Several local feature descriptors were introduced for facialimage analysis. Among those LBP, LGP, LTP, CENTRIST andNABP have been proposed for the classification of images. Thetechniques of those descriptors are traced in this section.Among all the feature descriptors, Ojala et al. first introduceoriginal LBP operator which thresholds n×n neighborhood ofevery pixel of an image with the center pixel value and considersthe result as a binary number 31.
LBPs were designedfor texture description, and induced for face representationin several applications including face detection 19, 11,18, face recognition 46, 1, 54, flower classification49,object classification , leaf classification , scene classification,expression recognition 41, 10, 44, gender classification36, 45 and texture classification 43.LBP is additionally utilized by Shan et al. 41 for expressionrecognition.
They divided the face image into manysubregions of different size and extracted options from solelyfew sub-regions which are classified using boosted SVM 3afterwards. Beyond that, real time gender recognition usingboosted LBP features was sought by C. Shan 40 . Inspiredby the tremendous performance of LBP operator, Tan andTriggs 46 amplify the binary pattern into ternary patternwhich encodes facial images using a fixed threshold (±5).Although LBP has gained popularity for its simplicity, itfails to differentiate a small difference and a large differencein acuities which deteriorates its preferential capacity. Anotherweakness of LBP is that it can be affected by the noise dueto local intensity fluctuation especially in uniform and nearuniform regions.Jun et al.
19 made an exploration to use adaptive thresholdin LBP and proposed LGP for face and human detectionwhere n × n neighborhood of a pixel is considered, and theneighbor having gradient greater than or equal to the averageof gradients of eight neighboring pixels, is set to a binary valueof “1”, otherwise is assigned a binary value of “0”.We et al. 49 introduced CENTRIST which is a visualfeature descriptor for scene and object classification whichperforms a census transform (CT) of an image and replacesthe image with its CT values 49. CT is a non-parametric localtransformation designed for establishing relationships betweenlocal patches 52, which is calculated similarly as LBP.Rahman et al.
37 proposed Noise Adaptive Binary Pattern(NABP) for facial image analysis such as face recognition,expression recognition and gender classification. NABP encodesthe face microstructures using an adaptive threshold andgenerates more discriminative patterns than other existing localfeature descriptors.The local feature descriptors so far we depicted were proposedfor different applications on image analysis.
Howevermore research should be conducted to increase the result ofthe accuracy for these applications.