Gender Recognition Of Multispectral Face Image Based On Deep Learning

In recent years,the gender information in population attribute is widely used in population research,content index,advertising and other fields.And the research of using computer vision technology to identify gender attribute has gained great attention.However,at the present stage,the traditional recognition method based on visible light image is vulnerable to the limitation of illumination changes and harsh atmospheric conditions,and it cannot meet the requirements in many practical application scenarios,which is a technical difficulty to be solved urgently.In these environments,if the infrared band image is used,the required image details can be collected to overcome the environmental problems in complex situations.In this kind of scene,the performance of infrared technology will be compared with that of visible light recognition technology,which involves the performance of this method under multiple spectral images,that is,the problem of multi-spectral image recognition.The problem usually involves the same image in two or more spectral imaging.Different spectral images have different characteristics.For example,visible image pays more attention to texture features while infrared image pays more attention to detail features.This phenomenon usually results in the high performance of one method in one spectrum and the sharp decline in the other spectrum;Another common problem is the sharp performance degradation of a technical approach in the face of different databases under the same spectrum,that is,the problem of the robustness of the identification method.Aiming at the above problems in the field of gender recognition of multi-spectral face image and cross-database gender recognition,this paper summarizes and reviews some classical mainstream methods and makes some innovative work,including the following aspects:(1)In this paper,a face gender recognition method based on LBP(Local Binary Patterns),HOG(Histogram of Oriented Gradients),SIFT(scale-invariant feature transform)and CNN(convolutional neural networks)is studied.LBP,HOG and SIFT use their respective feature operators to extract features from face images,and gender recognition is carried out based on the feature vectors obtained and the classifier.In the mainstream convolutional neural network,feature extraction is carried out by convolution layer and pooling layer,and recognition and classification is carried out by fully connected layer.In order to verify the advantages and disadvantages of each method,we use the same spectral images in the same database to conduct gender recognition experiments.The results show that the convolutional neural network method has better recognition performance,higher accuracy and stronger robustness than the traditional manual feature extraction operator in the field of gender recognition of multi-spectral face images.(2)In the face of the limited improvement of traditional methods in the field of gender recognition of multi-spectral face images,this paper proposes to use deep learning technology to widen the network structure and deepen the number of layers on the basis of convolutional neural network.The advantages and disadvantages of Gender recognition experiments on multi-spectral face images by VGG,InceptionNet and ResNet were studied,a new method for Gender recognition of face images based on multi-scale convolutional neural network was proposed,and a new structure IR-Gender V1 was designed for Gender recognition of multi-spectral face images.The results show that the multi-scale convolutional neural network IR-Gender V1 converges faster and has higher recognition accuracy in the same database than other deep learning methods.(3)To solve the problem of cross-database robustness,the feasibility of transfer learning in deep learning method to ensure the gender recognition performance of multi-spectral face image is studied.Traditional recognition methods,which mostly rely on the feature extraction operator designed by hand,often have poor performance in the cross-database performance.On the basis of adopting multi-scale convolutional neural network,this paper conducts structure splicing and parameter migration for the deep network with excellent performance in a single database through transfer learning.The experiment shows that on the multi-spectral face data set TINDERS and CASIA,the IR-gender V2 network not only outperforms the traditional feature extraction operator in recognition performance,but also achieves an accuracy of 86.75%in cross-database gender recognition of VIS,which is higher than the 82.69%accuracy of deep learning network ResNet and 97.75%in gender recognition of CASIA database under NIR.Its comprehensive recognition performance is better than the previous single deep learning network structure,the recognition accuracy is higher,has a strong practical significance.