Image Enhancement Based on Pigment Representation

TL;DR

This work tackles automatic image enhancement by introducing pigment representation, a content-adaptive, high-dimensional color space built per image. A visual encoder derives pigment expansion weights, reprojection offsets, and RGB reconstruction weights to convert RGB into $N$ pigments, reproject them via $L$ one-dimensional mappings, blend the results, and reconstruct RGB colors, all within a five-stage pipeline. The method achieves state-of-the-art results on photo retouching and tone mapping datasets (MIT-Adobe FiveK and PPR10K) with a compact model size and low latency, outperforming fixed color-space approaches and dense-pixel methods. The key contributions include enabling complex, component-correlated color mappings through abundant pigments and content-aware transformations, while maintaining efficiency and scalability. The approach promises practical impact for automatic image enhancement in consumer devices and professional workflows, by delivering high-quality, content-specific color refinements with minimal computational overhead.

Abstract

This paper presents a novel and efficient image enhancement method based on pigment representation. Unlike conventional methods where the color transformation is restricted to pre-defined color spaces like RGB, our method dynamically adapts to input content by transforming RGB colors into a high-dimensional feature space referred to as \textit{pigments}. The proposed pigment representation offers adaptability and expressiveness, achieving superior image enhancement performance. The proposed method involves transforming input RGB colors into high-dimensional pigments, which are then reprojected individually and blended to refine and aggregate the information of the colors in pigment spaces. Those pigments are then transformed back into RGB colors to generate an enhanced output image. The transformation and reprojection parameters are derived from the visual encoder which adaptively estimates such parameters based on the content in the input image. Extensive experimental results demonstrate the superior performance of the proposed method over state-of-the-art methods in image enhancement tasks, including image retouching and tone mapping, while maintaining relatively low computational complexity and small model size.

Figures (9)

• Figure 1: Comparisons of the (a) 1D LUT-based method, (b) 3D LUT-based method, and (c) the proposed pigment representation-based method. In (c), the input image is first converted into a set of pigments, which are then transformed using pigment reprojection functions. The reprojected pigments are subsequently combined to reconstruct the enhanced image.
• Figure 2: An overall framework of the proposed method.
• Figure 3: An example of pigment reprojection function.
• Figure 4: Qualitative comparisons on the MIT-Adobe FiveK dataset Adobe5K for image retouching: (a) shows GT (retouched by expert C) with its corresponding input image. (b), (c), and (d) show the resultant images and their corresponding error maps obtained by 4D LUT 4DLUT, CoTF CoTF, and the proposed method, respectively.
• Figure 5: Qualitative comparisons on the PPR10K dataset PPR10K for image retouching: (a) shows GT (retouched by expert a) with its corresponding input image. (b), (c), and (d) show the resultant images and their corresponding error maps obtained by AdaInt Adaint, RSFNet RSFNet, and the proposed method, respectively.