Color transmission method based on visual saliency and channel attention mechanism

Abstract

The invention discloses a color transfer method based on visual saliency and a channel attention mechanism, and belongs to the technical field of digital image processing. According to the invention, based on visual saliency feature extraction of the reference image and based on the generative adversarial network model combined with the channel attention mechanism, a color transmission method for coloring the color features of the reference image is provided, so that the accuracy of color feature extraction and the final color transmission effect are improved; the method can obtain a more real and natural coloring result, and the result can learn a reference image color style which is more consistent with the cognition of a human visual system.

Description

technical field [0001] The invention belongs to the technical field of digital image processing, in particular to a color transfer method based on visual salience and channel attention mechanism. Background technique [0002] The research goal of the image coloring project is to add color to grayscale images or black-and-white line drawings, which can be applied to a variety of scenarios, such as coloring medical images, color reconstruction of old photos, coloring line drawings of comics, etc., and has high application value. [0003] Before the development of image coloring technology, the early pure manual coloring method was often done by artists manually drawing colors on the film, but this process usually takes a long time and requires a lot of manpower and material resources. Have a certain art background. In 1970, Wilson Markle coined the term colorization to describe the computer-aided technique he invented to add color to black-and-white movies, television shows, ...

AI Technical Summary

Problems solved by technology

[0006] Using the GAN to train the shading model can obtain high-quality shading results, but there are still some problems in the coloring results of the current mainstream methods: First, because most of the GANs still use convolutional structures, However, due to the limitation of the receptive field of the convolutional neural network itself, the coloring model lacks the capture of long-distance features, resulting in insufficient understanding of each position in the image, which affects the coloring of the local area; the second is the lack of detail in the generated image, which is Since the generative network will have a large number of redundant features in the process of downsampling feature extraction and the image discrimination process of the discriminative network, this will affect the feature learning ability of the model, resulting in insufficient detailed information of the colored image

Images