Research On Image Restoration With Its Applications

Image restoration is an image reconstruction technology that image is rebuilt with some prior knowledge (degraded model) according to counter-process of degeneration. This dissertation involves the specific topics on the two models of degradation: noise jamming and motion blurring. The algorithms of noise removal and parameter identification for motion-blur are further studied and tested on the on-the-fly vision system of RFID flip chip bonder.Impulse noise can seriously deteriorate image quality and the performance of its filter affects directly the result of subsequent image proceeding. A novel filter based on mathematical morphology for high probability impulse noise removal is presented. Firstly, an impulse noise detector using mathematical residues is proposed to identify pixels which are contaminated by the salt or pepper noise. Then the image is restored using specialized open-close sequence algorithms that apply only to the noise pixels. Finally, black and white blocks which degrade the quality of the image will be recovered by a smart block erase method. Experimental results demonstrate that the proposed filter outperforms a number of existing algorithms and can remove most of the noises effectively while preserving image details very well.The point spread function for the image blur of uniform linear motion has been explored. Two methods based on frequency-domain and directional derivatives are proposed to identify two major blur parameters of the point spread function– blur direction and blur length. In the first method, the reason for the occurrence of black strips in the spectrum image is analyzed and the black strips that are perpendicular to the motion-blur direction are detected using radon transformation. The mathematical model of the relationship between blur length and the pitch of the dark lines is estimated based on the curve fitting algorithm. In the second method, the directional derivative is defined and the directional sub-pixel is calculated through bilinear interpolation. The angle corresponding to the minimum global directional derivatives is identified as blur direction. And the blur length is estimated by the minimal value of the derivative autocorrelation. Experimental results show that the frequency-domain method facilitates real-time calculation, while is not suitable for the noise image. The derivative method not only achieves accurate results, but also has strong noise immunity.A novel on-the-fly vision system applied to RFID flip-chip bonder has been presented. The feasibility and availability of the proposed algorithms above have been tested and verified on RFID devices. Based on the identified parameters, the Wiener filtering is carried out to recover the motion-blurred image. And the restored images are further analyzed by pattern matching. Examples of application show that the proposed algorithms can efficiently enhance image quality and also improve the accuracy and precision of pattern matching.