Image generation method

Abstract

The invention discloses an image generation method, which comprises the following steps: acquiring a training data set which comprises a plurality of first images and a plurality of second images, with the first images being original images, and the second images being images of a category to be generated; establishing a neural network model based on the CycleGAN and VAE; and training the neural network model through the training data set, wherein the trained neural network model is used for image generation. According to the method, a mode of combining the CycleGAN and the VAE is adopted, code distribution of the image is generated through the VAE network, the generated codes are input into the CycleGAN network to serve as noise signals of the generated image, category limitation is added to the generated image, it can be guaranteed that the generated image is an expected image, and the generated image is vivid and has good detail information at the same time.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image generation method. Background technique [0002] Vision is an important method for class recognition and detection of objects. With the development of artificial intelligence technology, visual inspection technology has developed rapidly, and image recognition and detection technology based on computer vision has also developed rapidly. At present, the machine learning method represented by deep learning has become the mainstream method of image recognition and detection. These methods first need to use a large number of images to classify or recognize and detect objects to learn, extract the characteristics of the category or object, and then use the features to complete the classification. Therefore, a large number of image samples is the basis for effective training of machine learning methods, and also the basis for the realization of machine learning meth...

AI Technical Summary

Problems solved by technology

The advantage of CycleGAN is that the image quality is high and the details are clear. The main problem is that the training is difficult and may not converge. In addition, the resulting image cannot guarantee the distribution characteristics of the sample.

[0007] Variational Auto-Encoder (VAE, Variational Auto-Encoder) is another method that can realize image generation, but since there is no confrontation process to judge the authenticity of the generated image, the VAE model only calculates the mean square error between the original image and the generated image The training of the model can be realized, causing the generated image to be more blurred than GAN

Images