DCUE-YOLO: A Lightweight Model in Industrial Defect Detection

Accurate and rapid identification of defects in industrial products is essential for ensuring quality and safety. However, the challenges presented by large-scale production environments, along with the difficulty in distinguishing between target defects and complex backgrounds, complicate defect detection. Consequently, most target detection models struggle to achieve an optimal balance between detection accuracy and efficiency. To improve detection accuracy and efficiency, this paper proposes a lightweight network architecture, DCUE-YOLO, based on YOLOv10. The primary objective is to improve both the accuracy and efficiency of industrial product defect detection. In addition, a feature extraction module with double convolutional path design with hidden channels is proposed under the premise of reducing the computational complexity; by capturing information of different scales, the model can enhance the ability to distinguish small target defects from complex backgrounds. In order to further improve the model's attention to small target defects, this paper also proposes a multifilter attention mechanism design. Meanwhile, in order to effectively solve the problem of partial feature information loss in the process of downsampling, this paper also uses a transposed convolution design. Extensive experiments were carried out using PCB, NEU-DET and mixed-type WM38 public data sets, producing mean mean average precision (mAP) scores of 94.3%, 90.5%, and 98.7%, respectively. Compared to the YOLOv10s model, our mAP has improved by 2.7%, 1.8%, and 1.2%, respectively, while the parameter count has decreased by 0.3M. Our model demonstrates advantages in recognition accuracy and inference speed, thus validating its effectiveness for industrial defect detection.