Title:Deep Colorization
Background: The current colorization methods are divided into two categories: scribble-based methods and examplebased methods.
Scribble-based methods: Need to provide a lot of graffiti on the grayscale image to be developed, which is time-consuming and labor-intensive .
examplebased method: In order to reduce the burden of manpower, this method is proposed to convert color information from similar reference images to grayscale images. The color rendering quality depends on the choice of reference images. The choice of suitable reference images is a difficult problem.
This paper treats the image coloring problem as a regression problem and uses deep learning to solve this problem.
Innovation:
1. Propose the first image dyeing method based on deep learning and prove its effectiveness.
2. It carefully analyzes the rich but distinctive image feature descriptors from low to high, which is a key part of the algorithm .
General idea:
The test of neural network is divided into the following steps:
1. First, the input layer of the neural network contains 128 neurons, corresponding to the 49-dimensional patch feature (Low-level), the 32-dimensional DAISY feature (Mid-level
), and the 47-dimensional semantic feature (High-level). The output layer contains 2 neurons, which are divided into the value of the UV (chroma) component of the pixel.
2. After the chrominance component is obtained, in order to eliminate potential artifacts, the original gray image is used as a guide to use the joint bilateral filtering technique to refine the value of the chrominance component.
3. Finally, combine the refined chrominance components with the original grayscale image to obtain the final dyed image.
Focus: Understanding of image feature descriptors.
Disadvantages:
1. Due to the reliance on machine learning methods, a large-scale database training network containing all possible targets to be dyed is required, which is impossible. For example, the current image is trained through natural images, so it is not suitable for synthetic images.
2. The information lost during the color-to-gray conversion cannot be recovered.