Image Deconvolution Using Mixed-Order Salient Edge Selection

Salient edge selection is a crucial technique to warrant the success of image deblurring. Current edge-based methods mainly focus on the single salient edge without exploiting the rich structural information of different levels of the image. With this in mind, we propose an effective mixed-order salient edge selection for blind image deblurring, i.e., besides the salient edge based on first-order gradient, we further consider combining zero- and second-order information. We find that the finer image structure inscribed at zero-order repairs the important structure missing in the latent image, while the strong structure of salient edges depicted at second-order further enhances the latent image. The union of these three increases the robustness of the intermediate latent image, which leads to an accurate estimation of the kernel. Also, the inclusion of the gradient L-0-norm improves the quality of the recovery by preserving the favorable edges and removing the detrimental details. Experimental results show that the proposed method is much faster than the prior-based ones, and it provides more satisfactory recovery than the single salient edge-based approaches (e.g., in terms of error ratio, PSNR, SSIM, SSDE). Compared with state-of-the-art works, our method achieves better results on quantitative datasets and real-world images.