Comprehensive Recognition Methods Based On Data Fusion

In recent years, with the rapid development of modern science, the informationcontained in data is huge and the complexity of data has been dramatically raised. Howto extract useful information from huge amount of data to provide comprehensiveunderstanding and recognition of target has attracted a lot of interest at home and abroad.Data fusion integrates big amount of data and represents it in a low-dimension form,thus can better reflect the essential characteristics of data. Brain tumor recognitionutilizing data fusion techniques has been a tough issue in the area of electronicinformation and biomedical engineering. Non-negative matrix factorization (NMF) is anadvanced blind source separation technique. Under the non-negative constraint, NMFalgorithm is able to reduce dimensions, extract features, and fuse information ofcomplex data, and then generates satisfactory results. According to practical needs, theimproved, expanded and optimized NMF algorithms have shown a great success inbiomedical signal processing, remote sensing data processing, speech signal processing,text data mining, et al.Based on data fusion method, this thesis focuses on tissue type recognition ofgliomas and target recognition using multi-sensor remote data. The main contribution ispresented as follows:1. To solve the inadequate target recognition accuracy using modern NMFalgorithms, the popular NMF algorithms are analyzed and their characteristics arediscussed. Their accuracy of target recognition is compared and analyzed usingmagnetic resonance spectroscopy imaging (MRSI) data of glioblastoma (GBM) ofhuman brain.2. The hierarchical NMF (hNMF) algorithm is proposed for two goals:1) toprovide a better interpretation of MRSI data;2) to give robust recognition of GBMtissue types (i.e. normal tissue, tumor, and necrosis). The algorithm constructs varies ofmasks under the assumption that spectra of the three tissue types are least correlated.NMF is applied hierarchically to recover tissue-specific spectra and visualization isrealized using non-negative least squares. The proposed algorithm has improved the accuracy of brain tumor tissue type recognition and their corresponding spatialdistribution.3. An unsupervised nosologic imaging method for glioma diagnosis is proposed toavoid large training set of labeled spectra like other nosologic imaging methods. Thismethod recognizes different tissues in MRSI data and estimates their spatial distributionusing NMF and hNMF and interprets them in one brain image using different colors,thus solves nosologic imaging for gliomas without any prior knowledge. Meanwhile, anerror-map estimation method based on linear least squares is proposed to provideconfidential information of nosologic images.4. A novel algorithm for the image registration between synthetic aperture radar(SAR) image and optical image is proposed. Under the level set framework, an energyfunction is defined with the matching function between two images, and the functioncan be solved by curve evolution. With this approach, the accuracy of registration couldnot be affected by speckle noises or the result of feature extraction.5. A time-sharing fusion method based on NMF is proposed for solving the targetconfusing problem in change detection using multi-band SPOT images. The changedarea is recognized with constructed residual images which are generated by imagefusion based on NMF. With the proposed method, the recognition accuracy of thechanged area is improved.