Image generation method for steel surface defects on improved CycleGAN

Abstract

Aiming at the difficulty of sample collection, high cost, and small sample problems caused by many types of defects and difficulty in covering all small samples in the process of industrial steel surface defect detection, a steel surface defect image generation method improved by cycle-consistent generation adversarial network (CycleGAN) was proposed. Firstly, channel attention map (CAM) and spatial attention map (SAM) were embedded in the CycleGAN model to enhance the feature extraction ability of the model. Secondly, the Weight Demodulation (WD) mechanism was introduced to fix feature artifacts and white spots, further improving the quality of the generated images. Thirdly, shape consistency loss was introduced to supervise the generator training process to solve the problem of inherent ambiguity in the process of image geometric transformation. Finally,the model before and after the improvement were experimented on the NEU-DET dataset. The results show that the improved model has more diversity and accuracy in the effect of defect image generation. PSNR and SSIM increase by 13.0% and 7.8% respectively, and FID values are reduced by 33.1%. This method can stably generate high-quality images of various steel surface defects, which can achieve the purpose of increasing training data, and also has reference value for the amplification of other defect datasets.

Keywords

• computer neural networks; image generation; attention mechanism; weight demodulation; loss of shape consistency