Evaluation and comparison of techniques for reconstructing the point spread function of images blurred by uniform linear motion

Abstract

In the field of digital image processing, it is common to find different types of degradation. One of them is motion blur, which is caused by the relative movement between the camera and the observed object. It produces a low-contrast trace on the image that follows the trajectory of the movement. If the relative velocity is constant and the blur is invariant across the entire image, the resulting blur can be modeled by means of the Point Spread Function (PSF) and using the trace’s length and the angle parameters. This work evaluated the accuracy of the estimation of the angle and length parameters, and the robustness to Additive White Gaussian Noise of a set of spatial and frequency approaches for reconstructing the PSF. It is important to highlight that the algorithms’ processing time was also considered. In total, 20 512x512 pixels synthetically-degraded images were used. Besides, five of the best-known techniques for estimating the angle and three for the length of the PSF were evaluated. The experimental results revealed the techniques with the lowest absolute mean error for estimating the angle and the length of the PSF in noise-free images: 2D Cepstrum Transform and 1D Cepstrum Transform, respectively.

Keywords

• Cepstrum
• Motion blur
• Steerable Filters
• Linear Point Spread Function
• Reconstruction
• Hough transform
• Radon transform