13‐1: A Lightweight Inference Network‐based Algorithm for Low‐Light Image Brightness Adjustment

Low‐light images pose a challenge due to their compressed dynamic range, often resulting in loss of detail. To address this, enhancement techniques are evolving, aiming to better represent these images on modern displays. In this paper, we propose a novel algorithm that utilizes two cascaded neural networks to adjust image illumination effectively. Our approach begins with the first architecture, which employs 2D separable convolutional layers, ReLU, and sigmoid functions to extract essential features from RGB images. Subsequently, the second architecture, a U‐ shaped network, hierarchically adjusts image illumination, particularly in low‐light conditions. A significant contribution of our method lies in its ability to process low‐illumination images efficiently using a simpler, lightweight neural network architecture. This characteristic is crucial for its implementation in application‐specific integrated circuits for edge devices, as it reduces the number of trainable parameters, thus facilitating integer inference. We validate our approach using the widely used LOw‐Light (LOL) dataset, containing 500 pairs of low‐light and normal‐light images. Through experimentation, we demonstrate the efficacy of our method in adjusting brightness. Furthermore, we discuss the feasibility of hardware implementation, particularly emphasizing the suitability for designing very large‐ scale integration circuits on System‐on‐Chip (SOC) platforms. This is made possible by leveraging integer‐based matrix operations and implementing appropriate data rounding techniques to handle overflow computations effectively.