An Improved GAN Model Based on Positive Samples for LED Die Defect Detection

The positive sample learning for defect detection can be applied to avoid the deficiency of deep learning networks. But there still exist the problems of simple defect generation and low detective accuracy. Thus, we put forward an improved GAN model for light emitting diode (LED) die defect detection. Firstly, we use GAN and the artificially generated defects to train the generator for fixing defects. Here, the real defect datasets and a random function are utilized in the defect generation module to generate multiple defects. An attention module is added to the discriminator to help the model detecting minor defects. Then, SSIM is used to compare the difference between the repaired image and the original image, and Otsu is utilized to detect and segment defects. What’s more, to improve the quality of the generated image, the SSIMLoss and NormalLoss are introduced into the original reconstruction loss function. Experiments on DAGM 2007 public dataset and the LEDWP dataset verify the efficiency of our proposed method compared to the one with no improvement introduction.