TPE: Lightweight Transformer Photo Enhancement Based on Curve Adjustment

Abstract

In recent years, learning-based methods have made great progress in the field of photo enhancement. However, the enhancement methods rely on complex network structures and consume excessive amounts of computing resources, which greatly increases the difficulty of their deployment on lightweight devices. Additionally, the methods have poor real-time performance when processing very large resolution images. In contrast to previous works on designing structurally diverse CNN networks, photo enhancement can be achieved through a lightweight self-attentive mechanism for global-local tuning. In this paper, we design a lightweight photo enhancement tool based on Transformer; we dub the tool, TPE. TPE captures long-range dependencies among image patches and can efficiently extract the structural relationships within an image. A multistage curve adjustment strategy overcomes the problem of the limited adjustment capabilities of the global adjustment function, allowing the method to combine both global modifications and local fine-tuning. Experiments on various benchmarks demonstrate the qualitative and quantitative advantages of TPE over state-of-the-art methods in photo retouching and low-light image enhancement tasks.