Four Way Local Binary Pattern for Gender Classification Using Periocular Images

Abstract

Human face can represent such information that would be difficult to express using thousands of words. For this reason, human face has always been an activefieldofresearchinComputerVisionandImageProcessing. Facecarriesvital information such as identity, gender, mood etc. of an individual. The power of discrimination and recognition come naturally to human being while machine does not have this trait. Machine cannot extract information from face with such ease like human being. Lots of painstaking efforts are required to make machine a strong competitor to human being. To make the machine performing close to human being we need to extract the discriminating features very efficiently. Several image descriptors have been proposed for facial image description. But they are not robust in all aspects. For an image descriptor to be a good one it should havesomequalityofrobustnessagainstdifferentchallengingenvironmentssuch as pose variation, illumination change, rotation, noise, scaling etc. So, finding a good image descriptor is a challenging task in Image Processing applications. To extract features from a face image, several descriptors such as global and local are proposed. Global descriptors such as Principal Component Analysis, Linear Discriminant Analysis, etc. adopt a holistic approach for representing the face image which does not encode local information and it degrades the performance of the algorithms. To encounter these problems, many local texture based approaches have been proposed such as Local Binary Pattern (LBP) which encodes the local texture information of an image and represents them globally. But, performanceofLBPdegradesinpresenceofnoise,rotationandnon-monotonicgray scale changes. These image descriptors are used in various Computer Vision applications such as face recognition, facial expression recognition, age estimation and gender classification etc. Among these, gender classification is a recent field of research that has grown interest among the researchers. Automatic gender classificationcanimprovetheperformanceofmanycomputervisionapplications such as security, surveillance, human interaction etc. Many facial image based gender classification techniques have been proposed by the researchers. They use the full face for gender classification. Most of them use Global or Local feature extractors for extracting the image features and then give them to machine learning algorithms such as Support Vector Machine, Nearest Neighbor Classifier,NeuralNetworkforclassifyinggender. Followingafullfacebasedapproach degrades the classification accuracy when the face is occluded. Occlusion is a naturalphenomenonthatisverycommoninrealworld. Fullfaceimagemaynot be available all the time for classification. So, the performance drops drastically. Moreover, the proposed approaches use the existing feature extractors. So, they incorporate all the drawbacks of these algorithms which is another performance degradingfactor. Periocularregionisattractingmuchinterestrecentlyduetothe occlusion factor. Periocular region refers to an area in immediate vicinity of eye. Duetotheavailabilityofperiocularimagesoverfullfaceimage,researchershave used this for different classification purpose. Considering all this limitations of genderclassificationfromfacialimages,weproposeanovelimagedescriptor,the Four Way Local Binary Pattern (FWLBP) for classifying gender from periocular images. ThisproposedmethodgeneratesabinarybitpatterncalledFWLBPcode by thresholding the image pixels of a 3 x 3 neighborhood in four ways (top, left, right, bottom) with respect to three horizontally and vertically centered pixels. This gives a better representation of edges in four ways which is a much more robustinformationthanintensityinformation. Theproposedmethod’sefficiency istestedonpubliclyavailableFERETdatabase. Periocularimageswereextracted fromfacialimagesandFWLBPwasusedtoextractthefeatureswhichweregiven toSupportVectorMachineforclassifyinggender. Fromexperimentalresults,itis seen that our proposed method gives better classification accuracy than the traditional LBP method both in general circumstances and in different challenging factors such as rotation, out of focus etc.