Image color expression mode migration method based on deep convolutional neural networks

Abstract

The invention provides a deep convolutional neural network capable of completing an image style recognition task, and provides an image color expression mode migration method. According to the method,content representation features of a to-be-processed content image and an initialized image are extracted through the deep convolutional neural network pre-trained in an object recognition task, calculation of a content loss function is carried out, style representation features of the initialized image and a style image are extracted through the deep convolutional neural network pre-trained in an image style recognition task, calculation of a style loss function is carried out, and finally, a result of a total loss function is obtained; and a gradient descent algorithm is used to start fromthe initialized image to carry out iterative optimization in an image domain according to the total loss function to obtain a corresponding image of a smallest result of the total loss function to usethe same as a final result image. According to the method of the invention, image color expression mode migration can be completed, a case where massive strokes similar to impressionist painting images exist in an output result image can be avoided at the same time, and structure information of original natural images can be retained.

Description

technical field [0001] The patent of the present invention relates to deep learning, in particular to a deep convolutional neural network and a method for image color expression pattern migration based on a deep convolutional neural network. Background technique [0002] With the development of deep learning, deep convolutional neural networks have shown a strong ability in the task of separating and combining the content and style of reconstructed images, and image style transfer has made great progress. However, in the image style transfer task, there is no precise definition of image style. The style of an image can be any visual attribute of the image, such as texture, color, etc., so there is no uniform standard for evaluating the results of image style transfer. . [0003] In the current style transfer method, even with the help of the powerful image feature representation ability of the deep convolutional neural network, the content and style of the image still canno...

AI Technical Summary

Problems solved by technology

[0003] In the current style transfer method, even with the help of the powerful image feature representation ability of the deep convolutional neural network, the content and style of the image still cannot be completely separated, so the current stylized image has some structural information of the style image, For example, when implementing style transfer between natural images and impressionist painting images, the output result image has a large number of strokes similar to impressionist painting images, which destroys the structural information of the original natural image

The main reason is that the current mainstream style transfer methods all use the VGG-19 (or VGG-16) network trained on the ImageNet dataset to extract the content representation and style representation of the image, and the network is based on the ImageNet dataset. The object classification task of 1000 categories is trained, and the feature extraction of the deep neural network model is task-oriented, so even if the statistical measure of the feature representation extracted by VGG-19 (or VGG-16) trained in this classification task is used, it is still Structural information in small regions is preserved, resulting in structurally distorted resulting images

Images