Semi-DinatNet: Two-Stage Underwater Image Enhancement With Semi-Supervised Learning

Abstract

Underwater image enhancement (UIE) is crucial for marine surveys, attracting significant research interest. The underwater environment often results in images with low illumination, blurriness, and color distortion, challenging the effectiveness of underwater image enhancement. Despite progress in deep learning-based UIE methods, many still struggle with brightness enhancement and detail restoration. Additionally, most methods rely on supervised learning, which performs poorly on unlabeled underwater images. We introduce a two-stage model for enhancing underwater images. The first stage employs a U-shaped Transformer with the dilated neighborhood attention transformer (DNAT) as the bottleneck layer and the channel-wise multi-scale feature fusion transformer (CMSFFT) in the up-sampling and down-sampling phases. DNAT captures long-range dependencies and models global features, Improving the network’s capacity to concentrate on both overall image areas and detailed textures. CMSFFT improves attention to attenuated color channels. The second stage includes an illumination-perception processing branch as a subsidiary network. By introducing an illumination guidance block (IGB) between the multi-dilated convolution block (MDB) and residual contextual block (RCB), the network better perceives color and light source information. Feature fusion is achieved through attention feature fusion (AFF), integrating the illumination-perception processing branch with the backbone network. Our training paradigm uses an optimized knowledge distillation approach for semi-supervised learning, enhancing the model’s efficiency with unlabeled underwater images and ensuring robust performance across various scenarios. Numerous experiments across various underwater image datasets confirm the exceptional performance of the proposed method.