Research Of Automatic Target Recognition Technology Based On Image Invariant Features

With the rapid development of photonics, intelligent control, pattern recognitionand computer vision, Automatic Target Recognition (ATR) technology is more andmore important in modern weapon systems. In ATR field, the feature extractionmethod based on image invariants is one of the key technologies for achieving ATR.In this paper, on the basis of studying target recognition and image processing, theinvariant features are summarized and expanded. Moreover, taking aircraft as anobject of study, a series of key technologies on ATR from feature extraction to targetclassification are studied. So the research of this paper can provide some beneficialmethods for ATR technology. The main contributions are as follows:1. The invariants, which include global invariants and local invariants in aircraftrecognition, are summarized;2. For traditional aircraft recognition, the research works are mainly based onaircraft type recognition, and the collected image target has been thought as aircraftin advance. However, in practical applications, the collected image target isunknown in advance, so the image target recognition system can’t know that thecollected target is or not aircraft. For detecting the aircraft from a series of targetsautomatically, an automatic target recognition algorithm base on the improved bag ofword model, which is used by Fuzzy C-Mean (FCM) clustering, is proposed;3. A two-stage classification framework on aircraft automatic recognition isproposed: Coarse classification and Fine classification. Coarse classification detectsaircraft from a series of targets by the above improved bag of word model. Fineclassification identifies the aircraft type from a series of aircraft targets by universalinvariants with Support Vector Machine (SVM) or traditional neural network.4. The aircraft recognition performance is compared by the three kinds of globalinvariants: Hu moments, affine moments, normalized Fourier descriptors, with SVMclassifier. Then the above three kinds of invariants are combined as moment-fourierdescriptors. Moreover, for resolving the wide value range between feature vectors, the four normalization methods, which are combined with traditional neural networkor SVM separately, are studied and the choice principle of the normalizationmethods is proposed.5. A new kind of local invariant features, which is called as SIFT SequenceScale (SIFT-SS), are proposed. The invariance of the SIFT-SS is analyzed. Moreover,the recognition performance of SIFT-SS invariants, affine moments, Multi-ScaleAutoconvolution (MSA) are compared in clear images, added noised images andocclusion images. Experiments are shown that the SIFT-SS invariants are the highestrecognition rate in the above three kind of invariants.6. As feature fusion lie in some questions, such as the too large feature vectordimensions, some poor invariants can reduce the recognition rate of aircraftautomatic recognition system. So a new decision fusion method is proposed: first,for an aircraft image, different kinds of invariants are extracted and each kind ofinvariants is used to construct a SVM classifier. Second, many SVM classifiers arefused by the adaptive weights vote method and used for aircraft type recognition.In this paper, although the research object is aircraft, the proposed methods areuniversal for all image targets. So the methods can be easy to expand the imagerecognition for many mobile rigid objects, such as satellite recognition, missilerecognition, ship recognition, vehicle recognition.