The main contribution of this work is that we extend the cepstral analysis based PSF estimation to handle non-linear motion PSF. To introduce the cepstral approach for non-linear motion PSF estimation, we will analyze the cepstral behavior of non-linear motion. Then, I propose a non-linear motion PSF estimation method based on the analyzed cepstral behavior. The proposed method is validated with both synthesized images and real world images. The experimental results show that cepstral analysis based method can work for non-linear motion PSF estimation.
Simulation results for piecewise linear motion: Examples of the PSF shape estimation experiment. From top to bottom, latent images, blurred images, restored images by estimated PSFs, restored images by ground truth PSFs, and zoomed up of the restored images are shown. Red framed figures in restored images are PSFs used for deconvolution (for better visualization, we enlarge the PSFs 3 times the normal size). From left to right, PSF size is increasing.
Simulation results for non-linear motion: Experimental results: Restored images of the entire method. From left to right, MOUNTAIN, WOMAN, SHIP cases are shown with NCC values between the restored images. From top to bottom, original images, blurred images, restored images by estimated PSF, restored images by ground truth PSF, and zoomed up of the restored images are shown. Red framed figures in restored images are PSFs used for deconvolution (for better visualization, we enlarge the PSFs 3 times the normal size).
Real image results for piecewise linear motion: PLM PSF estimation for real world images. From top to bottom, TREE, FLOWER, DOLL, and TEXT scenes are shown. From left to right, blurred images, estimated PSFs, restored images. Captions of middle column images denotes estimated PSF size.
Real image results for non-linear motion: Experimental results: Restored images of the real world experiment. From left to right, DOLL, ORANGE, SIGN BOARD, and TEXT scenes are shown with the image resolution. From top to bottom, blurred images, restored images by our method, restored images by Fergus's method [Fergus et al., 2006], restored images by deconvblind [MathWorks], and Zoom up of restored images are shown. Red framed figures in restored images are PSFs used for deconvolution (for better visualization, we enlarge the PSFs 3 times the normal size) and each caption of restored images denotes the size of the estimated PSF. Other framed figures correspond to zoomed up regions of restored images.
Please contact me if you want any pre-prints.