Research on Generating Method of Fiber Pill Images Based on Wasserstein Deep Convolution GAN

The detection of fiber pills has always been a challenging task in the production of pharmaceutical factories. However, due to the tiny size and complex structure of fiber on pills, traditional detection methods face the problem of low detection accuracy. To improve the detection accuracy, the deep learning algorithm which could accurately capture the tiny features of fiber pills is a considerable method for fiber detection. However, the number of fiber pill samples are insufficient for the training of deep learning model, which makes it difficult to achieve high-precision detection. In this work, an exploratory study of fiber pill image generation is carried out based on Wasserstein Deep Convolution GAN (WDC-GAN). The WDC-GAN introduces deep convolutional networks into the generator and discriminator to compensate for the insufficient dataset problem of fiber pills. The Fréchet Inception Distance (FID) results have shown that the feature distribution of fiber pills generated by WDC-GAN model are very close to that of the real fiber pills, which has verified the generation ability of the proposed method.