정확한 얼굴/사람 분석을 위한 국부 특징 변환과 융합

Abstract

Face/Human analysis is one of the important topics in computer vision. Many researchers have introduced many different approaches using local transform features: specifically local binary patterns (LBP) and histograms of oriented gradients (HOG). Each approach has its own advantage in that LBP is robust to local illumination change and HOG is robust to local pose change. However, the proposed methods have some limitations such that LBP is sensitive to locally changing intensity changes due to makeup, wearing of glasses, and a variety of background and HOG requires a long processing time to compute the feature transformation.To overcome these limitations, we propose two new local transform features: local gradient patterns (LGP) and binary HOG (BHOG). LGP assigns one if the neighboring gradient of a given pixel is greater than the average of eight neighboring gradients, and zero otherwise, which makes the local intensity variations along the edge components robust. We show that LGP has a higher discriminant power than LBP in both the difference between face histogram and non-face histogram and the detection error based on face/face distance and face/non-face distance. BHOG assigns one if the histogram bin has a higher value than a given threshold, and zero otherwise, where threshold is just an average value of the total histogram bins. This makes the feature computation time fast due to no further postprocessing and SVM classification.We also propose a hybridization of local transform features that combines them by AdaBoost feature selection method, where the best local transform feature among several local transform features (LBP, LGP, and BHOG) is sequentially selected by the degree of its classification error until we obtain the required classification performance. This hybridization makes the face/human analysis robust to the local illumination change by LBP, the locally changing intensity change by LGP, and the local pose change by BHOG, which improves the detection/recognition performance considerably.We apply the proposed local transform features and the hybridization technique to the face detection, human detection and face recognition to validate the improvement of detection/recognition performance as follows. In the face detection, experimental results using the MIT+CMU and the FDDB databases show that the face detection rates using LBP, LGP, and LBP+LGP+BHOG hybrid features are 90\%, 93\%, and 96\% at one false positives per image (FPPI). LGP feature achieves better face detection performance than LBP feature and hybrid feature achieves the best face detection performance among other features. In the human detection, experimental results using the INRIA database show that the human detection rates using HOG, BHOG, and LBP+LGP+BHOG hybrid features are 79\%, 80\%, and 86\% at one FPPI. BHOG feature achieves the similar human detection performance with HOG feature, but BHOG feature is 10 times faster than HOG feature, and hybrid feature achieves the best human detection performance among other features. In the face recognition, experimental results using the PF07 database show that the face recognition rates using HOG with 30 non-overlapping blocks, HOG with 40 selected blocks, LBP(2)+LGP(4)+HOG(34), and LBP(12)+LGP(15)+HOG(40) hybrid features are 72.2\%, 77.3\%, 79.3\%, and 84.3\%, respectively. HOG feature with selected blocks achieves better face recognition performance than HOG feature with non-overlapping blocks and hybrid feature achieves the best face recognition performance among other features.In summary, the proposed LGP and BHOG feature show the accurate detection and recognition performance and a fast computation time, respectively, and the hybrid feature provides a considerable improvement of face detection, human detection, and face recognition.