YUVDR: A residual network for image deblurring in YUV color space

Motion blur removal caused by camera shake and object motion in 3D space has long been a challenge in computer vision. Although RGB images are commonly used as input data for CNN-based image deblurring, their inherent issues of color overlap and high dimensionality can limit performance. To address these problems, we propose YUVDR, a residual network based on YUV color space, for image deblurring. By using YUV images, we mitigate the issues of color overlap and mutual influence. We introduce novel loss functions and conduct experiments on three datasets, namely GoPro, DVD and NFS, which offer a wide range of image quality levels, scene complexities, and types of motion blur. Our proposed method outperforms state-of-the-art algorithms, yielding a 3-5 dB improvement in the PSNR of test results. In addition, utilizing the YUV color space as the input data can greatly reduce the number of training parameters and model size, by approximately 15 times. This optimization of GPU memory not only improves training efficiency, but also reduces testing time for practical applications.