neural network, we design a weight-adaptive coarse and fine-grained feature fusion module for semantic supplementation. Comprehensive experimental results on three common datasets demonstrate the effectiveness of the proposed method.

A lightweight speech enhancement network fusing bone- and air-conducted speech.

Air-conducted (AC) microphones capture the high-quality desired speech and ambient noise, whereas bone-conducted (BC) microphones are immune to ambient noise but only capture band limited speech. This paper proposes a speech enhancement model that leverages the merits of BC and AC speech. The proposed model takes the spectrogram of BC and AC speech as input and fuses them by an attention-based feature fusion module. The backbone network of the proposed model uses the fused signals to estimate mask of the target speech, which is then applied to the noisy AC speech to recover the target speech. The proposed model adopts a lightweight design of densely gated convolutional attention network (DenGCAN) as the backbone network, which contains encoder, bottleneck layers, and decoder. Furthermore, this paper improves an attention gate and integrates it into skip-connections of DenGCAN, which allows the decoder to focus on the key areas of the feature map extracted by the encoder. As the DenGCAN adopts self-attention mechanism, the proposed model has the potential to improve noise reduction performance at the expense of an increased input-output latency. Experimental results demonstrate that the enhanced speech of the proposed model achieves an average 1.870 wideband-PESQ improvement over the noisy AC speech.

BiSTNet: Semantic Image Prior Guided Bidirectional Temporal Feature Fusion for Deep Exemplar-Based Video Colorization.

How to effectively explore the colors of exemplars and propagate them to colorize each frame is vital for exemplar-based video colorization. In this article, we present a BiSTNet to explore colors of exemplars and utilize them to help video colorization by a bidirectional temporal feature fusion with the guidance of semantic image prior. We first establish the semantic correspondence between each frame and the exemplars in deep feature space to explore color information from exemplars. Then, we develop a simple yet effective bidirectional temporal feature fusion module to propagate the colors of exemplars into each frame and avoid inaccurate alignment. We note that there usually exist color-bleeding artifacts around the boundaries of important objects in videos. To overcome this problem, we develop a mixed expert block to extract semantic information for modeling the object boundaries of frames so that the semantic image prior can better guide the colorization process. In addition, we develop a multi-scale refinement block to progressively colorize frames in a coarse-to-fine manner. Extensive experimental results demonstrate that the proposed BiSTNet performs favorably against state-of-the-art methods on the benchmark datasets and real-world scenes. Moreover, the BiSTNet obtains one champion in NTIRE 2023 video colorization challenge (Kang et al. 2023).

Ancient paintings inpainting based on dual encoders and contextual information

Deep learning-based inpainting models have achieved success in restoring natural images, yet their application to ancient paintings encounters challenges due to the loss of texture, lines, and color. To address these issues, we introduce an ancient painting inpainting model based on dual encoders and contextual information to overcome the lack of feature extraction and detail texture recovery when restoring ancient paintings. Specifically, the proposed model employs a gated encoding branch that aims to minimize information loss and effectively capture semantic information from ancient paintings. A dense multi-scale feature fusion module is designed to extract texture and detail information at various scales, while dilated depthwise separable convolutions are utilized to reduce parameters and enhance computational efficiency. Furthermore, a contextual feature aggregation module is incorporated to extract contextual features, enhancing the overall consistency of the inpainting results. Finally, a color loss function is introduced to ensure color consistency in the restored area, harmonizing it with the surrounding region. The experimental results indicate that the proposed model effectively restores the texture details of ancient paintings, outperforming other methods both qualitatively and quantitatively. Additionally, the model is tested on real damaged ancient paintings to validate its practicality and efficacy.

LWSDNet: A Lightweight Wheat Scab Detection Network Based on UAV Remote Sensing Images

Wheat scab can reduce wheat yield and quality. Currently, unmanned aerial vehicles (UAVs) are widely used for monitoring field crops. However, UAV is constrained by limited computational resources on-board the platforms. In addition, compared to ground images, UAV images have complex backgrounds and smaller targets. Given the aforementioned challenges, this paper proposes a lightweight wheat scab detection network based on UAV. In addition, overlapping cropping and image contrast enhancement methods are designed to preprocess UAV remote-sensing images. Additionally, this work constructed a lightweight wheat scab detection network called LWSDNet using mixed deep convolution (MixConv) to monitor wheat scab in field environments. MixConv can significantly reduce the parameters of the LWSDNet network through depthwise convolution and pointwise convolution, and different sizes of kernels can extract rich scab features. In order to enable LWSDNet to extract more scab features, a scab feature enhancement module, which utilizes spatial attention and dilated convolution, is designed to improve the ability of the network to extract scab features. The MixConv adaptive feature fusion module is designed to accurately detect lesions of different sizes, fully utilizing the semantic and detailed information in the network to enable more accurate detection by LWSDNet. During the training process, a knowledge distillation strategy that integrates scab features and responses is employed to further improve the average precision of LWSDNet detection. Experimental results demonstrate that the average precision of LWSDNet in detecting wheat scab is 79.8%, which is higher than common object detection models and lightweight object detection models. The parameters of LWSDNet are only 3.2 million (M), generally lower than existing lightweight object detection networks.

