摘要
针对复杂背景下细小裂纹图像检测难、噪声干扰多和裂纹宽度信息易丢失的问题,提出一种基于U-Net改进的方法。利用残差块解决网络退化,加入BN层改善梯度弥散,融入深度可分离卷积以及高尺度的转置卷积,实现特征信息由浅入深的传递;改进注意力机制,实现细节特征的优化;延伸U-Net特征向量长度,在底部加入由最大池化层、小尺度深度可分离卷积与上采样层构建的层融合模块,实现分辨率和感受野之间的平衡。实验结果表明,在客观标准下,改进的方法比U-Net的IoU的值提高0.1873,Recall的值提高了0.1127,Precision提高了0.1359,F1-score提高了0.0687,并且实验结果皆优于其他方法对于U-Net的改进,减少了伪分割现象,完成对细小裂纹分割,获得更加精细的裂纹宽度信息。
In the complex backgrounds,image detection of fine cracks is difficult,noise interference is high,and crack width information is easy to lose.To solve these problems,we proposed an improved method based on U-Net.We used the residual block to solve network degradation,added BN layer to improve gradient dispersion,and integrated depth-wise separable convolution and high scale transpose convolution,so as to transfer the feature information from shallow to deep.The attention mechanism was improved to optimize the details.We extended the length of the U-Net feature vector,and added a layer fusion module at the bottom constructed by the max pooling layer,small scale depth-wise separable convolution and the upsampling layer,thus realizing the balance between resolution and receptive field.The experimental results show that in the objective standard,the improved method is 0.1873 higher than the IoU value of U-Net,the Recall value is increased by 0.1127,the precision is increased by 0.1359,and the F1-score is increased by 0.0687.The results are better than other improved methods for U-Net.The method reduces the false segmentation,completes the segmentation of fine crack and obtains the more detailed crack width information.
作者
封晓晨
李宁
顾玉宛
符心宇
王雨生
徐守坤
Feng Xiaochen;Li Ning;Gu Yuwan;Fu Xinyu;Wang Yusheng;Xu Shoukun(School of Information Science and Engineering,Changzhou University,Changzhou 213164,Jiangsu,China;School of Computer and Information,Hohai University,Nanjing 210098,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2022年第3期193-200,共8页
Computer Applications and Software
基金
国家自然科学基金项目(61906021)
江苏省教育厅2018年师资队伍建设第二批专项经费项目。
作者简介
封晓晨,硕士生,主研领域:图像处理;李宁,副教授。顾玉宛,讲师;符心宇,硕士生;王雨生,硕士生;徐守坤,教授。
