Research On The Relationship Between Face Recognition Accuracy And Image Properties

As one of the fastest growing biometric technologies,face recognition has been widely used.As an important step of face recognition,image acquisition is very important to the recognition accuracy.From the perspective of image,a variety of face recognition algorithms are compared and analyzed to explore the relationship between face recognition accuracy and various common properties of images.The face recognition model based on Inception-ResNet-V1 network structure for extracting features and SVM(Support Vector Machine)classifier is used in unrestricted environment,on the other hand the recognition algorithms like Eigenface,FisherFace and LBPH(Local Binary Patterns Histograms)are applied in restricted environment,which measures the recognition accuracy in a new dataset changed in the image-related properties such as hue and saturation.The purpose is to explore the adaptability of various face recognition methods to image changes.The correspondence between the property parameters and the recognition accuracy of each face recognition system can be applied to evaluate the influence of each property on the recognition accuracy of the specific recognition system,which could be used for selecting the training images in the future.The main work is as follows:1.More than 10 images of a single class image are selected from the LFW dataset to construct a dataset in unrestricted environment,and 15 images of a single class image are selected from the Caltech Faces face database and the Georgia Tech Face face database to create a dataset in restricted environment;2.The HSV color space,gray world method and other algorithms are used to process the constructed dataset,so that the image can be changed to different degrees in hue,saturation and other properties,finally the changed datasets are saved;3.Based on the idea of transfer learning,the frozen pre-training models trained in CASIA-WebFace face database,VGGFace2 face database with softmax loss training provided by FaceNet are combined with the SVM classifier to train the datasets under the previously constructed unrestricted environment,and our own face recognition systems are obtained,which performs well under the original dataset.Then the datasets changed in different properties are trained and tested in the systems.From the experimental results obtained,the recognition systems based on the two pre-training models work wellin different regions;4.Eigen-Face,Fisher-Face and LBPH are used to observe the performance of the three algorithms in the changed dataset under the restricted environment.From the recognition results,the three algorithms have unique correspondence with each image property parameter.