MSF-SegFormer: a feature fusion algorithm for magnetic leakage image segmentation

Abstract

Traditional segmentation networks have low segmentation accuracy for flux leakage images, often leading to missed or false detections of small defects, which significantly affect the evaluation of defect severity. Based on the SegFormer network, a high-accuracy decoder based on multi-scale feature fusion is proposed, which is more suitable for the segmentation of small defects in flux leakage and replaces the multi-layer perceptron (MLP) decoder of the original network. The new network model is called MSF-SegFormer. MSF-SegFormer introduces a feature fusion network MSF that integrates high-resolution and low-resolution features and introduces feature pyramid fusion, which can merge output features at different levels across different scales. A cascaded attention module is proposed, combining two local attention mechanisms in a cascade and using a residual network to enhance the local feature representation of flux leakage images, improving the accuracy and stability of the task. In the application of flux leakage defect data, compared with benchmark models such as CNN and SegFormer, this model can accurately segment target edges with fewer parameters, maintain high accuracy, reduce false detection probability, and improve the Miou value of the traditional MLP decoder from 88.21% to 90.44%.