Restoration And Coloring Of Historical Photos Of Lanzhou University Based On Deep Learning

Lanzhou University was founded in 1909.Over the past hundred years,a large number of historical photos have been stored to witness the growth and changes of Lanzhou University.Through the photos,we can feel the unique campus spirit and cultural elements of each era,and also see the historical mission undertaken by the students of Lanzhou University in different eras.However,limited to the photography technology and storage means at that time,most of the existing historical photos of Lanzhou University are black and white,and there are blurred,scratches,light spots and other damage.In order to better inherit the historical memory and cultural value in the photos,this paper realizes the restoration and coloring of the historical photos of Lanzhou University by means of deep learning.The specific work contents are as follows:Firstly,the restoration of historical photos of Lanzhou University was realized.In the restoration process of historical photos,the models trained only by synthetic photos cannot achieve better generalization effect in real historical photos.In this paper,Variational Auto Encoder(VAE)is adopted to design conversion models among real historical photos,synthetic historical photos and clean photos.According to the complex degradation process of historical photographs and the content of most people-oriented features,the mechanism supporting multiple degradation restoration and the restoration details of face detection and facial enhancement are set in the model,so as to obtain a historical photo restoration model with strong generalization ability.The historical photos of Lanzhou University to be restored are from the alumni website of Lanzhou University and the News website of Lanzhou University,with a total of 196 photos.The quality of 168 photos was significantly improved after restoration by the restoration algorithm,among which the restoration of black-and-white photos was mainly reflected in the elimination of scratches and spots,the resolution of yellowing and fading problems,and the improvement of overall photo clarity;the restoration of color photos was mainly reflected in the removal of scratches and face enhancement.Second,the coloring of black-and-white historical photos is completed.In this paper,the coloring of historical photos is done based on the generative model of Generative Adversarial Networks(GAN),in which the generator is improved from the U-net network and the discriminator is a simple binary discriminator.In this paper,128 black-and-white historical photos of Lanzhou University are collected,and 56%of them achieve the expected coloring effect after being colored by the coloring algorithm.In order to improve the coloring effect,the method of code fusion is used to combine the repair function and coloring function,and the effect of different order of restoration and coloring on the coloring effect is investigated.After experimental comparison,it was found that the best coloring sequence was repair first and then coloring,which not only improved the coloring effect,but also increased the ratio of reaching the expected coloring effect to 73%.