Research on HLS Implementation of Self-Calibrated Illumination Neural Network: Low-Light Image Enhancement

Low-light image enhancement is crucial in applications such as nighttime surveillance, autonomous driving, and medical imaging. Traditional image enhancement methods like histogram equalization and Laplacian transformation can improve image brightness and contrast to some extent but struggle with complex lighting variations and detail preservation. Recently, deep learning-based methods have demonstrated superior performance but typically rely on high-performance computing resources, making real-time processing on resource- constrained embedded devices challenging. The paper addresses this problem by proposing a Self-Calibrated Illumination Neural Network (SCINN) based low-light picture enhancement method and implementing it with hardware acceleration through the use of High-Level Synthesis (HLS) tools. SCINN employs a self-calibration mechanism to adaptively adjust illumination parameters based on input image features, enhancing low-light images. In particular, SCINN is composed of three primary modules: image improvement, illumination correction, and feature extraction. It makes use of a Convolutional Neural Network (CNN) to improve the original image, extract features, and produce illumination adjustment parameters. Experimental results demonstrate that SCINN excels in improving image brightness and contrast while preserving details. Compared to traditional methods, SCINN shows significantly better enhancement effects. Additionally, the hardware-accelerated implementation using HLS tools on an FPGA shows that SCINN achieves efficient image enhancement with low hardware resource consumption, significantly improving processing speed and meeting real-time requirements.