Analytical-Heuristic Modeling and Optimization for Low-Light Image Enhancement

TL;DR

The paper tackles low-light image enhancement by formulating an analytical-heuristic framework that fuses physics-based gamma correction with a novel dichotomy function, optimized via a genetic algorithm. The proposed Dichotomy Tuna pipeline processes images through HSV/YCbCr color spaces, color restoration, guided-filter denoising, and a linear combination of transformed channels, with a final gamma correction. Results on the LOL benchmarks show strong performance, including first place in PSNR on LOLv2-synthetic, achieved without any learning-based training. The work demonstrates that interpretable, analytically grounded models can rival deep-learning approaches in challenging vision tasks and suggests future directions at the intersection of swarm optimization and analytical vision.

Abstract

Low-light image enhancement remains an open problem, and the new wave of artificial intelligence is at the center of this problem. This work describes the use of genetic algorithms for optimizing analytical models that can improve the visualization of images with poor light. Genetic algorithms are part of metaheuristic approaches, which proved helpful in solving challenging optimization tasks. We propose two analytical methods combined with optimization reasoning to approach a solution to the physical and computational aspects of transforming dark images into visible ones. The experiments demonstrate that the proposed approach ranks at the top among 26 state-of-the-art algorithms in the LOL benchmark. The results show evidence that a simple genetic algorithm combined with analytical reasoning can defeat the current mainstream in a challenging computer vision task through controlled experiments and objective comparisons. This work opens interesting new research avenues for the swarm and evolutionary computation community and others interested in analytical and heuristic reasoning.

Figures (6)

• Figure 1: This flowchart shows all the steps we consider to improve the original Dichotomy model to highlight the visual appearance of images with under and over-exposed light regions.
• Figure 2: This collage shows the fifteen low-light testing images in the LOLv1 database, and the photos improved with the dichotomy model plus a Gaussian filter of $\sigma = 1$ presented in this work. Surprisingly, it reproduces tones like daytime, with a simple mathematical operation that contrasts state-of-the-art models based on complex reasoning.
• Figure 3: This collage shows the results of twelve low-light image enhancement algorithms tested on one image of the LOLv2-real database studied in this work. We can observe the problem of inverting the pixel values of Dichotomy+filter for overexposed regions and how Dichotomy Tuna solves the problem, although filtering and color issues remain.
• Figure 4: This collage shows the results of twelve low-light image enhancement algorithms tested on one image of the LOLv2-synthetic database studied in this work. Surprisingly, Dichotomy Tuna reproduces ground truth with simple analytical-heuristic reasoning that contrasts state-of-the-art models based on abstruse reasoning.
• Figure 5: This collage shows the results of Dichotomy Tuna applied to 15 images selected from LOLv2-synthetic to illustrate the algorithm output.

Theorems & Definitions (2)

• Definition 1
• Definition 2