Research On Visual Perception Inspired Face Recognition Method

Automatic face recognition has attracted great interests of researchers and hasbecome a hot subject in computer vision and pattern recognition, as it has a broadapplication prospect in a variety of fields, such as military affairs, security, lawenforcement, finance, etc. Currently, automatic face recognition is able to achieverelatively satisfactory results under some specific circumstances. However, due to theimpact of a series of uncontrollable factors (e.g. illumination, expression, pose, imagingcondition, etc.), the performance of face recognition in a reality environmentdeteriorates drastically, and cannot meet the requirement of the practical application.The human visual system has an incomparable advantage in face recognition, which canidentify a person through the face without any efforts even in a complex environment.The face recognition performance of the human visual system is much better than thestate-of-the-art computer vision system. Thus, it is an important way to improve theperformance of automatic face recognition that designing recognition methods withreference to the mechanisms of human visual system. The work of this thesis isconducted following this idea, which explores new recognition methods inspired byvisual perception mechanisms to improve the accuracy and reliability of automatic facerecognition.The creative work of this thesis can be summarized as the following3points:①As the face recognitions methods based on statistics are difficult to distinguishthe image variations caused by interference factors and different individuals, aninvariant facial feature extraction model inspired by visual cortex mechanisms isproposed in this thesis, which eliminates the image variations caused by the interferencefactors, such as illumination, expression, etc., to improve the face recognitionperformance. This model constructs a two-layer hierarchical network, which simulatesthe functions of the primary visual cortex (V1), to eliminate image variations caused bythe interference factors, and therefore derives a set of invariant facial features. The firstlayer simulates the function of V1simple cells to extract a type of facial featuresinsensitive to illumination variation. In this layer, the learning mechanism of V1simplecells is simulated using the sparse coding model, which learns a set of filters thataccount for the local spatial structures of face from face images. And a few ofrepresentative band-pass filters are selected for facial feature extraction according to their reconstruction energy and frequency properties. Then this layer emulates the linearanalyzing function of the V1simple cells to extract the edge features that lie in differentspatial frequencies. As the edge features correspond to high frequency components ofthe image, and the illumination condition affects the high frequency components little,this type of edge features is invariant to illumination variation. The second layer of thenetwork, which simulates the function of V1complex cells, pools the output of the firstlayer in neighborhoods of position and scale by local maximum operation to furtherincrease the facial features’ robustness to expression, slight pose change and variationscaused by local transformation of facial parts. This type of invariant features eliminatesthe impacts of illumination, expression, etc effectively, and thus the accuracy andstability of face recognition are boosted.②In order to improve the face recognition accuracy by the important local facialfeatures, an attention mechanism inspired invariant facial feature extraction method isproposed by combining the visual attention mechanism and the hierarchical networkinspired by visual cortex mechanisms. The facial features extracted by the first layer ofthe hierarchical network are utilized to compute a salience map based on an informationmaximization attention theory. Then the salience map, which evaluates the importanceof different regions of the face, is converted to a weight map. In the computationprocess of the second layer of the network, the invariant facial features are weightedaccording to the weight map, and thus the contribution of the salient regions is enhanced.The holistic structure of the face plays an important role in face recognition, and theway of implementing the attention mechanism as a weight map is able to enhance theimportant regions while keeping the holistic structure of the face. The defect of treatingthe face image uniformly is avoided in this method, and the recognition accuracy isfurther improved.③A color face recognition method inspired by the opponent color perceptionmechanism is proposed to improve the comprehensive recognition performance, whichmakes extensive use of the information provided by the color image. In order toalleviate the impact of illumination variation to the color information, this method firsttransforms the color face image in RGB space to a form of opponent color, which hasone luminance component and two chromatic components, according to the opponentcolor perception mechanism of human. For the luminance component, the hierarchicalnetwork inspired by visual cortex mechanisms is employed for feature extraction, andthe features derived are treated as the texture features of the face. For the chromatic components, their low frequency components are extracted as the color features of theface. The texture features and color features are classified respectively, and the finalresult of face recognition is obtained by fusing the classification similarities of the twotypes of features. The color features are robust to image blur, and they can overcome thedefect of sensitivity to image blur of the texture features. The texture features areinvariant to illumination variation, and they can further make up the flaw of sensitivityto illumination condition of the color features. In addition, the texture features and colorfeatures describe different properties of the image, and they are complementary to eachother. Therefore, the proposed method acquires three advantages by fusing therecognition results of texture features and color features: first, it is robust to image blur;second, it overcomes the sensitivity to illumination variation of color image; third, itimproves the recognition accuracy effectively both when the image is clear and whenthe image is blurred.This thesis put forward new ideas for the developing of automatic face recognition,and offers some meaningful references for the application of visual perceptionmechanisms.