ZHAO Ziyu1，2， WANG Xuquan1，2， MA Jie3， XING Yujie1，2， DUN Xiong1，2， WANG Zhanshan1，2， CHENG Xinbin1，2
1. Institute of Precision Optical Engineering， School of Physics Science and Engineering， Tongji University， Shanghai 200092， China； 2. MOE Key Laboratory of Advanced Micro-Structured Materials， Tongji University， Shanghai 200092， China； 3. College of Electronics and Information Engineering， Tongji University， Shanghai 200092， China
Abstract：By integrating intelligent algorithm-driven image processing techniques, computational imaging has the potential to transcend the limits of conventional hardware-centric optical systems, enabling optical systems to achieve high performance and a compact design. Focusing on the image reconstruction requirements in lightweight infrared single-lens computational imaging, this study investigated lightweight model deployment methodologies tailored for edge AI chips. Through targeted operator optimisation, model pruning, and quantisation implemented on edge devices, the deployed U-Net reconstruction model achieved a 52.3% reduction in parameters and a 60.3% reduction in computational operations, resulting in a 56% acceleration in edge processing frame rate while sacrificing only 0.91 dB in PSNR and 0.021 in SSIM. Further architectural simplification allowed ultra-high-speed video-rate on-chip image reconstruction exceeding 95 FPS, at the cost of just 1.3 dB PSNR and 0.018 SSIM. The experiments examined edge hardware acceleration for computational single-chip infrared camera reconstruction algorithms. This study provides technical references for engineering applications of lightweight infrared computational imaging systems.
