Patch-Based Transformer for Low-Light Image Enhancement

Images captured in low-light settings often manifest issues, notably reduced brightness and amplified noise. Addressing these concerns, our research elucidates a sophisticated technique crafted to enhance the fidelity of images taken under such adverse circumstances. During the image acquisition phase, the camera sensor meticulously logs the raw RGB data. This information subsequently undergoes a comprehensive optimization and refinement process facilitated by the camera’s Image Signal Processor (ISP), culminating in the finalized RGB depiction. We delve deeply into the intricacies of the ISP processing trajectory and unveil a methodology termed PLET (Patch-Based Low-light Enhancement Transformer), dedicated to refining low-light image quality to mirror that of well-lit conditions. For an intricate processing and comprehensive analysis of the image constituents, we employ the image patch approach, achieving not only pinpointed image refinement but also a profound comprehension of image content. Rigorous experimental assessments reveal that our approach excels in elevating image luminance and curtailing noise, cementing its efficacy in low-light image enhancement endeavors.

I. Introduction

With the proliferation of the Internet of Things (IoT) devices, capturing and transmitting visual data has become an integral part of modern interconnected systems. As a security surveillance cameras, wearable smart devices, and other IoT technologies penetrate our daily lives, the demand for high-quality imagery across various lighting conditions continues to grow [1]. Particularly, the challenges presented by low-light scenarios underscore the significance of image clarity and quality in the domain of digital image processing and computer vision [2]. Both in daily photography and specialized professional imaging applications, the quest for achieving impeccable image quality remains unceasing. Conventional enhancement strategies, such as histogram [3] equalization and Retinex [4] theory, have a provided foundational methodologies to ameliorate images with low luminance. Yet, these approaches often confront challenges in maintaining the image’s intrinsic visual aesthetics and in effectively curtailing noise amplification. The advent of deep learning has ushered in transformative shifts in this domain [5]. Leveraging the power of extensive data sets and state-of-the-art computational paradigms, deep neural architectures excel at discerning and amplifying intricate image features. Notably, Convolutional Neural Networks (CNN) [6] and Generative Adversarial Networks (GAN) [7] have emerged as frontrunners in the image enhancement arena, adeptly isolating nuanced features from dimly lit images to deliver them with unparalleled clarity and quality.