AI driven Wide Dynamic Range CMOS Image Sensor

This work proposes the implementation of a wide dynamic range CMOS image sensors using on-chip DNN. A given pixel is read in the linear mode under low light illumination and in logarithmic mode under high light where the mode of operation of pixel is decided using comparator. The compressive nature of logarithmic pixel is leveraged using amplifier to enhance the dynamic range. To further enhance the dynamic range, three hardware efficient DNN models, SiDCNet, SiDCNet (inr.) and SiDCNet (fire), are implemented and compared. On an average, the PSNR and SSIM of 31.17 dB and 0.95, respectively, is observed on standard test images. A comparison between post DNN inference is drawn with the linear-logarithmic acquisition, resulting an increase in image quality of 8.5 dB and 0.14 in terms of PSNR and SSIM.