COLEA: A Low-light Enhancement Accelerator with Memory-usage Reducing Strategy

Convolutional neural networks (CNNs) have revolutionized many fields, including low-light enhancement, an important part of computer vision. Compared with traditional modelbased methods, the low-light enhancement with CNN achieves better performance. It not only improves the luminance component of images but also adjusts images’ chromatic components. However, the hardware resource and power requirement limit the deployment of low-light enhancement on embedded systems. In this brief, we present a CNN-based low-light enhancement accelerator, namely COLEA, with a hardware-algorithm co-design technique. We develop a down-sample and up-sample computing flow to reduce memory usage, making the COLEA process $4 \times$ bigger images. Based on the FPGA evaluation, our COLEA not only achieves superior performance but also offers significant benefits in terms of memory usage reduction, making it a convincing solution for low-light enhancement on embedded systems.