Resonator defect detection based on single image generative adversarial network and YOLO

In the manufacture of ceramic resonators, defects on surfaces can directly affect the working stability of resonator products. Due to extreme scarcity of abnormal resonator image samples containing defect textures, traditional computer vision algorithms have difficulty learning the key data distribution characteristics of abnormal resonators, leading to poor detection results. To solve this problem, this paper proposes a resonator defect detection method based on single image generation and deep learning classification. By training SinGAN model on a single real resonator defect image, we learn the spatial and textural features on the image and then we can utilize the trained model to create seemingly realistic fake surface crack images from several sketch map drawings, therefore effectively increasing the number of abnormal resonator samples. We train a YOLOv3 object detection model on the enlarged dataset only containing fake abnormal resonator samples and try to detect cracks on new real defect resonators. Experiments show that our proposed method has better image generation quality compared with previous methods and the YOLOv3 model bounds the real cracks successfully, proving that using a single defect sample to detect more and more defective images is feasible. More importantly, our method can be commonly used in classification and detection tasks of industrial products which have a similar data distribution.