Research And Its Application On Grayscale Image Colorization Algorithm Based On Dense Neural Network

In most scenes,color images have richer information than grayscale images.The existing gray image coloring algorithm mainly includes: user-guided color spreading algorithms? color mapping algorithms based on a specified function or parameter?data-driven pseudo coloring algorithms.Due to their own characteristics,the first two algorithms can not adapt varible kinds of images very well or be used on a large scale.In the third type of algorithm,the algorithm based on the example image reference method had poor adaptability on the large-scale usage.With the improvement of the performance of machine learning and image computing devices(GPUs,etc.),the grayscale image colorization algorithm based on deep learning has excellent performance and stands out.This type of method uses neural networks to build different network architectures and train with large data sets.By taking convolution operation,the content and features of the image are extracted and analyzed,and the mapping relationship between the gray image and the color image is sought.Then the corresponding colorization model is constructed.This paper adopts the idea of deep learning and studies the gray image colorization algorithm.The specific content is as follows:(1)This paper presents a method of grayscale image pseudo coloring that constructed and trained an end-to-end deep learning model based on dense neural network aims to extract all kinds of information and features(such as classification information and detail feature information).Just enter a grayscale picture to the trained network,you can generate a full and vibrant vivid color picture.By constantly training the entire network on a wide variety of data sets,you will get the most adaptable,high-performance pseudo color network.The experiments show that the method proposed in this paper has a higher utilization of features and can obtain a satisfactory coloring effect.Compared with the current advanced pseudo color methods,it has also made remarkable improvements,and to a certain extent,the problem during the coloring processing have been improved,such as color overflow,loss of details,low contrast etc.(2)By introducing multi-scale training,multi-loss optimization,and through comparison experiments,the performance of existing algorithms is optimized.(3)By using transfer learning,the color human anatomical slice image was taken as the reference image,and the medical image colorization experiment was carried out,and the effect was acceptable,which provided a new way for medical image coloring.