Facial Atrribute Recognition Based On Deep Matric Learning

Facial images contain a wealth of valuable information such as gender,age,etc.,which plays an extremely important role in the accurate identification of human identity and human-computer interaction.The attribute recognition based on facial image is a cross-domain issue involving pattern recognition and computer vision.Based on the results of face detection,the thesis uses Convolution Neural Network to extract features of face area,and then carries out gender identification and age estimation.The work done by this thesis is as follows:1.With the low performance of age recognition for facial images,the thesis pro-poses an algorithm for facial age estimation based on deep triplet loss.Specifically,a deep embedding network with powerful distinguishing ability is constructed to ex-tract high-efficiency features from facial images,which joins the classification loss and triplet loss functions to optimize and supervise the training.With triple loss,more dis-criminative features can be extracted by the constraint that the difference between two facial images of the same age is smaller than that of different ages.The experimental results show that the mean absolute value of facial age can be reduced by 28.8%,which exceeds the figure for other current facial age recognition methods.2.With the difference between negative samples as well,the thesis proposes an age estimation algorithm based on deep quadruple ordering loss.The age estimate method based on deep triples loss only considers the relative distance between positive and negative sample pairs.In order to introduce the absolute distance between positive and negative sample pairs,a negative minus sample is added to form a deep quadruple.The experimental results show that the age average absolute error declines by 6.8%compared with the former algorithm based on deep triplet loss.3.With the low recognition accuracy and poor generalization ability for facial attributes through multiple single networks,the thesis proposes the method for facial attributes recognition based on multi-task metric learning.Facial attribute recognition includes two tasks:facial age and gender,which joins the quadruple loss function and the classification loss function to optimize the training of the deep embedding network aiming at a deep feature metric space.The accuracy of facial age and gender and their generalization ability are jointly optimized by sharing the characteristics of the previous layers,and using the relationship between attributes.The experimental results show that the age average absolute error sees a drop by 13.7%and the gender recognition accuracy can be improved by 1.4%.4.We built a video big data analysis platform for retail business,which integrates the facial attribute recognition algorithm based on multi-task metric learning proposed in the thesis,and provides an in-store heat map analysis function.The customer facial data collected in the store was used for testing.The test results show that the platform can meet the needs of the actual project in terms of time and accuracy.