Competency Bank-Based Underwater Image Enhancement Framework

Recent efforts have made notable advancements on underwater image enhancement. However, due to the attempt to handle diverse degradation indiscriminately with a unified network, most methods are caught in a dilemma of low generalization capability and high resource consumption. Theoretically, retraining the mediocre network on each degradation-specific dataset and combining them can improve generalization, but this leads to an explosion in network parameters. Motivated by the observation that during the differentiation of a network from mediocre to specialized, only a small subset of parameters plays the critical role, we attempt to leverage the focused powerful competencies of specialized networks by identifying these key parameters. In this paper, we propose a novel Competency Bank based Underwater image enhancement framework (CBU) for efficient and robust processing. Specifically, our framework establishes a Clustering-Locating-Constructing (CLC) training strategy which constructs the Competency Bank by sequentially locating key degradation-specific parameters, forcing the differentiation of networks and harnessing the highly targeted capabilities of specialized networks. Furthermore, to endow the network with appropriate competencies, degradation maps are introduced to ensure the reliability of competency allocation and dynamical adjustment, termed as degradation-guided classifier. Comprehensive experiments demonstrate that switching the competencies allows the network to maximally enhance network’s generalization and performance with minimal parameter increase.