A semi-supervised generative adversarial segmentation model for detecting defects from images

Insulation boards are one of the most commonly used electrical safety tool in both power generation and operation. Nevertheless, various defects are always produced in the production, transportation, and use of insulation boards due to factors such as human operation, equipment aging, and natural factors. These defects will significantly impact the performance of the insulation boards and pose safety risks, potentially leading to electrical safety incidents. Therefore, the detection of surface defects on insulation boards has become an essential step in their production and pre-use. However, due to the extensive human and financial resources required for annotating the dataset significantly hinder the progress of the detection task. Moreover, the current methods lack sufficient capability to characterize the features of small and weak defects on insulation boards surfaces, resulting in a low detection accuracy. In this paper, we propose a surface defect detection method for insulation boards based on a semi-supervised generative adversarial segmentation model, referred to as the SSGD-Seg model. This model comprises a segmentation network and a discriminator network, which employs a semi-supervised generative adversarial strategy by training only with a small amount of annotated training dataset. Furthermore, the Omni-Dimensional Dynamic Convolution module proposed in this article addresses the issue of overfitting in network training and enhances the network’s capability for feature extraction. The SSGD-Seg model has achieved outstanding defect detection performance on the insulation board surface defect dataset. The model trained with 20% annotated training data achieved an mIoU value of 83.52%, a DI score of 87.25%, and a recall rate (RE) of 96.83% on the test set.