Semantic-Driven Context Aggregation Network for Underwater Image Enhancement

Abstract

Recently, underwater image enhancement has attracted broad attention due to its potential in ocean exploitation. Unfortunately, limited to the hand-crafted subjective ground truth for matching low-quality underwater images, existing techniques are less robust for some unseen scenarios and may be unfriendly to semantic-related vision tasks. To handle these issues, we aim at introducing the high-level semantic features extracted from a pre-trained classification network into the image enhancement task for improving robustness and semantic-sensitive potency. To be specific, we design an encoder-aggregation-decoder architecture for enhancement, in which a context aggregation residual block is tailored to improve the representational capacity of the original encoder-decoder. Then we introduce a multi-scale feature transformation module that transforms the extracted multi-scale semantic-level features, to improve the robustness and endow the semantic-sensitive property for the encoder-aggregation-decoder network. In addition, during the training phase, the pre-trained classification network is fixed to avoid introducing training costs. Extensive experiments demonstrate the superiority of our method against other state-of-the-art methods. We also apply our method into the salient object detection task to reveal our excellent semantic-sensitive ability.

This work is partially supported by the National Natural Science Foundation of China (Nos. 61922019, 61733002, and 61672125), LiaoNing Revitalization Talents Program (XLYC1807088), and the Fundamental Research Funds for the Central Universities.

D. Shi—Author is a student.