Multibranched Dilated Convolution With Feature Fusion and Dropout for Accurate Wafer Surface Defect Detection

Product defect detection is a critical evaluation process in intelligent manufacturing systems, providing essential analysis for improving product quality. In semiconductor manufacturing, identifying defects on the surface of wafers is crucial for determining the root causes of the defects, which helps improve chip yield. We propose a deep learning network for detecting surface defects on wafers, composed of three synergistic innovative components. First, the multibranch dilated convolution module (MDCM) enhances the network’s ability to extract defect features by expanding the receptive field of feature extraction. Next, the residual feature pyramid network (ReFPN) facilitates the fusion of defect features across different network layers, significantly improving the detection capabilities for small-scale targets by enriching semantic content between layers. At last, the use of reciprocating channel dropout (RCD) removes redundant feature information, significantly increasing training stability and accuracy. Our model surpasses existing state-of-the-art models by 1.8% in average precision, achieves 37.56 fps, and performs best in detecting defects of various sizes and aspect ratios.