Research and Application on Face Recognition Algorithm Based on FWLD Method and Deep Belief Nets
Yihong Zhang, Xiang Li, Zhijie Wang · 2017
Aiming at the problem of insufficient feature extraction and sensitive to noise for traditional face recognition algorithms, a face recognition algorithm based on improved Weber local descriptor and deep belief nets is proposed.First analyze the shortcomings of Weber local descriptor, and based on the fuzzy logic, an improved FWLD face description method is proposed.To make full use of the domain pixels in WLD direction component, based on introduction of LBP coding, the fuzzy logic is used to optimize the LBP coding, and then the improved WLD algorithm is used as the input of the deep belief network.And the network parameters are obtained through the layer-by-layer greedy pre-training network.Finally, the BP neural network is used to fine tune and optimize the DBN network, and the test samples are predicted by the trained network.In the ORL data set, the correct rate is 95%.The simulation results show that the face recognition algorithm proposed in this paper is higher in recognition rate and more robust than the traditional recognition. IntroductionFace recognition technology, as an important part of pattern recognition and computer vision, is concerned by more and more researchers.At the beginning of the study, the researchers proposed some methods, including the geometric feature based method, the model-based method and the statistics-based method, which is better than other methods, such as Principal Component Analysis (PCA) [1], and Linear Discriminant Analysis (LDA) [2], etc.Under the limiting conditions, there are always better recognition effects, but under the non-limiting conditions, such as changes in light, attitude changes, and occlusion, the recognition rate drops sharply.In response to this problem, people are increasingly studying local features.In the literature [3], local Binary Pattern (LBP) is proposed.The LBP is featured by simple calculation, invariant rotation and constant gray scale, but the description ability is insufficient and is susceptible to noise interference.In literature [4], (Weber Local Descriptor, WLD) is proposed, unlike the LBP method, the WLD feature consists of the differential excitation component and the direction vector.The WLD differential excitation component counts the difference between the neighborhood pixel and the central pixel, and then divides the center pixel, which further improves the WLD robustness.In literature [5], WLD is proved to be insensitive to light and convenient to calculate.In literature [8], WLD is used for face recognition, and good experimental results are achieved.In literature [9], a multi-scale WLD algorithm is proposed to further increase the robustness of WLD.In literature [10], a rotation-invariant WLD algorithm is proposed to enhance its rotation invariance by introducing angle calculation on the direction component.In literature [11], Gabor and WLD combined algorithm is proposed, and successfully applied in the iris recognition.Different from the current improved method, the algorithm in this paper introduces the fuzzy logic, and fuzzy LBP coding is conducted on the WLD direction component.This not only makes full use of the domain pixels on the direction component and improves the description ability of WLD, but also improves the robustness of WLD by fuzzy logic.Experiments show that the algorithm proposed in this paper has stronger description ability and robustness than traditional WLD.