Abstract: Aiming at solving the problems of low accuracy of small-scale defect detection, poor edge segmentation effect in complex texture background and slow detection speed of traditional semantic segmentation model for aluminum-plastic blister package defect detection, this paper proposes an improved surface defect detection method for aluminum-plastic blister package. Firstly, the backbone feature extraction network replaces the original Xception network with the lightweight MobileNetV2 network, which significantly reduces the number of model parameters. Secondly, the efficient channel attention (ECA) module is cascaded in the feature extraction module and atrous spatial pyramid pooling (ASPP) module to accelerate the global feature fusion and reduce the loss of detail information, thus improving the segmentation accuracy of the model for small-scale defects. Finally, a boundary refinement module is added to the decoder of DeepLabv3+ to improve the segmentation accuracy of the model on the edges of the defect region under the complex texture background of the aluminum-plastic surface. Experimental verification is carried out on the self-built capsule board image dataset, and the
Keywords: aluminum-plastic blister package; defect detection; semantic segmentation; attention mechanism; boundary refinement
合肥工业大学学报自科版