Semi-supervised Semantic Segmentation Algorithm for Video Frame Corruption

To address the problems of lack of labeled data and inaccurate segmentation in semantic segmentation of corrupted frame in surveillance video, a semi-supervised semantic segmentation method based on pseudo label filter with weak-strong perturbation and horizontal continuity enhancement module are proposed. The weak-strong perturbation-based pseudo label filter method performs selective re-training via prioritizing reliable unlabeled images based on holistic image-level stability. Concretely, weak-strong perturbations are applied on unlabeled images, and the discrepancy of their predictions serves as a measurement for stability of pseudo label. In addition, the horizontal continuity enhancement module is designed to make the model learn clearer inter-class boundaries of corrupted frame data. To validate the proposed method, the corrupted frame data from the security surveillance system are collected to produce a dataset for validation experiments. The experimental results show that our work outperforms other semi-supervised methods in terms of mean intersection over union (MIoU), demonstrating the effectiveness of the proposed method and the horizontal continuity enhancement module on the semantic segmentation task for corrupted frames.