In the surface detection recognition algorithms based on deep learning, a large amount of sample data was often required. For some newly established production lines, it was impossible to collect enough samples in a short period of time. This would result in inefficient detection which will affect efficiency in the early stage of a detecting system. In order to solve this problem, an enhanced generative adversarial network was adopted to perform image translation on image samples in other production lines to obtain the defect samples of new production lines, which was also called as cross domain image conversion. The surface defect samples of hot rolled steel sheets and the non defective samples of cold rolled steel strips were fusion converted into cold rolled strip surface defects samples. The experiment conducted cross domain conversion on the surface defects of six different types of hot rolled steel sheets. The results showed that the image conversion results with heavier background texture are better. For some defects with small defect scales, such as pitting, the detection results still have room for improvement. In order to quantitatively determine the detection result, a neural network was introduced to classify the original image and the translated image. The accuracy of classification results reached 96%, indicating that the image cross domain conversion effect was good and has certain application value.