PDE: Gene Effect Inspired Parameter Dynamic Evolution for Low-light Image Enhancement

TL;DR

This work identifies a gene-like phenomenon in low-light image enhancement where resetting learned parameters to random values can improve some images, revealing that static parameters limit adaptability. It introduces Parameter Dynamic Evolution (PDE) with Parameter Orthogonal Generation (POG) to dynamically evolve parameters, simulating gene mutation and recombination and reducing the gene effect. Empirical results on LOL datasets show PDE reduces the gene-effect metric (DGE) and yields PSNR gains of about $1.0$ dB over Restormer and $0.3$ dB over CIDNet, with ablations validating the benefits of PDE and POG. The study offers a biology-inspired, plug-and-play approach to enhancing LLIE robustness and adaptability, with code slated for public release.

Abstract

Low-light image enhancement (LLIE) is a fundamental task in computational photography, aiming to improve illumination, reduce noise, and enhance image quality. While recent advancements focus on designing increasingly complex neural network models, we observe a peculiar phenomenon: resetting certain parameters to random values unexpectedly improves enhancement performance for some images. Drawing inspiration from biological genes, we term this phenomenon the gene effect. The gene effect limits enhancement performance, as even random parameters can sometimes outperform learned ones, preventing models from fully utilizing their capacity. In this paper, we investigate the reason and propose a solution. Based on our observations, we attribute the gene effect to static parameters, analogous to how fixed genetic configurations become maladaptive when environments change. Inspired by biological evolution, where adaptation to new environments relies on gene mutation and recombination, we propose parameter dynamic evolution (PDE) to adapt to different images and mitigate the gene effect. PDE employs a parameter orthogonal generation technique and the corresponding generated parameters to simulate gene recombination and gene mutation, separately. Experiments validate the effectiveness of our techniques. The code will be released to the public.

Figures (9)

• Figure 1: Overview of gene effect. Resetting certain parameters to random values can even improve enhancement performance for some images. Inspired by biological genes, where random mutations benefit some individuals while harming others, we name this peculiar phenomenon “gene effect”.
• Figure 2: Biological gene evolution. Biological individuals rely on gene mutation and gene recombination to dynamically evolve gene and adapt to new environments. For gene mutation, new genetic information are generated. For gene recombination, different genetic information are changed. Inspired by biological gene evolution, we propose PDE, which employs a parameter orthogonal generation technique and the corresponding generated parameters to simulate gene recombination and gene mutation, separately.
• Figure 3: Gene effect in low-light image enhancement models. From left to right, the images are as follows: the low-light image, the image enhanced by the original well-trained Restormer Restormer, the image enhanced by the Restormer in which certain attention mechanism parameters have been reset to random values, and the reference image. The image enhanced with the well-trained parameters exhibits overexposure and fading color, with only 12.72 dB. In contrast, the images enhanced with random parameters show even higher PSNR values, along with more accurate color and brightness.
• Figure 4: Comparison of the generated dynamic parameters.The top row presents the basic framework of current parameter mechanism, where the dynamic convolution employ convolutions to extract weights based on the input image to weight the candidate parameters. The bottom row presents comparison of the generated dynamic parameters. The dynamic parameters generated by ours for each row and column image exhibit gradual evolution processes, indicating the ability to recognize differences and understand similarities between these images. (The input low-light images in the comparison have been brightened for better visibility.)
• Figure 5: The similar low light images and corresponding different high light images. The training images do not even satisfy the same distribution, as similar low-light images are likely to be mapped to different high-light image. Yet LLIE methods typically are composed of static parameters. This forces the model to learn a general mapping, only well-suited for a part of images but become maladaptive and even harmful when facing other particular images.