TreEnhance: A Tree Search Method For Low-Light Image Enhancement

TL;DR

TreEnhance is an innovative low-light image enhancement method combining Monte Carlo Tree Search and deep reinforcement learning, capable of producing high-quality enhanced images with minimal tuning and flexible application scenarios.

Contribution

It introduces a novel tree search-based framework for low-light image enhancement that integrates reinforcement learning and offers two inference strategies for accuracy and speed.

Findings

Achieved good qualitative and quantitative results on two datasets.

Provides a flexible, resolution-agnostic enhancement method.

Demonstrates effective guided search for reverse enhancement.

Abstract

In this paper we present TreEnhance, an automatic method for low-light image enhancement capable of improving the quality of digital images. The method combines tree search theory, and in particular the Monte Carlo Tree Search (MCTS) algorithm, with deep reinforcement learning. Given as input a low-light image, TreEnhance produces as output its enhanced version together with the sequence of image editing operations used to obtain it. During the training phase, the method repeatedly alternates two main phases: a generation phase, where a modified version of MCTS explores the space of image editing operations and selects the most promising sequence, and an optimization phase, where the parameters of a neural network, implementing the enhancement policy, are updated. Two different inference solutions are proposed for the enhancement of new images: one is based on MCTS and is more accurate but more time and memory consuming; the other directly applies the learned policy and is faster but slightly less precise. As a further contribution, we propose a guided search strategy that "reverses" the enhancement procedure that a photo editor applied to a given input image. Unlike other methods from the state of the art, TreEnhance does not pose any constraint on the image resolution and can be used in a variety of scenarios with minimal tuning. We tested the method on two datasets: the Low-Light dataset and the Adobe Five-K dataset obtaining good results from both a qualitative and a quantitative point of view.