Nanjing University of Information Science and Technology
The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)
In view of the lack of samples caused by the difficulty in obtaining crack images and the insufficient feature space of traditional data expansion methods to enhance the sample feature, a crack sample expansion method based on generative adversarial network is proposed. This method improves the DCGAN network, namely MDCGAN (Modified DCGAN). Firstly, the data set is preprocessed, and the sliding window method is used for data dimensionality reduction and cleaning; Secondly, the activation function is optimized to improve the diversity of generation features. At the same time, spectral normalization is introduced for weight standardization to improve the stability of network structure, so as to generate high-quality crack data set. Finally, the improved Alexnet network is used to extract and classify the extended mixed sample set. The results show that the data enhancement performance of MDCGAN network is significantly improved compared with the traditional expansion method, and it is suitable for expanding crack images.