Research And Implementation Of Face Recognition Methods Based On Local Features

With the development of related disciplines, the application fields of facerecognition are extending continuously, and the market share is improving fast. Butlimit to the environmental changes, different image acquisition devices, face changesand more other elements, there are still many key problems remain to be solved inface recognition technology. In order to improve the robustness of face image featuresagainst expression changes, variation of illumination, and accessories (shield) andother noise, this thesis focus on face features extract methods. After fully study anddeeply research the methods such as local binary patterns, weber local descriptor andmonogenic multi-scale representation, some new face features extract methods areproposed. In this paper, the main works are as follows:1. Based on the monogenic multi-scale representation, a face features extract methodwhich based on the monogenic local phase XOR patterns is proposed. After that, wefuse the two complementary features: monogenic magnitude local binary patternsand monogenic local phase XOR patterns using block-based strategy in score stage.Lastly, we discuss how quantization parameters and block numbers influent themonogenic local phase XOR patterns and how the algorithm performance changeswith different fusing weight of the two complementary features.2. In this part, we put forward multi-modal weber features descriptor methods basedon weber local descriptor. Multi-modal weber features mainly contain two parts:weber difference stimulation direction decomposition local binary patterns andweber direction difference patterns. We obtain the weber difference stimulationdirection decomposition local binary patterns by the way that dividing the weberdifference stimulation into some images which are in the same orientationsaccording to the quantification results of weber gradient directions. And then, codethe directional images with local binary patterns. As for weber gradient differencepatterns, we first calculate the difference of two pixels, then gain the absolute valueand code the result into a binary bit according to a empirical threshold. At the end,fuse the two features by block-based strategy in feature layer, and calculate thesimilarity with cosine distance rule. Finally, nearest neighbor classifiers are used toclassify the samples.3. We raise a new recognition algorithm called monogenic local difference binarybased on monogenic representation. This method can avoid quantization error using different modes, and it form a hybrid phase encoding with scale-scale informationand pixel encoding information by itself. In order to reduce the dimensions of thefeature vectors and resource consumption of algorithm and enhance the robustnessagainst noises, we calculate the dominant patterns of monogenic local phasedifference binary patterns by statistical strategy.4. Based on the monogenic representation and weber local description, monogenicweber difference excitation binary patterns and monogenic dominant orientationXOR patterns are presented in this paper. As for monogenic weber differenceexcitation binary patterns, we get weber difference excitation on monogenicmagnitude, and divide it into two parts which each other represent uniformvariation, then, we encode the result images by binary method. Another, we canobtain monogenic dominant orientation XOR patterns in the follow way. Firstly, weuse principal component analysis method to extract the dominant orientationinformation exactly. Secondly, encode the dominant information with XOR operator.Lastly, we calculate the histogram. Usually, we regard the two pattern features ascomplementation elements. So, in order to enhance the performance, we fuse themwith block-based strategy in score level.We test our algorithms on AR, CAS-PEAL, Yale and ORL databases. Theexperiment results show that the proposed face feature extract methods have strongrobustness to noises such as expression changes, light changes, accessories and facevariations.