Efficient aggregate distribute network for tiny defect detection

Industrial products are indispensable in dailylife, real-time surface defect detection is crucial for ensuring product quality and optimizing production line efficiency. However, the complex background of surface defects of industrial products, diverse defect types, and irregular defect shapes make it challenging for general object detectors to effectively classify and locate defects in defect detection tasks. Therefore, this paper proposes an efficient aggregate distribute network (AD-Net) to optimize performance of defect detection in intricate industrial scenes. First, considering that defects have random distribution and irregular shape, this paper introduces an enhanced linear deformable convolution (ELDConv) in the backbone network stage of extracting deep feature. ELDConv expands the receptive field of the defect feature extraction network, helps the network capture comprehensive and key defect semantic feature. Secondly, a lightweight aggregate distribute feature pyramid network (AD-FPN) is designed in the neck to effectively aggregate and distribute cross-layer feature. Finally, a multi-scale adaptive-aware detection head (MASH) is constructed, which can dynamically assign appropriate local context to tiny targets of different scales to improve detection accuracy. Experiments show that the mean average precision (mAP) of the proposed AD-Net reaches 80.8% on the alibaba tianchi fabric dataset. 98.8% mAP on the printed circuit board (PCB) defect dataset. 78.6% mAP on the NEU-DET defect dataset. In addition, taking account into the detection accuracy, real-time detection speed and model size, AD-Net is suitable for deployment on embedded devices for real-time defect detection.