Research On The Technology Of Face Recognition Data Augmentation For Unbalanced Datasets

In the era of deep learning,face recognition is one of the most successful cases,showing great advantages in many aspects.Face recognition tasks based on deep learning require a classification network with superior performance and large-scale face datasets.In recent years,due to the introduction of large-scale face datasets such as Celeb-1M,Deep Face,Mege Face,etc.,the difficulty of insufficient data has been gradually overcome,but the related problems surrounding face datasets are still full of challenges,one of which is the unbalance of the face datasets.The unbalance of face datasets is reflected in the huge difference in the number of images in different categories.The categories with a small number of samples may only be20-30,while the categories with a large number of samples are as high as 300-500.The difficulty of obtaining different types of face data results in large differences in the number of face images of each type in the face datasets.We take the unbalanced face dataset as the starting point,and through the explicit and implicit design of face data augmentation algorithms,the small sample types of face data are expanded,so that the overall face datasets are balanced,and the decision boundary deviation of the network model is reduced.We analyze the reasons for the unbalanced data,clarifies its influence on the convolutional network model,and proposes solutions.We investigate 14 basic image transformations and GAN-related image generation methods that act on unbalanced face datasets,and summarizes the characteristics and shortcomings of these methods,and designs a set that can extract low-level information and integrate high-level information generate a network,and integrate the Fourier transform module into the generated network,so that the network can make better use of global and local information.Through the algorithm designed in this paper,the small sample category data of the face can be expanded explicitly to obtain more balanced face data.The GAN training strategy designed in this paper is different from the basic GAN strategy.We use a dual discriminator network,one for content and one for clarity,to simplify the task of the discriminant model in a decoupling way,so that the generation network can generate face images well at both the content and clarity levels.From the algorithm level,this paper makes implicit data augmentation to the trained face data at the feature level,improves the metric space,extends the original Euclidean space to the cosine space,expands the feature layer data,and reduces the decision boundary deviation of the network model.