基于深度学习的金属表面腐蚀图像语义分割模型性能分析

Semantic Segmentation Experiment and Model Performance Analysis of Metal Corrosion Images Based on Deep Learning

基于计算机视觉的金属表面腐蚀检测技术是利用深度学习模型对金属腐蚀图像进行像素级语义分割,通过模型训练提取腐蚀区域的纹理、形状等特征,根据腐蚀面积判断腐蚀程度,解决了人工检测带来的效率低、误差大等问题,便于实现后续自动化检测、智能化评估与管具服役失效数据库的建立。为了更好地实现腐蚀区域的精确定位和分割,基于公开的金属腐蚀图像数据集,建立了512×512像素的全图数据集和切片图像数据集,通过实验研究了FCN、U-Net、PSPNet、DeepLabv3+、HRNet和SegFormer等深度学习模型在金属腐蚀检测方面的性能。研究结果表明,所有模型在金属腐蚀图像分割任务中使用切片图像数据集的性能优于全图数据集。对比其他模型,SegFormer模型在切片图像数据集训练下,mIoU和mPA分别为66.20%、78.55%,参数量为3.72 MB,取得了出色的分割效果。

The metal surface corrosion detection technology based on computer vision uses the deep learning model to perform pixel-level semantic segmentation on metal corrosion image.It extracts the texture,shape and other features of the corrosion area through model training,and judges the corrosion degree according to the corrosion area,which solves the problems of low efficiency and large error caused by manual detection.It is convenient for the subsequent automatic detection,intelligent evaluation and the establishment of pipe service failure database.In order to better locate and segment the corroded area,based on the public metal corrosion image dataset,the 512×512 pixel full image dataset and slice image dataset were established and the performance of deep learning models such as FCN,U-Net,PSPNet,DeepLabv3+,HRNet and SegFormer in metal corrosion detection were studied through experiments.The results indicate that all models perform better in metal corrosion image segmentation task using slice image dataset than full image dataset.Compared with other models,the SegFormer model trained on the slice image dataset has the mIoU and mPA of 66.20% and 78.55%,and the parameter amount is 3.72 MB,which gets excellent segmentation results.