融合双注意力机制和U-Net网络的锈蚀图像分割

A Segmentation Method Based on Dual Attention Mechanism and U-Net for Corrosion Images

针对传统方法难以精确分割出金属构件锈蚀区域特征的难题,构建了一种融合双注意力机制和U-Net深度学习网络的锈蚀图像区域分割模型。首先,基于U-Net网络的对称编解码架构搭建骨干网络,采用VGG16网络的预训练权重对模型参数进行初始化;其次,在下采样和上采样之间的跨层连接中融合双注意力机制使网络聚焦于局部锈蚀特征,同时在上采样中使用深度可分离卷积加速模型的运算效率;最后采用锈蚀图像数据集对该网络进行训练从而得到锈蚀图像分割模型。通过实采的金属构件锈蚀图像对所提模型进行验证,结果表明:所构建的锈蚀图像分割模型能够有效地从复杂背景图像中分割出锈蚀区域特征,锈蚀区域特征的识别准确率达到95.5%,交并比指标为81.4%;相较于传统U-Net方法,识别准确率和交并比指标分别提升了3.3%和9.2%。

A novel segmentation model based on dual attention mechanism and U-Net deep learning network for corrosion images is constructed to solve the problem that traditional methods are difficult to accurately segment the features of corrosion areas from corrosion images of metal components.Firstly,the symmetrical code-decode architecture of U-Net is used to construct the backbone network,and the pre-training weights of VGG16 network are adopted to initialize parameters for the proposed model.Secondly,the dual attention mechanism is integrated in the cross layer connection between down-sampling and up-sampling to make the network focus on local corrosion features.At the same time,the deep separable convolution is used in up-sampling to improve the computational efficiency.Finally,the network is trained by using the corrosion image data set to obtain the corrosion image segmentation model.The effectiveness and practicability of the proposed model are verified by the actual corrosion images of metal components.Experimental results show that the proposed segmentation model effectively segments the corrosion area features from the complex background images.The recognition accuracy of corrosion area features reaches 95.5%,and the index of the intersection over union is 81.4%.Compared with the traditional U-Net model,the accuracy and the index of intersection over union of the proposed model are improved by 3.3%and 9.2%,respectively.