| Vision based object recognition is widely used for both civil used and nationaldefense used. For decades, the related technique including feature extraction andfeature recognition has been the hotspot of research in and abroad. Because thecorrected targets are usually small object with the scale of several to tens ofdiffraction limit, and also only partical error correction can be achieved of dynamicrandom pertubation atmospheric turbulence, instead of perfect correction, real-timecorrection images acquired by Adaptive Optics system contain high-frequenct error,represented by blurry contour and unrecognizable detail. Because of this peculiarity,on studied images, general algorithm cannot perform as well as on common digitalimages. In particularly, this degradation of performance manifests as therepeatability decreasing of point detector and discrimination decline of descriptor.Therefore, it is necessary to study the method of feature extraction and recognition,focus on images obtained in Adaptive Optics imaging system.Aiming at the difference of performance of general algorithm working oncommon images and the studied images, in this article, leading edge technique ofpoint detectors and local feature descriptors is combined innovatively to study andexplore several problems of this filed.Saliency provides an available approach to improve point detectors; binarydescriptor system opens a new road to increase the discrimination between featuredescriptions. Since the former is sensitive to local patches of images which owe highamount of information, it is used in point detectors, respectively as the method offeature selection and adaptor of a detector itself. The latter derived from randomforest classification, has the advantage of fast speed and well describing-ability, sothat invariance and discrimination of local feature descriptor can be guaranteed. Theunion of these two gives improvement in feature extraction and recognition used instudied images, compared with traditional methods, and also provides notablesuperiority in computing speed.Evaluation of general point detectors and descriptors is programmed, on theplatform of Matlab and C++/OpenCv. Recall rate and1-precision is used as thecriterion of the performance both in common images and studied images, andsolution is proposed based on the result of evaluation. Several point detectors are involved, including Harris, GFTT, FAST, DoG and DoH. And also, severaldescriptors are participated in, including SIFT, SURF and BRIEF. This workestablishes the theoretical basis of the following works.The mainly content studied in this article is listed as following:First, feature selection method of points obtained by point detectors is studied.Saliency is adopted for the reason that saliency is sensitive to high amount ofinformation local patch. And as the tool of feature selection, SURF features arefiltered and their characteristic scales are also refined, to get stable features.Experiment shows: this method has good tolerance of invariance of featureappearance; discrimination of selected features is increased, which indicates thatabandoned features are less stable. Besides, the weakness is reduced with assistant offast point detectors, that salient region extraction method cost plenty of time, and thebalance between algorithm speed and sufficient feature extraction is improved.Second, a method is studied of injecting anti-noise into a general detector itself.Entropy is used as the weight of alternative features aiming to image noises, so that2-step Harris is proposed based on classical Harris. On one hand, Harris threshold isauto selected in order to gain sufficient alternative corners; on the other hand, localentropy as the weight, provides corners with more possibility to locate on object andtherefore has higher amount of information. Experiment shows: compared withgeneral algorithm, this method can still obtain corners on object stably even understrong noises. Especially under salt&pepper noise, this method achieves more thantwice of repeatability compared with classical Harris. Repeatability increasessignificantly from20%~30%of classical Harris to70%~90%, and the rate of cornerlocating in noises is reduced.Third, a method is studied of achieving feature recognition without featuredescriptor calculation. Classifier is used for feature matching, with the combinationof Ferns classifier and Adaboost framework, in order to locate object stably. Therobustness to feature variance can be configured by restricting the variance oftraining set on scale and rotation; Adaboost brings the solution of decrease of Fernsrecognition, leaded to discrimination decline of images patch as the lack of imagedetail. Experiment shows: boosted classifier achieves more stable feature recognitionthan SURF, and has notable higher recognition rate in the restricting condition oftraining set. The improvement of boosted classifier is also considerable comparedwith Ferns. Forth, a method of fast feature recognition is studied. Making use of theadvantage of both the binary descriptor BRIEF which is fast, low-dimensional andgood in discrimination, and high-speed corner detector FAST, fast feature detect andrecognition is achieved. Focusing on the weakness that BRIEF is highly sensitive torotation, rotation invariance is introduced in descriptor, so that the performance onstudied images of this method is comparable to that on common images. Experimentshows: this method is10times in speed compared with SURF, while it performsequal level recognition. This indicates that this method achieves efficient featurerecognition.In is article, general point feature detect and feature recognition methods arecropped and improved, aiming to the studied images which owe the peculiarity ofblurry contour and unrecognizable in detail. The improvement of algorithm is gainedin practice. The studied content in this article focus on stable feature extraction andefficient feature description, involves the analyses of problems, the propose ofsolution and implement. Experiment shows: for studied images, features with highvisual saliency are more probably stable and have better robustness under the changeof rotation, scale and view. Visual saliency can be used to improve the repeatabilityof detectors. Introduce of binary descriptor and method without descriptorcalculation can be more propriety for studied images. Compared with traditionalhistogram based method, these two perform comparable with lower computecomplexity, and hence solve the problem of decrease of feature recognition rate,caused by lack of image detail. Besides, the evaluation of general algorithm on bothcommon images and studied images provides theoretical basis and projectexperience for further study on this kind of images, such as feature tracking,category recognition of object, etc... |
PDF Full Text Request