Image Generator Comparison for Abnormality Inspection Without Defect Data

Instead of using a discriminator to directly identify abnormalities, the invention inverts the approach by using a generator to create normal images and comparing them with actual inspection images. The generator creates images of normal products, and the discriminator then identifies abnormalities by detecting differences between generated normal images and actual inspection images, eliminating the need for extensive abnormal teaching data.

The invention performs preliminary action by pre-training a generator model on normal product images before actual inspection. The generator is trained in advance to create realistic normal product images, which are then used during inspection to compare against actual images. This preliminary training on normal data alone enables the system to detect abnormalities without requiring pre-collected abnormal images.

The invention inverts the traditional approach by training the generator only on normal product images rather than collecting diverse abnormal images. The generator learns to create normal images, and abnormalities are detected by comparing actual images against these generated normal images, eliminating the time-consuming process of collecting and preparing abnormal teaching data.

The generator creates synthetic copies of normal product images that can be used for comparison during inspection. These generated images serve as reference copies of what normal products should look like, enabling accurate abnormality detection without needing physical samples of every possible defect type.

The invention simplifies the system by inverting the training approach: instead of training a discriminator to recognize multiple abnormal modes directly, a generator is trained only on normal modes. This single training process provides universal adaptability because the generator learns the fundamental characteristics of normal products, enabling detection of any deviation from normality regardless of the specific abnormal mode.

The generator trained on normal images serves a universal function for detecting all types of abnormalities. Rather than requiring separate training for each abnormal mode, the single generator model provides multi-functional capability to detect any deviation from normal product appearance, simplifying the inspection system while maintaining versatility.