基于动态特征蒸馏的水工隧洞缺陷识别方法  

Hydraulic tunnel defect recognition method based on dynamic feature distillation

作　　者：黄继爽 张华[1,2,3] 李永龙 赵皓 王皓冉[3,4] 冯春成[1,2,3] HUANG Jishuang;ZHANG Hua;LI Yonglong;ZHAO Hao;WANG Haoran;FENG Chuncheng(School of Information Engineering,Southwest University of Science and Technology,Mianyang Sichuan 621010,China;Robot Technology Used for Special Environment Key Laboratory of Sichuan Province(Southwest University of Science and Technology),Mianyang Sichuan 621010,China;Sichuan Energy Internet Research Institute under Tsinghua University,Chengdu Sichuan 610213,China;State Key Laboratory of Hydroscience and Engineering(Tsinghua University),Beijing 100084,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]特殊环境机器人技术四川省重点实验室(西南科技大学),四川绵阳621010 [3]清华四川能源互联网研究院,成都610213 [4]水沙科学与水利水电工程国家重点实验室(清华大学),北京100084

出　　处：《计算机应用》2021年第8期2358-2365,共8页journal of Computer Applications

基　　金：国家重点研发计划项目(2019YFB1310503);国家“十三五”核能开发科研项目(20161295);四川省科技创新创业苗子工程(2020JDRC0130);中国大唐集团有限公司科学技术项目(CDT-TZK/SYD[2018]-010)。

摘　　要：针对水工隧洞缺陷识别任务中现有深度卷积神经网络(DCNN)对缺陷图像特征提取能力不足、识别种类少、推理耗时长的问题,提出一种基于动态特征蒸馏的缺陷自主识别方法。首先,通过深度曲线估计网络对图像进行优化,从而改善低照度环境下的图像质量;其次,构建加入注意力机制的动态卷积模块取代传统静态卷积,并且把得到的动态特征用于训练教师网络以获得更好的模型特征提取能力;最后,在知识蒸馏框架中融合鉴别器结构,以构造一种动态特征蒸馏损失,并通过鉴别器将动态特征知识从教师网络传递到学生网络,从而在大幅减少模型推理时间的同时实现六类缺陷的高精度识别。在四川某水电站水工隧洞缺陷数据集上对该方法和原有残差网络进行对比实验,结果表明该方法可达到96.15%的识别准确率,其模型参数量和推理时间分别降低到原来的1/2和1/6。通过实验结果可知,将缺陷图像的动态特征蒸馏信息融合到识别网络中能够提高水工隧洞缺陷的识别效率。Aiming at the problems that the existing Deep Convolutional Neural Network(DCNN)have insufficient defect image feature extraction ability,few recognition types and long reasoning time in hydraulic tunnel defect recognition tasks,an autonomous defect recognition method based on dynamic feature distillation was proposed.Firstly,the deep curve estimation network was used to optimize the image to improve the image quality in low illumination environment.Secondly,the dynamic convolution module with attention mechanism was constructed to replace the traditional static convolution,and the obtained dynamic features were used to train the teacher network to obtain better model feature extraction ability.Finally,a dynamic feature distillation loss was constructed by fusing the discriminator structure in the knowledge distillation framework,and the dynamic feature knowledge was transferred from the teacher network to the student network through the discriminator,so as to achieve the high-precision recognition of six types of defects while significantly reducing the model reasoning time.In the experiments,the proposed method was compared with the original residual network on a hydraulic tunnel defect dataset of a hydropower station in Sichuan Province.The results show that this method has the recognition accuracy reached 96.15%,and the model parameter amount and reasoning time reduced to 1/2 and 1/6 of the original ones respectively.It can be seen from the experimental results that fusing the dynamic feature distillation information of the defect image into the recognition network can improve the efficiency of hydraulic tunnel defect recognition.

关 键 词：水工隧洞 缺陷识别 动态卷积 知识蒸馏 模型压缩 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]