Research On Processing Techniques For Low-quality Images

In our research,low quality images include:(1) Images containing uncomplete information of interest;(2) Images of low resolution.It is very impotant to do some researches on low quality images.In this paper, we select two typical images:low qulity convex and concavo characters on metal label and satellites images captured in simulated outspace environment.Our researches focus on the following topics:image preprocessing,features extraction and recognition classification.We proposed a method on metal label character area loacion based on prior information.Metal labels are made by special label-printer,so metal labels have a reliable standard which can be used for convex and cancavo characters location.Metal label images are interfered by some random natural elements,so the captured images' gray distribution is often asymmetrical,which takes a bad effect on the next processing.To this problem,we proposed a new method on charaters segmentation.For images captured by CCS lighting,we proposed a new method on illumination compensation based on spatial domain to solve the problem of asymmetrical gray distribution.In image preprocessing,some characters are often broken which takes a bad effect on next recognition,for this we proposed a new method on mending broken metal label characters based on structure and prior information.For images captured in simulated outspace environment,resolution is very low, so it is not proper to be processed with ordinary segmentation method.For this,we proposed a method on satellite segmentation based on gray and grids distribution.For circumferential projection,it is important to locate the projection center.Otherwise,metal label characters often have gaps,deforms when being captured.What is worse is this can take a bad effect on recognition,for this problem, we improved original features extraction based on circumferential projection. In our researches,we found some similar metal label characters are often recognized wrongly.So we proposed two levels recogtion.In the second level recognition,some boundary features are extracted for recognition.We testified a methon on space targets recognition based on wavelet moments and BP network.Experimental results showed we could not get sastisfactory recognition using this method.So we did some research on features combination and furthmore testified a method on space targets recognition based on image sequence using DS evidence theory.Our research is financially supported by a grant from the Ph.D.Programs Foundation(20060422011) and 683 state plan.Here we express thanks from the bottom of heart for them.