Research On The Quality Evaluation Method Of GAN Image Generation Based On Deep Learning

The Generative Adversarial Network(GAN)uses adversarial learning to fit the distribution of real samples and obtain the ability to generate data,which is of great significance to the development of artificial intelligence.In recent years,there have been more and more innovative researches on GAN,and various derivative models are constantly improved and developed.They have been widely used in the image field,and many breakthrough results have been achieved.The essence of the GAN is to learn the real sample distribution during the game of two neural networks,however,the images generated by GAN have no absolute correspondence with the real images and GAN lacks a clear likelihood measurement.So,traditional methods are difficult to quantify the generated image quality.At present,there are relatively few researches on the evaluation of GAN generated images,and the more popular indicators still have problems such as the uncertainty of the low GAN score and relying on pre-trained models,which hinder the standardized development of GAN to a certain extent.Therefore,the research on this basis has a very important application prospect.Based on the basic theory and practical application of GAN,this paper studies the principle,structure and development of GAN and its derivative model,and proposes a new GAN quality evaluation model,and designs indicators to evaluate the stability of GAN.(1)In order to quantitatively evaluate the authenticity and diversity of the images generated by the GAN,the currently popular evaluation indicators of the GAN are analyzed,and the evaluation indicators with reference value and development potential are discussed,and then a new quality evaluation model,Squeeze-and-Excitation GAN Quality Model(SEGQM),is proposed.The model uses SE-Res Net to classify images and designs two scores to describe the authenticity and diversity of the generated images.On this basis,the comprehensive evaluation score of the generated adversarial network is determined,and an evaluation index triplet Squeeze-and-Excitation GAN Quality Index(SEGQI)is designed.Experimental results show that compared with Inception Score,Fré chet Inception Distance and GQI,SEGQI can objectively and comprehensively evaluate the authenticity and diversity of generated images,and has very important guiding significance for the design and development of GAN.(2)In order to better detect the stability of the GAN model,the reasons for the instability of GAN and the problems existing in the training process are discussed.The common pattern collapse phenomenon is analyzed in detail,and a stability evaluation index based on Siamese Network is proposed from the perspective of image similarity.The similarity measurement model uses a feature extraction network to extract image features and uses the idea of Siamese Network to fit the average feature distance of images to obtain the similarity between images.A GAN Stability Index(GSI)is designed by calculating the ratio of the in-class and inter-class similarity between the generated data set and the real data set.The experimental results show that the stability evaluation index GSI can detect whether the GAN model has a mode collapse,which reflects the stability of the model.