NOLSAGAN: A NO-Label Self-Attention Segmentation Method Based on Feature Reconstruction Using Generative Adversarial Networks for Optical Fiber End-Face

Surface flaws on the optical fiber end-face cause signal attenuation, transmission loss, or an unstable optical fiber connection. Most deep-learning-based defect detection systems are inextricably linked to large defect images with pixel-level annotations, which are impossible to label precisely without massive manpower. To reduce faulty sample gathering and labeling tasks, this research proposes a novel unsupervised segmentation approach based on generative adversarial networks (GANs). We propose an anomaly detection mechanism based on convolution feature reconstruction, which exploits the discrimination and sparsity of convolution representation to effectively detect and identify potential anomaly patterns. To capture the long-distance dependence between pixels in different spatial positions, a defect attention module is proposed to supplement convolution, which is applied to the generator to strengthen the elimination of defect areas and suppress the change in background areas. Simultaneously, a new loss function considering the gray level and structural characteristics is proposed to constrain the distance between the true and false images, which can better simulate the internal structure of defects. The experimental results demonstrate that the proposed method is conspicuously effective compared with other unsupervised methods and yields excellent comprehensive performance on optical fiber datasets with a mean pixel accuracy (mPA) of 70.92%, with a mean intersection over union (mIoU) of 66.02%.