Lightweight YOLOv7-tiny for PCB defects small object detection

In the PCB defect detection, the defect targets to be detected are small. In order to avoid a large number of target detector parameters and low efficiency of detection, a lightweight defect detection model for PCB based on improved YOLOv7- tiny is proposed. First, due to the complex fragmentation operation and insufficient across spatial scales down-sampling operation of the backbone network ELAN, we propose a fast backbone network constructed by a cross-scale max pooling layer with zero parameter and a partial channel re-parameterized module, utilize the cross-scale interaction, multi-branch training of the backbone network to improve feature extraction capability and achieve simplify reasoning. At the same time, to address the complex connectivity of the PANet and the deep network is difficult to capture the feature information of small targets, a single-scale prediction head method based on the global and local hybrid feature condensing module is proposed, which models the multi-scale information of the backbone network globally and locally, and shares the information interactively, to better capture the global context correlations and dependencies of images, and enhance the location information of small targets , as well as to further reduce the computational costs and the complexity of the model. Experiments were carried out on the DeepPCB dataset. And the results show that the number of parameters and flops of the improved model are reduced to 29% and 51% compared to the original model, the inference speed is increased by 62%, which is better than other mainstream lightweight target detectors.