Research On Data Augmentatiom Technology Based On Generative Adversarial Network

In recent years,with the continuous advancement of computer technology and the popularity of deep learning technology,the application of artificial intelligence in various fields of society has been more widely promoted and landed.Among them,the face detection algorithm is the basic step of many applications such as authentication,security monitoring,etc.It is used for further face matching by characterizing,framing and labeling different instances of faces in the image.However,in specific scenes,the diversity of faces in the image data is very rich,including factors such as size,light,angle,which have a serious impact on the face detection rates.In the current public dataset,only WIDERFACE's face data is the most detailed,but it also can not cover all kinds of real-life scenarios.Therefore,automatic enhancement of the data set is a general operation of the training model.There are two ways of data augmentation including traditional method and GAN.In the traditional enhancement methods,the number of images is mainly augmented by segmenting,inverting,and shifting the images,but these ordinary linear or radiometric transformations cannot change the details of face representation,and they are powerless to the above problems.On the contrary,the advanced data augmentation method based on GAN has shown unparalleled attraction at present.GAN is a generative adversarial network.It learns the simulated distribution of images through game training and generates false-true images and also has broad prospects in the field of image migration and image editing.Therefore,this thesis implements an enhanced network of face detection training data based on GAN,which has the following two aspects:First,I generate datasets through multiple different deep neural networks.The implementation of the variational auto-encoder(VAE),learning the mean and variance of multiple distributions ensures to generate the corresponding relationship of the faces,resulting in a natural but fuzzy set of faces.Implementation of DCGANQP model using square potential as a loss function with encoder section,trained using Celeb A and LSUN datasets to generate realistic sets of faces and scenes.Implementing a Cycle GAN-based day and night image migration model to enhance the WIDERFACE data set.Implementing an Att GAN-based face editing model to generate multiple attributes for a partial collection of faces.Second,through the use of the generated data,the multi-image synthesis method is used to generate the data set and annotation information for the face detection model of Retina Face and Center Face,and the converted image and the existing WIDERFACE annotation are used to form the training dataset in the case of harsh conditions.Then using the pre-trained models of the above two face detectors for the training of additional data set generation to obtain a better performing detection model.Therefore,this paper argues that using GAN for automatic data augmentation and adding detector training in the face detection tasks can effectively enhance the robustness of the model and improve the detection rate of the model.