Research On Skin Injury Data Augmentation Technology Based On Improved Generative Adversarial Network

Image classification and object recognition hold significant importance in the field of deep learning medical imaging.However,in the task of automatic skin lesion classification,the limited availability of annotated medical images presents challenges in training models with robustness and generalization capabilities.To address the issues of insufficient image data and neural network overfitting,this study combines traditional data augmentation methods with generative adversarial network techniques in deep learning to expand the skin lesion image dataset.The main content of this article includes the following aspects:(1)Construction of medical image datasets.In this paper,skin cancer was selected as the research object and based on the public dataset of the medical image dataset ISIC 2018 Skin Classification Challenge.Preprocessing operations are performed on various skin injury images,and appropriate center cropping is performed to complete the production of the dataset.During the training process,the number of parameters of the network can be reduced more effectively,thereby reducing the amount of computation of training and inference and improving the training speed.(2)Choosing representative models for data augmentation,this paper conducts a comparative study on traditional augmentation,Deep Convolutional Generative Adversarial Networks(DCGAN),and Wasserstein GAN(WGAN)models.We analyze the advantages and disadvantages of each network model,and propose an improved version of the DCGAN model called ECA-GAN,which is tailored to our specific application scenario.By adding the channel focus(ECA-Net)module to the DCGAN network model generator,the channel information of the input feature map is better utilized,and the clarity and realism of the generated image are improved.In addition,residual blocks are added to the DCGAN generator and discriminator and combined with the deconvolution blocks in the original model to effectively alleviate the checkerboard effect generated in the composite image and improve the image generation quality.(3)The nature of the data allows us to compare directly with the expanded image.The proposed method is finally classified in the residual network(Res Net),and the experimental results show that the accuracy of the ECA-GAN network model compared with the original algorithm reaches 99.5%,which is about 1% and 15.4% higher than before and without expansion.In summary,this paper takes skin lesion image augmentation as the research task,and the generative adversarial network model designed effectively solves the problem of insufficient data,and effectively improves the accuracy of the model in the final medical classification.