MAA-YOLOv8: enhanced steel surface defect detection through multi-head attention mechanism and lightweight feature fusion

In the process of industrial production, product defects often arise due to improper operations among other reasons, rendering the detection of such flaws an indispensable procedure. However, the vast array of defect types, coupled with their complex characteristics, poses ongoing challenges for contemporary defect detection algorithms within industrial settings. To solve this problem, the present study introduces an enhanced steel surface defect detection model based on the modified YOLOv8 algorithm-termed the MAA-YOLOv8 model-to augment the accuracy and practicality of the algorithm. Initially, a multi-head attention mechanism was incorporated into the C2f to bolster the feature extraction capabilities within the backbone network and diversify the attention maps. Secondly, in the neck structure, we design a multi-channel feature fusion module (McPAN) to solve the problem of balance between computational efficiency and the ability to capture useful features. A series of experiments conducted on the NEU-DET dataset reveal that the MAA-YOLOv8 model achieves a mean Average Precision (mAP) of 94.4%, representing an enhancement of 11.1% over the original YOLOv8s model. The MAA-YOLOv8 model proposed in this study substantially elevates the performance of steel surface defect detection while ensuring the speed of detection.

Hybrid modeling for vehicle lateral dynamics via AGRU with a dual-attention mechanism under limited data

A precise vehicle dynamics model is critical for simulation and algorithm testing. Neural networks have been widely used to build high-fidelity vehicle dynamics models due to the excellent learning ability. However, data starvation is a common phenomenon in neural networks. With limited data, it is difficult for neural networks to achieve precise predictions. To address these problems, a hybrid model combining physics and dual attention neural networks is developed to model vehicle lateral dynamics. First, due to the interpretability of the physical model, linear lateral dynamics model is regarded as a prior model. However, due to the imperfect prior knowledge, there are residuals between the prior and the actual vehicle dynamics. Therefore, neural networks are used to characterize the residuals to achieve recalibration of vehicle dynamics model. Modeling vehicle residual dynamics with neural networks is a time series forecasting problem. The GRU with a dual attention mechanism and adaptive initial hidden states (DA-AGRU) is designed to capture spatial and temporal correlations in the vehicle dynamics data. In particular, considering the unique auto regressive structure of the hybrid model, a spatial attention mechanism with a feature fusion module is designed, so as to globally compute the weights of different channel features. The dataset used to train and validate the model is recorded from a vehicle platform, and the experimental results show that the proposed hybrid model can accurately predict vehicle dynamics states in a data-scarce environment.

Self-Supervised Monocular Depth Estimation for Endoscopic Imaging.

Endoscopy holds a pivotal role in the early detection and treatment of diverse diseases, with artificial intelligence (AI)-assisted methods increasingly gaining prominence in disease screening. Among them, the depth estimation from endoscopic sequences is crucial for a spectrum of AI-assisted surgical techniques. However, the development of endoscopic depth estimation algorithms presents a formidable challenge due to the unique environmental intricacies and constraints within the dataset. This paper proposes a self-supervised depth estimation network to comprehensively explore the brightness changes in endoscopic images, and fuse different features at multiple levels to achieve an accurate prediction of endoscopic depth. First, a FlowNet is designed to evaluate the brightness changes of adjacent frames by calculating the multi-scale structural similarity. Second, a feature fusion module is presented to capture multi-scale contextual information. Experiments show that the average accuracy of the algorithm is 97.03% in the Stereo Correspondence and Reconstruction of Endoscopic Data (SCARED dataset). Based on the training parameters of the SCARED dataset, the algorithm achieves superior performance on the other two datasets (EndoSLAM and KVASIR dataset), indicating that the algorithm has good generalization performance.

A duplex transform heterogeneous feature fusion network for road segmentation

Detecting roads in automatic driving environments poses a challenge due to issues such as boundary fuzziness, occlusion, and glare from light. We believe that two factors are instrumental in addressing these challenges and enhancing detection performance: global context dependency and effective feature representation that prioritizes important feature channels. To tackle these issues, we introduce DTRoadseg, a novel duplex Transformer-based heterogeneous feature fusion network designed for road segmentation. DTRoadseg leverages a duplex encoder architecture to extract heterogeneous features from both RGB images and point-cloud depth images. Subsequently, we introduce a multi-source Heterogeneous Feature Reinforcement Block (HFRB) for fusion of the encoded features, comprising a Heterogeneous Feature Fusion Module (HFFM) and a Reinforcement Fusion Module (RFM). The HFFM leverages the self-attention mechanisms of Transformers to achieve effective fusion through token interactions, while the RFM focuses on emphasizing informative features while downplaying less important ones, thereby reinforcing feature fusion. Finally, a Transformer decoder is utilized to produce the final semantic prediction. Furthermore, we employ a boundary loss function to optimize the segmentation structure area, reduce false detection areas, and improve model accuracy. Extensive experiments are carried out on the KITTI road dataset. The results demonstrate that, compared with state-of-the-art methods, DTRoadseg exhibits superior performance, achieving an average accuracy of 97.01%, a Recall of 96.35%, and runs at a speed of 0.09 s per picture.

