Erienet Achieves Efficient RAW Image Enhancement for Low-Light Conditions

Low-light image enhancement increasingly relies on RAW images due to their superior performance compared to standard formats, yet current methods often struggle to balance processing speed with image quality. Jianan Wang, Yang Hong, and colleagues from the Beijing Institute of Technology, along with Hesong Li, Tao Wang, Songrong Liu, and Ying Fu, address this challenge with the development of ERIENet, an efficient RAW Image Enhancement Network. This innovative approach simultaneously processes information at multiple scales using streamlined processing modules and uniquely leverages the rich data contained within the green channels of RAW images to guide reconstruction. The result is a network that not only outperforms existing state-of-the-art methods in enhancing low-light RAW images, but also achieves remarkable processing speeds, exceeding 146 frames per second for 4K resolution images on readily available hardware.

However, most existing methods for RAW-based low-light enhancement sequentially process multi-scale information, which limits lightweight models and high processing speeds. Additionally, they often overlook the advantages of green channels in RAW images and fail to fully utilize this information for improved reconstruction performance. The study pioneers a fully-parallel, multi-scale architecture that simultaneously processes information at different resolutions, unlike sequential methods that hinder processing speed and model efficiency. This approach utilizes novel channel-aware residual dense blocks to extract feature maps, significantly reducing computational costs and enabling real-time performance. The team engineered a system that specifically exploits the superior information content within the green channels of RAW images, recognizing that green pixels are sampled at twice the rate of red and blue pixels in Bayer pattern images.

They introduced a dedicated green channel guidance branch to extract illumination-sensitive features, employing spatial adaptive normalization to refine batch normalization parameters and optimize performance. This branch effectively leverages the increased spatial resolution and brightness information inherent in green channels, improving reconstruction quality with minimal additional parameters and computations. Experiments involving images from commonly used low-light enhancement datasets demonstrate that ERIENet outperforms state-of-the-art methods in enhancing RAW images with greater efficiency. The system achieves an optimal speed exceeding 146 frames-per-second (FPS) for 4K-resolution images on a single NVIDIA GeForce RTX 3090 with 24G memory, a significant advancement over existing techniques. Researchers validated the method on both the SID and ELD datasets, achieving superior performance in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and overall visual quality while minimizing memory usage and computational costs. This combination of high speed and accuracy highlights the effectiveness of the optimized network architecture and the innovative utilization of RAW data characteristics.

ERIENet Enhances RAW Images at High Speed

Scientists have developed a new image enhancement network, ERIENet, that delivers superior performance in low-light RAW image processing while maintaining high efficiency. The work addresses limitations in existing methods that often struggle to balance image quality with processing speed, particularly when dealing with the high resolution and complexity of RAW images. Experiments demonstrate that ERIENet outperforms state-of-the-art techniques in enhancing low-light RAW images, achieving an optimal speed of over 146 frames-per-second (FPS) for 4K-resolution images on a single NVIDIA GeForce RTX 3090 with 24G memory. The breakthrough lies in a multi-scale parallel architecture that processes image information simultaneously across multiple branches, significantly reducing computational time.

Researchers incorporated channel-aware residual dense blocks, which efficiently extract features and focus on important image details. Furthermore, the team introduced a novel green channel guidance mechanism that leverages the increased spatial resolution and brightness information inherent in the green pixels of RAW images. This guidance branch optimizes feature maps across different scales, using spatial adaptive normalization to refine batch normalization parameters. Testing on the SID and ELD datasets, ERIENet consistently achieved superior results in key performance indicators, including PSNR and SSIM, while also demonstrating improved visual quality.

The network processes 4K images at 146.2 FPS on the SID dataset, highlighting its ability to combine high speed and accuracy. This achievement is particularly significant as it allows for real-time processing of RAW images, even on resource-constrained devices, offering a practical solution for a wide range of applications. The method’s efficiency is also demonstrated by its lower memory and computational costs compared to existing techniques.

ERIENet Achieves Fast, High-Quality RAW Enhancement

This work presents ERIENet, a novel network designed for efficient and high-performance low-light enhancement of RAW images. The team developed a fully parallel architecture incorporating an efficient channel-aware residual dense block, which enables compact feature extraction with reduced computational costs and significantly improves processing speed, achieving over 146 frames per second for 4K images. Furthermore, ERIENet uniquely utilizes information from the green channels of RAW images to guide the overall feature extraction process, enhancing reconstruction quality. Experiments on standard datasets demonstrate that ERIENet outperforms existing state-of-the-art methods in both speed and efficiency, while also achieving strong results in objective image quality metrics and visual appearance.

Ablation studies confirm the effectiveness of the green channel guidance branch in improving feature extraction and pixel recovery, leading to better low-light enhancement. While ERIENet excels at enhancing single images, the authors acknowledge that exploring the potential of frame relationships in video remains an open challenge for future research. They also suggest integrating low-light enhancement techniques into broader computer vision tasks such as object detection and depth estimation, potentially advancing both high- and low-level visual processing.