HIERA: High-Quality and High-Throughput Dehazing Hardware Accelerator with Reconfigurable Computing Unit
Junhao Zhang, Dongqi Fan, Liang Juan Chang · 2024
Haze environment hide the real objectives in the image, hindering the deployment of various vision tasks in a complex environment, such as object detection, surveillance, and remote sensing. Dehazing algorithms with neural networks can produce high-quality images with more algorithm parameters. However, existing hardware dehazing systems is difficult to meet the requirements of high image quality, low power, and high hardware utilization. In this paper, we propose a high-quality and high throughput dehazing hardware accelerator architecture, namely HIERA. The HIERA uses a neural network instead of traditional theory and contains reconfigurable computing units using DSP chains to reduce source usage and power consumption. In addition, we provide a 3D-LUT-based illumination enhancement module to obtain images with more natural illumination. Our network optimization strategy reduces memory usage and MAC operations by 26.7% and 57.3 %, respectively. The proposed HIERA accelerator is tested on the RESIDE dataset, achieving dehazed images with 18.03/21.27/21.24 dB PSNR, and the power consumption is 2.64 W. After combination with the illumination enhancement module, the average PSNR increased by 1 dB.