SGDFormer: One-stage transformer-based architecture for cross-spectral stereo image guided denoising

Cross-spectral image guided denoising has shown its great potential in recovering clean images with rich details, such as using the near-infrared image to guide the denoising process of the visible one. To obtain such image pairs, a feasible and economical way is to employ a stereo system, which is widely used on mobile devices. In this case, it is necessary to first model the stereo correspondence between two images before guided denoising. Current works attempt to generate an aligned guidance image to handle the disparity. However, due to occlusion, spectral differences and noise degradation, the aligned guidance image generally exists ghosting and artifacts, leading to an unsatisfactory denoised result. To address this issue, we propose a one-stage transformer-based architecture, named SGDFormer, for cross-spectral Stereo image Guided Denoising. The architecture integrates the correspondence modeling and feature fusion of stereo images into a unified network. Our transformer block contains a noise-robust cross-attention (NRCA) module and a spatially variant feature fusion (SVFF) module. The NRCA module captures the long-range correspondence of two images in a coarse-to-fine manner to alleviate the interference of noise. The SVFF module further enhances salient structures and suppresses harmful artifacts through dynamically selecting useful information. Thanks to the above design, our SGDFormer can restore artifact-free images with fine structures, and achieves state-of-the-art performance on various datasets. Additionally, our SGDFormer can be extended to handle other unaligned cross-model guided restoration tasks such as guided depth super-resolution.

A method of dense point cloud SLAM based on improved YOLOV8 and fused with ORB-SLAM3 to cope with dynamic environments

With the development of society and the advancement of technology, intelligent robots have been widely used in various fields. At the same time, Simultaneous Localization and Mapping (SLAM) technology is a key technology in the research field of intelligent robots. However, in dynamic environments, achieving accurate and robust visual SLAM remains a major challenge. In this paper, we propose a method based on improved YOLOv8 fused with ORB-SLAM3 to address dense point cloud SLAM in dynamic environments. Our proposed method successfully integrates real-time object detection and image segmentation technologies of YOLOv8 into the ORB-SLAM3 framework, achieving high-precision and robust visual SLAM in dynamic environments. In the YOLOv8 framework, we use a balanced convolution method, GSConv, instead of some traditional convolution layers (Conv), which balances accuracy with computational load. Based on the GSConv convolution method, we adopt a new feature fusion module, VoVGSCSP, to replace traditional C2f feature fusion modules, thereby improving the Neck structure of YOLOv8 and achieving a lightweight network model. We compare our proposed method with ORB-SLAM3 and some computer vision algorithms on the TUM dataset. Experimental data confirms that our method outperforms existing visual SLAM algorithms in dynamic environments. In fast-moving dynamic environments, the RMSE of absolute pose estimation of our method is 96.28% lower than that of ORB-SLAM3, and the RMSE of relative pose estimation is 51.57% lower than that of ORB-SLAM3. The experimental results demonstrate that our method significantly improves the accuracy of pose estimation in dynamic environments and greatly enhances the performance compared to ORB-SLAM3.

REDef-DETR: real-time and efficient DETR for industrial surface defect detection

Abstract Industrial surface defect detection is an important part of industrial production, which aims to identify and detecting various defects on the surface of product to ensure quality and meet customer requirements. With the development of deep learning and image processing technologies, the surface defect detection methods based on computer vision has become the mainstream method. However, the prevalent convolutional neural network-based defect detection methods also have many problems. For example, these methods rely on post-processing of Non-Maximum Suppression and have poor detection ability for small targets, which affects the speed and accuracy of surface defect detection in industrial scenarios. Therefore, we propose a novel DEtection TRansformer-based surface defect detection method. Firstly, we propose a Multi-scale Contextual Information Dilated module and fuse it into the backbone. The module is mainly composed of large kernel convolutions, which aims to expand the receptive field of the model, thus reducing the leakage rate of the model. Moreover, we design an efficient encoder which mainly contains two important modules, namely feature enhancement based on cascaded group attention module and efficient feature fusion module based on content-aware. The former module effectively enhances the high-level semantic information extracted by the backbone, thus enabling the model to better interpret features, and it can improve the problem of high computational cost of transformer encoder, thus increasing the detection speed. The latter module performs multi-scale feature fusion across the feature information of various scales, thus improving the detection accuracy of the model for small-size defects. Experimental results show that the proposed method achieves 80.6%mAP and 80.3FPS on NEU-DET, and 98.0%mAP and 79.4FPS on PCB-DET. Our proposed method exhibits excellent detection performance and achieves real-time and efficient surface defect detection capability to meet the needs of industrial surface defect detection.