A DCGAN image generation algorithm based on AE feature extraction

Current image generation algorithms have problems such as excessive noise in the generated images, poor image clarity and slow convergence of the generation network. To address the limitations of traditional image generation methods, we propose an image generation algorithm by combining an autoencoder (AE) and a deep convolutional generative adversarial network (DCGAN).The method first uses the feature representation capability of the AE to encode the training images, with the aim of compressing and downscaling the image to obtain an efficient data feature representation of the image. This has significantly less dimension than the original image and carries important features of the original image. This feature representation is then combined with random noise and used as the input to the DCGAN network after data processing. In the recognition phase, the network is continuously adjusted to optimize the model by using gradient descent and receives feedback from the discriminator. Experiments show that the network can generate images with better results to the extent that the human eye cannot separate them from the real data. The advantages of the proposed method are highlighted by comparing it with traditional generative networks.