Construction of training data set of portrait cartoon stylized model and model generation

Abstract

The invention relates to the construction of a training data set of a portrait cartoon stylized model and the generation of the model. Through the data augmentation of a face color image and a small amount of cartoon face images, and through the combination of the detection of face key points and the adjustment of the sizes and facial forms of the five sense organs by a liquefaction algorithm, the five sense organs of all cartoon face images are deducted to form a material library, and the five sense organs are randomly combined and pasted on the face of a blank cartoon. Under the conditions of high cartoon data drawing cost and long consumed time, the cartoon face image data is amplified through the method, the data volume of a training set can be multiplied on the basis of a small amount of data, the identity diversity of cartoon face image data is greatly enriched, and a good portrait cartoon stylized model training effect is achieved.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to the production of portrait cartoon stylization, and in particular to the construction of a training data set and model generation of a portrait cartoon stylization model. Background technique [0002] With the development of smart mobile terminals and social networking, the cartoon drawing technology based on computer graphics processing is more and more popular among users, for example, it is possible to take or select an image image and input it into image processing application software (mobile terminal App or desktop level) software), and output cartoon-style image output through software processing. In the prior art, such application software is usually implemented using a cartoon model based on human face contours, and the cartoon model is obtained through pre-training with a training model using a large number of portrait photos and cartoon-style images as...

AI Technical Summary

Problems solved by technology

Portrait photos are relatively easy to take, but cartoon images need to be carefully drawn by experienced painters, enriching different face contours, facial features, etc., to form a variety of training data, and the economic and time costs are high, which limits the data scale. A small amount of cartoon portrait data for training, the training effect is relatively rough, it is difficult to obtain a rich cartoon style

[0004] In order to expand the data set, traditional methods such as random cropping, horizontal flipping, random rotation, and offset are used in existing methods to change the position information of the image, but this method cannot modify the identity information of the cartoon face, and the face key The point information is basically unchanged, such as the face shape and the distribution of facial features, which cannot be truly effective and improve the diversity of the data. In fact, it is still training corresponding to the key point information of the same face.

Images