Semi-supervised Learning with GAN for Semantic Segmentation of Wood Defect Image

Automatic defect detection on wood surface is essential for ensuring product quality. Semantic segmentation methods have shown outstanding performance in wood defect detection. However, it is costly to acquire correct pixel-level labels of wood surface defects due to the complicated characteristics. To mitigate the cost of acquiring high-quality datasets, this study proposes a semi-supervised defect detection approach based on generative adversarial networks, which contains a generator and a discriminator. The generator is utilized to generate pixel-level segmentation results of wood defect images. To improve the segmentation performance, the discriminator learns against the generator by judging the goodness of the prediction results and generates supervised signals for the prediction results of unlabeled images. With this learning model, the proposed method can enhance segmentation performance using unlabeled images and alleviate the burden of data annotation. We validate the effectiveness of the proposed approach on a collected dataset of wood defects. Experimental results demonstrate the remarkable performance of our method even when utilizing only a small amount of annotated data.