Research On Recipe Generation Technology Based On Generative Adversarial Network

With the changes of the times,people's living standards have gradually improved,and their eating habits have changed drastically.In the past,most people just wanted to eat enough and only a few people paid attention to the food pairing and tastes.Now,more and more people have higher requirements on recipes.China is a country with a large population and vast area,and the eating habits of different regions are not the same.In order to enrich the diversity of our country's recipes and meet people's growing dietary needs,we need to conduct research on the basis of existing recipes,and then accurately and scientifically explore the internal laws of food pairing and create new recipes.Nowadays,Generative Adversarial Networks(GAN)has attracted widespread attention and has achieved good results in many applications.The GAN model is still being gradually improved.On the basis of the original GAN model,many researchers have developed many different GAN variants,such as WGAN,LSGAN,BEGAN,Stack GAN,etc.These variants are suitable for solving practical problems encountered in various scenarios,which demonstrate the powerful capabilities of the GAN's family and provide new ideas for us to study the generation of recipes.In this article,we try to apply the Seq GAN framework to achieve the task of recipes generation.We first obtain Chinese recipes from the website through crawler technology,and then train the Seq GAN model so that it can generate realistic recipes.In order to expand the number of recipes data and improve the generalization ability of the model,we use text-based data enhancement technology and hope to obtain a better training effect.In order to verify the recipes generation effect of the Seq GAN model,we compared the recipes generated by the Seq GAN model and the RNN model.In addition,we are considering adding labels to recipes to meet people's personal needs.The labels indicate the taste characteristics of the recipes,such as "sauce flavor","sour and sweet and flavor " and "spicy flavor".At this time,a training Condition Generative Adversarial Network(CGAN)is required.In CGAN,we use the taste labels of the recipes data as the condition information to guide the recipes generation.By analyzing the relationship between recipe tastes and ingredients,we can find the internal connection between tastes and ingredients(mainly minor ingredients).The experimental results show that the two generative models trained in this article can realize recipes generation by learning the collocation relationship between foods,which provides a scientific means for people to choose ingredients in daily life.