Generative Adversarial Network Based Face Attribute Recognition And Regeneration

In the field of pattern recognition and multimedia search,deep learning convolutional neural network is a new technology in recent years.With its ad-vantages of simplicity,efficiency and easy training,it has been widely used in image processing field.Especially in the face-related fields,the emergence of convolutional neural networks has greatly improved the accuracy of face recog-nition and face recognition,which has become the mainstream technology and the most promising technical direction in the current face area.As an alternative to the evolution of neural networks,the early days of confrontation generation networks were to explore the internal construction of neural networks.With the continuous evolution of related technologies,with its characteristics that can generate realistic images,it also shows strong practical value in the field of image reconstruction.At the same time,with the arrival of performance bottle-neck in supervised learning,the concept of Migration Learning,as a transition between unsupervised learning and supervised learning,proposes that data and methods of existing scenes can be used to explore recognition tasks in unknown scenarios.Out of a higher research significance.In this paper,both engineering practice and theoretical research take into account,first introduced the convolution neural network infrastructure,the tra-ditional theory of training and testing methods;then introduced how to use the project to accelerate the training of distributed multi-card neural network train-ing;actual Optimization techniques used in network feedforward,such as con-volution optimization,merging multiple computation steps,and so on.Finally,with the help of deep learning framework and instruction set technology,train-ing speed and 10 times feedforward speed increase can be greatly improved.In the aspect of research,this paper mainly studies two aspects of the ba-sic recognition of face attributes and the migration of face attributes.There are two basic problems in face recognition,such as the difficulty of multi-dataset tasks and the credibility of network output.By adjusting the network struc-ture,improving the way of image pre-processing and designing self-assessment modules,Relevant issues have been targeted to solve,but also enhance the ac-curacy of face recognition of property images.The absolute error in the morph age dataset was only 3.5 years old,more than 90%in the chalearn fotw gender and smile dataset accuracy,and the top 5 accuracy of the 5 age-related datasets annotated by the laboratory reached 93.6%.On the other hand,we combine the countermeasure generation network to explore the application of relocation learning to face attributes.Firstly,we gen-erate the neural network that generates the real face by generating the face im-ages and optimizing the authenticity and universality of the synthesized data.In order to apply the facial attribute model to different usage scenarios,the 40 types of face attributes that can perform well on both celeA and lfwA datasets are constructed with the help of the theory of face-super-resolution combined with migration learning Model training methods.Compared to the original model increased by 10 percentage points.