Mask RCNN Chip Surface Defect Detection based on Multi-Scale Grouped and Dual Attention Mechanism

Addressing the issue of low detection precision caused by multiple types of defects on chip surfaces that often overlap, this paper proposes an improved Mask RCNN algorithm for chip surface defect detection based on multi-scale grouped and dual attention mechanisms. The standard convolution in the original ResNet50 network structure is modified to four group convolutions of different scales, with each group using different convolutional kernels. The feature maps obtained from each group are then concatenated to achieve channel stacking. Additionally, the SENet channel attention mechanism and the GCNet spatial attention mechanism are introduced at different stages of the Feature Pyramid Network (FPN) to fully utilize the hierarchical feature representations, thereby enhancing the network's ability to perceive and accurately detect surface defects. In the MPU6050 chip surface defect detection experiment, the average detection precision of the original Mask RCNN algorithm is 93.16%, and the proposed algorithm reaches the average detection precision of 96.35%, which significantly improves the detection performance of MPU6050 chip surface defects.