Detection and estimation of image blur

The airborne imagery consisting of infrared (IR) and multispectral (MSI) images collected in 2009 under airborne mine and minefield detection program by Night Vision and Electronic Sensors Directorate (NVESD) was found to be severely blurred due to relative motion between the camera and the object and some of them with defocus blurs due to various reasons. Automated detection of blur due to motion and defocus blurs and the estimation of blur like point spread function for severely degraded images is an important task for processing and detection in such airborne imagery. Although several full reference and reduced reference methods are available in the literature, using no reference methods are desirable because there was no information of the degradation function and the original image data. In this thesis, three no reference algorithms viz. Haar wavelet (HAAR), modified Haar using singular value decomposition (SVD), and intentional blurring pixel difference (IBD) for blur detection are compared and their performance is qualified based on missed detections and false alarms. Three human subjects were chosen to perform subjective testing on randomly selected data sets and the truth for each frame was obtained from majority voting. The modified Haar algorithm (SVD) resulted in the least number of missed detections and least number of false alarms. This thesis also evaluates several methods for estimating the point spread function (PSF) of these degraded images. The Auto-correlation function (ACF), Hough transform (Hough) and steer Gaussian filter (SGF) based methods were tested on several synthetically motion blurred images and further validated on naturally blurred images. Statistics of pixel error estimate using these methods were computed based on 8640 artificially blurred image frames --Abstract, page iii