MEM-RCNN: A Deep Learning Based Image Segmentation Algorithm for Pipeline Magnetic Leakage Defects

Proposes a deep learning based pipeline magnetic flux leakage defect image segmentation algorithm. Aim to address the issues of small target defects being prone to loss and low segmentation accuracy in pipeline magnetic flux leakage image segmentation. Algorithm is based on the Mask-RCNN network and modifies the backbone of the network. The core is to propose a margin enhancement module (MEM). Replace the head of the ResNet residual network with the proposed MEM module, which enhances the edge features of objects in the input image by increasing the weight of image edge position pixels. Experiments have shown that compared to the original MaskRCNN network, the MEM-RCNN network proposed in this paper performs more detailed object edge segmentation and has good segmentation performance for small defects.