High-Frequency-Aware Multi-Task Learning and Transformation Consistency for Semi-Supervised Electromagnetic Shielding Optical Window Image Segmentation

Abstract

The electromagnetic shielding optical windows (OWs) are critical components in modern aircraft and electronic instruments. Accurate segmentation of crack template images can quantitatively analyze their structural parameters, thereby facilitating OWs design and manufacturing. However, pixel-level annotation is costly and labor-intensive. To address these problems, we propose a high-frequency-aware multi-task learning and transformation consistency for semi-supervised electromagnetic shielding optical window image segmentation network (HAMTC-Net) that enhances unlabeled data utilization through transformation consistency and high-frequency feature guidance. Specifically, CutMix and Mixup augmentations are incorporated to improve consistency regularization. An equivariant loss is introduced through an auxiliary classification task, which increases the global perception of the encoder. Furthermore, wavelet-based high-frequency features guide pixellevel consistency learning, enabling progressive learning from simple to complex patterns. Experiments on OWs crack image datasets demonstrate that HAMTC-Net outperforms existing state-of-the-art semi-supervised learning methods in segmentation accuracy.