Exposure Correction Framework via Vector Quantization for Image Enhancement

Photographs taken under improper exposures can appear either excessively dark or excessively bright. Most existing methods attempt to correct exposure in continuous representation, which often leads to low-quality results. In this paper, we introduce a novel exposure correction framework known as the Discretizing Exposure Network (DICE), which is designed to learn discrete exposure representations. To achieve this, our proposed framework is comprised of two key components: Exposure Discretization Module (EDM) and Color Condition Module (CCM). The EDM initially separates the feature into detail and exposure representations, subsequently learning discrete exposure features through vector quantization. Meanwhile, the CCM models the color distribution inherent in natural scenes, as improper exposure images lack color or detail information. Ex-tensive experiments demonstrate the effectiveness of the proposed method against state-of-the-art approaches both quantitatively and qualitatively.