Research On Optimization And Application Of Computer Vision-oriented Generative Countermeasure Network

Generative Adversarial Networks has achieved great success in synthesizing real images and modeling.It is widely used in domain adaptation,image visual computing,language processing,etc.In supervised machine learning or deep learning,too few labeled samples will cause the built model to be prone to overfitting,and the representation ability and generalization are not strong.In practice,it is very difficult and unrealistic to obtain a large number of labeled samples.Generative confrontation networks are relatively mature in image generation,but the existing generation confrontation networks have shortcomings such as instability,mode collapse,and low computational efficiency.We conduct research on the generative confrontation network,and it is of great significance to use the network to generate sample examples and enhance the data set.This paper aims at the problem that it is difficult to obtain a large number of training samples due to the difficulty of data collection and the high cost of sample labeling,and explores the optimization of generative confrontation network and its application in image generation and small sample tumor recognition tasks.The main research contents are as follows:(1)An Attentive Evolutionary Generative Adversarial Network based on the attention mechanism is proposed to improve the shortcomings of traditional generative adversarial networks that are unstable,low in computational efficiency and easy to collapse.In AEGAN,the generator continuously evolves through three independent mutation operators to resist the current environment-the discriminator,and only the offspring with high fitness(ie the generator)is retained in each iteration.In addition,a normalized self-attention module is embedded in the discriminator and generator of AEGAN to adaptively assign weights according to the importance of features.Furthermore,a learning and training algorithm for AEGAN is proposed.Through this algorithm,AEGAN overcomes the shortcomings of traditional generative adversarial networks that only use a single loss function and rely on deep convolution,greatly improving the stability and statistical efficiency of training.Extensive image synthesis experiments were conducted on CIFAR-10,Celeb A and LSUN datasets to verify the performance of AEGAN.Experimental results and comparisons with other generative adversarial networks and their variants show that this model is better than existing models.(2)Taking the practical application of tumor image diagnosis as the background,aiming at the problem of unsupervised domain adaptation for small samples,a general model of domain-adaptive tumor recognition with enhanced data set is designed.First,a larger source domain data set is created by using standard data augmentation techniques and AEGAN,and the enhanced data set is used as the source domain to train the label classifier.Secondly,a self-adversarial domain based on the attention mechanism is designed.Adapt to narrow the gap between the source domain and the target domain,and use the label classifier to test the target tumor data set to improve the accuracy of small sample tumor recognition.Simulation experiments show that the domain-adaptive method based on data set enhancement can effectively improve the accuracy of malignant tumor image recognition.