Robots with cameras provide a non-contact information acquisition solution for hydraulic tunnels, while manual damage-related information extraction is time-consuming and costs labor. This study proposes a robust real-time framework for identifying hydraulic tunnel underwater structural damage using deep learning and computer vision. First, a high-performance detector is built via the You Only Look Once v5s and adaptively spatial feature fusion module. A series of comparative experiments are used to explore the setting of the sparsification and pruning ratios to trade off a balance between accuracy and efficiency. Model sparsity ratio of 0.01 with a pruning ratio of 0.3 can be combined to change the weight distribution in the batch normalization layer and reduce network redundant parameters for slimming. Then, model fine-tuning with knowledge distillation is utilized to recover the accuracy degradation caused by pruning. A hydraulic tunnel is utilized as the case study, and three defects including rust, exfoliation, and calcification precipitate are utilized as research items. The performance of the models was evaluated based on detection accuracy, robustness, and efficiency. Five extreme attributes of underwater scenes, including oblique angles, high brightness, uneven illumination, low visibility, and obstacle interference, were considered to test model generalization and efficacy. Experimental results show it achieves good detection performance in complicated underwater scenes, achieving 0.814 precision, 0.980 recall, 0.889 F1_score, and 0.894 Mean Average Precision (mAP)@0.5 in the test set. Moreover, the proposed method achieved a 50 Frames Per Second (FPS) detection speed when detecting video with 1080p, indicating its real-time detection capability.

中文翻译：

---

使用深度学习和计算机视觉识别水工隧道结构缺陷的鲁棒实时方法

带有摄像头的机器人为水工隧道提供了一种非接触式的信息获取解决方案，而人工损伤相关信息的提取费时费力。本研究提出了一个强大的实时框架，用于使用深度学习和计算机视觉识别水工隧道水下结构损坏。首先，通过 You Only Look Once v5s 和自适应空间特征融合模块构建高性能检测器。一系列比较实验用于探索稀疏化和修剪比率的设置，以权衡准确性和效率之间的平衡。0.01 的模型稀疏度和 0.3 的剪枝比可以相结合，改变 batch normalization 层的权重分布，减少网络冗余参数以实现瘦身。然后，利用知识蒸馏的模型微调来恢复由修剪引起的精度下降。以某水工隧道为例，将生锈、剥落、钙化沉淀3种缺陷作为研究项目。根据检测精度、鲁棒性和效率评估模型的性能。考虑了水下场景的五个极端属性，包括倾斜角度、高亮度、照明不均匀、低能见度和障碍物干扰，以测试模型的泛化性和有效性。实验结果表明，它在复杂的水下场景中具有良好的检测性能，在测试集中实现了 0.814 精度、0.980 召回率、0.889 F1_score 和 0.894 Mean Average Precision (mAP)@0.5。而且，