Research On Food Image Generation Based On Generative Adversarial Networks

In the current daily life,artificial intelligence has gradually entered the human vision,and its efficient work ability has brought great convenience to people.With the development of science and advancement of technology,people’s demand for image data in their lives has begun to increase day by day.In most cases,due to the lack of raw data,people’s personalized needs for images are difficult to meet.However,with the success of the Generative Adversarial Networks(GAN)in the fields of optical image reconstruction and image generation,the resolution and fidelity of its generated image have gradually improved.It is possible to study the generation technology of personalized images based on the adversarial generation network.And has important value and significance.This paper is based on the generation of dish map images and relies on the principle of controllable generation of images by Conditional Generative Adversarial Networks(CGAN).This paper studies the generation of dish images by three CGAN algorithms.In GAN,its discriminator model can be viewed as a simple binary neural network.Therefore,we propose a method that combines the model structure of a multi-classification neural network with the model structur e of a traditional discriminator model,thereby indirectly improving the ability of the model to generate images based on conditions by improving the capability of discriminator models.The main work of this article is as follows:First,This paper analyzes the advantages and disadvantages of the three CGAN models,and try to avoid problems such as gradient exploding and mode collapse during model training.Secondly,This paper conducted multiple sets of comparative experiments on the three improved CGANs using the MNIST dataset and the CIFAR-10 dataset,and made optimal adjustments to the network structure and model parameters of the three conditional GANs.Finally,according to the results of the control experiments,several suitable models are selected,and the CF80 dataset collected and sorted by our laboratory is used for training.The results are compared with the original CGAN model and the conclusion is drawn.The experimental results show that the improvement of the discriminative model of CGAN improves the data recognition ability of the discriminative model,and effectively controls the gradient exploding problem and mode collapse,thereby improving the ability of generating model data.