Image Fusion And Recognition On Multi-scale Analysis And Subspace Learning

Image fusion and recognition have always been research focuses in imageprocessing. Images can be acquired through various ways and possess rich contents.The relationships between images are complex. Different images generally containdifferent contents,or contain the same contents showing different forms. Based onmulti-scale technology and subspace technology combined with machine learning, thearticle proves into their applications in image fusion and recognition from someaspects to improve reliability and accuracy.Based on local features, combined with watershed and semi-supervisedalgorithms, two semi-supervised multi-object image segmentations are proposed.Watershed as a multi-scale method in spatial domain is used to segment images intoseparate enclosed areas as basic units. In image segmentation based onsemi-supervised clustering, objective functions that contain data items and penaltyitems are built. Labeled data and unlabeled data are assigned different weights duringiterative processes. Through obtaining optimal solutions, images are segmented. Inimage segmentation based on Multiway cut, a node hierarchy and a weightedundirected network are built based on interactions. The method introducesclass-terminal nodes and adopts two upper thresholds in watershed segmentation topre-segment images. It can achieve a coarse-to-fine segmentation.Feature registration and global registration may result in inaccurate registrationsor miss some registrations. Aiming at these problems, an adaptive image localregistration, based on SIFT (scale invariant feature transform) point registration andedge registration, is proposed. SIFT point registration is extended and scale invariantedge registration is implemented. The points with high registration reliability arechosen to calculate all the transforms hidden between two images. The method canprovide more registration information and reduce registration errors.A non-reference image quality assessment, based on SIFT density, and an imagefusion, based on SIFT relative difference maps and image enhancement trend maps,are proposed. Experiments show that SIFT density of an image, which is processed byneighborhood enhancement and double-size interpolated magnification, decreaseswith noise, bur and block effects increasing. It can accurately reflect image quality.The SIFT relative difference map reflects local quality difference for each pixelbetween two images. According to the differences, the images are partitioned intothree types of different regions. For the regions whose absolute relative differences are larger, corresponding image regions with higher quality are directly chosen asfusion results. Otherwise, images are fused according to an image enhancement trendmap. The image enhancement trend map, based on SIFT point registrations and edgeregistrations, is built to keep more common information and show pixel grayenhancement trends. The fusion method can enhance registration information, restraininterference signals and acquire more useful information.For image recognition, an image recognition, based on subspace and SIFT, andan image recognition, based on incremental and decremental subspace learning, areproposed. Subspace recognition is a global feature recognition method, which canrealize rapid image recognition and improve processing speed, but will ignore thedetails of image characteristics. SIFT provides a feature-point-registration voterecognition method, which use details of local regions to achieve registrations, but isnot fit for image batch processing. The combination of both of them can effectivelyimprove recognition accuracy for large number of images. Incremental anddecremental subspace learning can deal with dynamic sample databases, in whichsubspace is updated without having to retraining across all sample data when thesamples are increased or reduced. Based on the variances of sample projections, itestablishes approximate incremental and decremental learning formulas to increaseprocessing speed, reduce the dimensionality of feature vectors. The method canrealize online subspace learning and accurate image recognition.The article mainly centers on multi-scale analysis and subspace learning. Itproves into image fusion and recognition and provides some new approaches inrelated fields of image processing. All the approaches are verified by simulations.