Image Generation and Recognition Technology Based on Attention Residual GAN

Abstract

In accordance with the concept of game antagonism, Generative Adversarial Network (GAN) is a popular model in current image generation technology. However, GAN has problems such as unstable training and difficult convergence, which seriously affect the effectiveness of input feature extraction and image recognition. The study introduces residual network structure and self attention mechanism to calculate the weight parameters of features, and then guides image generation through image label information. The improved GAN model classifier is applied to image recognition. The final experimental data shows that the Fréchet Inception Distance (FID) values of the iGAN in facial expressions and behavioral actions are 77.68 and 176.84, respectively, which are closer to the distribution of real image data. In behavioral image recognition, the accuracy of the model is 96.8%, and the required time is 30 seconds. In facial expression recognition, the accuracy and recognition time of the model are 90.1% and 24 seconds, respectively. This indicates that it can generate high-quality images, has stronger feature extraction capabilities, and has higher recognition efficiency. This model provides a new technical reference for the further improvement of image processing technology, and has certain application potential and value.

Keywords

• Attention mechanism
• residual network
• GAN
• image generation
• identification technology