Robust parametric blur identification for motion blurred image under noisy conditions

Abstract

Identification of blur parameters is the key problem towards restoration of noisy motion blurred image. This paper presents a novel algorithm to estimate linear motion blur parameters such as motion direction and blur length under noisy conditions. The blurred image is pre-processed, dual Fourier transformed and bit-plane sliced subsequently. The blur angle is estimated accurately by radon transform of a specific bit plane selected through criteria based on image entropy. To mitigate the effect of noise on blur length estimation, the image is processed in bispectrum domain as the bispectrum has a characteristic to suppress additive Gaussian noise. A robust exponential model is proposed that represents the behaviour of blur length in bispectrum domain. The proposed algorithm is tested on several standard images. The maximum errors observed in the estimated blur angle and the blur length is less than 1.6° and 2 pixels respectively in presence of additive Gaussian noise. Experimental results obtained from the proposed algorithm are also compared with the previous algorithms to demonstrate its superior performance.