PCB surface defect fast detection method based on attention and multi-source fusion

PCB board defect detection is a necessary part of the PCB manufacturing process and needs to be repeated several times to ensure the quality of the PCB board. However, existing PCB surface defect detection methods suffer from high computational effort, low robustness and slow speed. To this end, this paper proposes the ShuffleNetV2-YOLOv5 model, a fast PCB surface defect detection method based on an attention mechanism and multi-source information fusion. The model uses a modified ShuffleNetV2 as the backbone network and incorporates an attention mechanism in multiple stages of the backbone network to improve the model's focus on valid information at different depths. Parameters for each stage of the Efficient NAS are introduced to limit the effect of degrees of freedom on the robustness of the network and to effectively reduce the number of parameters of the model. In the feature fusion section, an enhanced feature fusion structure, S3Head structure, is proposed to incorporate information from the sampled part of each stage into the feature fusion, enriching the information source for feature fusion.In addition, a CBAM attention mechanism is introduced at each sampling output stage to enhance the contextual information interaction capability. A migration learning algorithm is employed in the training process to further improve the training of the model. Experimental results show that the ShuffleNetV2-YOLOv5 model has higher accuracy and lower floating point operations than several lightweight models such as YOLOv3tiny and YOLOv4tiny. To further improve the human-computer interaction experience, an upper computer interface was designed using PyQT.