PCB Image Segmentation Algorithm Based on Cross-Layer Non-Local Fusion and DeepLabV3+
For the issues of low smoothness,poor continuity and low segmentation efficiency of target edges in the PCB image segmenta-tion process,a lightweight image segmentation model combining attention mechanism is proposed.First,a MobileNetV2 network is used for deep feature extraction.Second,a branch of the features is input to the void space pyramid pooling module for multi-scale feature ex-traction and fused to obtain high-level features.Finally,a cross-layer non-local module is introduced to fuse the bottom-level features ob-tained by convolution through another branch with the above high-level features.The mean intersection over union of the method is 96.176%,the precision is 97.59%,the recall is 95.912%,the segmentation speed is 0.062 s,and the number of parameters is 25.39 Mbyte.The proposed method takes into account the problem of small target detection and boundary information loss in the image,and improves the accuracy and real-time performance of image segmentation.
万方数据
维普
