Abstract: The rust detection algorithm is the key to the transformation of steel bridge maintenance from artificial vision to machine vision, and it is also the technical foundation for the construction of intelligent steel bridge detection robots. In order to meet the practical requirements of low energy consumption and high accuracy of rust detection algorithms for steel bridge intelligent detection robots, the DeepLabV3+ model was improved. The MobileNetV2 backbone network was used to replace the Xception backbone network in the original model, making the model lightweight and adaptable to mobile devices. The void rate in the ASPP module was optimized to improve the network's ability to extract rust of different sizes. The CBAM attention mechanism was added to enhance the model's perception and capture of key features. A comparative experiment was conducted between the improved DeepLabV3+ model and the original DeepLabV3+ model, PSPNet model, and U-Net model. At the same time, ablation experiments were conducted. Finally, the improved model was mounted on a visual robot and field engineering tests were conducted. The results show that the improved DeepLabV3+ model has an average accuracy improvement of 7.5%, an average intersection to intersection ratio improvement of 14.7%, and an average recall rate improvement of 9.1%, compared to other models in segmenting steel bridge rust images.

Key words: bridge engineering; DeepLabV3+; steel bridge rust detection; convolutional neural networks; image segmentation

摘要： 锈蚀检测算法是钢桥管养从人工视觉向机器视觉转型的关键，更是智能化钢桥检测机器人构建的技术基础。面向钢桥智能检测机器人对锈蚀检测算法低能耗和高精度的实际需求，针对DeepLabV3+模型加以改进，采用MobileNetV2主干网络替换原模型中的Xception主干网络，使模型轻量化以易适配移动端设备，优化ASPP模块中的空洞率以提高网络对不同尺寸锈蚀的提取效果，添加CBAM注意力机制增强模型对关键特征的感知和捕捉；将改进后的DeepLabV3+模型与原DeepLabV3+模型、PSPNet模型和U-Net模型进行了对比，同时开展了消融实验；最后，将改进模型搭载于视觉机器人上，并开展实地工程测试。结果表明：相比于其它模型，改进的DeepLabV3+模型对钢桥锈蚀图像的分割准确率平均提高了7.5%，平均交并比平均提高了14.7%，召回率平均提高了9.1%。

关键词: 桥梁工程；DeepLabV3+；钢桥锈蚀检测；卷积神经网络；图像分割