Diffusion Model for Motion Deblurring with Directional Separated Decoders

Various types of blur that can occur in images pose problems in computer vision tasks. Image deblurring has been studied extensively to address these problems. With recent advances in deep learning, image deblurring using U-Net methods has demonstrated an improved performance. However, these models have limitations in dealing with motion blur with directional properties. In this study, we propose a diffusion model for motion deblurring with directional separated decoders to address these limitations. Our model learns the horizontal and vertical directions of motion blur with dual decoders and uses multiscale structure guidance to improve performance. Our model shows improved results with a PSNR of 32.47 when tested on the GOPRO data. This confirms that D6 effectively performs motion blur image restoration and produces clean images.