Image Frequency Separation Residual Network for End-to-end RAW to RGB Mapping

Due to the limitations of hardware specification of smartphones' camera system, there is still a visible gap in imaging quality between smartphones and digital singlelens reflex (DSLR) cameras. Sophisticated learning-based image processing becomes a promising solution to close this gap. In this paper, we propose an Image Frequency Separation Residual Network (IFS Net) to perform the end-to-end RAW to RGB image mapping. Different from existing methods that directly train the input image and the ground truth image one-to-one as a whole, our proposed method first divides the input image and the ground truth into high-frequency and low-frequency parts by discrete wavelet transform (DWT). These two parts are then trained separately using different networks for details and global information, and finally synthesized into the output image using inverse DWT. Experimental results show that the proposed IFS Net outperforms other existing algorithms in both PSNR and SSIM. Visual comparison shows that the images produces by IFS Net preserves more details and look close to that captured by DSLR cameras.