Corrosion damage segmentation network based on multi-scale feature mapping fusion

• Linjie Zhou,
• Likun Hu,
• Lu Ding,

Vol. 19, No. 7, July 31, 2025
10.3837/tiis.2025.07.009, Download Paper (Free):

• Defect segmentation
• Cross attention
• Transformer
• Convolutional Neural Network
• Hybrid structures

Abstract

In this work, we propose a novel architecture designed to enhance the segmentation of metal coating defects named CAFTNet (Cross Attention Fusion Transformer Network). The network integrates the strengths of both Transformer and Convolutional Neural Networks (CNNs) to address the challenges of weak feature extraction and suboptimal segmentation accuracy in existing methods. We use the Swin Transformer-Base as the encoder, selected for its ability to capture global dependencies in image data. To overcome the limitations of the receptive field during network propagation, we design a cross-attention feature mapping fusion module. This module effectively integrates encoded and decoded features, ensuring feature alignment and representation, thus facilitating robust modeling for the segmentation task. In the decoder phase, we craft an efficient convolutional module to decode the fused features, ensuring the accurate translation of high-level representations into precise spatial information. On two distinct datasets characterized by metal coating flaking and corrosion, CAFTNet achieved F1 scores of 90.0% and 96.6%, mIoU scores of 82.3% and 93.5%, and overall accuracy (OA) of 93.6% and 95.3%, respectively. The source code will be available at https://github.com/touchfish-z/CAFTNET.

Statistics

Cite this article