Attention-Guided Shape-Aware Double-branch Segmentation Network

In order to solve the problem of low segmentation accuracy caused by defect interference, dirty noise and lens blur under complex conditions, an attention-guided shape-aware double-branch segmentation network is proposed. Firstly, to solve the problem of low accuracy caused by deep to shallow semantic propagation errors, a semantic flow alignment module is proposed, which learns offsets between feature maps to assist information alignment. Secondly, an attention-guided self-selection fusion module is proposed, which combines the characteristics of deep information and shallow information to guide more accurate segmentation. At the same time, the shape-aware loss function is proposed to solve the problem of noise and target adhesion. This function uses shape features to guide the network to focus on the boundary region that is difficult to partition to improve segmentation performance. Comprehensive experiments on a self-built chip dataset confirm that this method improves feature representation and segmentation performance, with mIoU up to 94.4% (2.1%) and FPS up to 21%. In the CamVid dataset, mIoU is 65.1% (3.0% increase), and the number of parameters is reduced by 4.6%, which achieves a good balance between real-time and accuracy.