关键词: 深度学习, 生成器, 鉴别器, 马尔可夫过程, 插值

Abstract: Manually marking the defect of the pin in the aerial survey of the drone is a time-consuming and labor- intensive task． To solve this problem，the deep learning algorithm ＲetinaNet was used to realize the automatic labe- ling of defects in pins． Given that if the unclear data collected by the drone is directly detected，the problem of low recognition accuracy can not be avoided，so a method for intelligent identification of defects based on generative ad- versarial network was proposed． That is to improve the local texture，edge and other details by the mutual game be- tween the generator and the discriminator，so as to improve the quality of such pictures and combine the defect in- telligent recognition algorithm model to extract accurate features to achieve defect detection． Finally，since the fuzzy- clear picture for generating network training is difficult to obtain，a fuzzy-clear picture pair was constructed by com- bining Markov process and sub-pixel interpolation． The experimental results show that the RetinaNet show excellent performance when detecting clear pictures，but for fuzzy picture，it can also lead to detection error．The constructed fuzzy-clear picture pair can effectively train the generated confrontation network to have deblurring function，which is beneficial to the convolutional neural network to extract more abundant features，thereby further improving the recognition rate of blurred pictures．

Key words: deep learning, generator, discriminator, Markov process, interpolation