Railway Anomaly detection model using synthetic defect images generated by CycleGAN

An Overview of Image Generation of Industrial Surface Defects

“…A self-encoder is a semi-supervised or unsupervised neural network which contains two parts, an encoder and a decoder; after the encoder learns the characterisation information of the input data, the corresponding characterisation information can be restored to the input information by the decoder, and its structure is shown in Figure 15: This structure can sufficiently extract the feature information of the image, which is very suitable for the image generation field, so the introduction of the self-encoder structure in the GAN framework can help in the generation of faulty images. For example, Hoshi T et al [65] added encoder and decoder structures to their generator and discriminator structures within the framework of CycleGAN network and added an attention mechanism between them, while Zhang G et al [66] added a weight-sharing self-encoder structure to the CycleGAN model, added adaptive noise to increase the diversity of generated defect samples in training, and added adaptive noise to increase the diversity of generated defect samples in training. By adding adaptive noise to increase the diversity of error samples generated, and adding a spatial and categorical control map to control the category and location of errors in the model, Figure 16 shows the model structure of the method.…”