Research On Restoration Method Of Old Movies Based On Deep Learning

Many filmstrips of early movies were not properly stored and protected.Repeated screenings caused scratches,flashes,frame dropping and other damage.What’s more,under the erosion of the years,problems of mildew,adhesion,deformation,fouling,oxidation,cracking and fading may arise.It has become a pressing issue to mitigate the damage to films and to restore thes e valuable video materials.In addition to restoration of problems such as blotches,scratches,and flicker,high-definition restoration of heritage films is needed to make them blossom in value in today’s era and bring it closer to the public to meet user s’ requirements for picture quality.The purpose of this study is to apply a deep learning approach to black-and-white old movie restoration.In order to train the networks,a novel data simulation process is used to automate the generation of artificially simulated old movie data.Meanwhile,an old film stylization process based on generative adversarial networks is proposed,in which the network learns to simulate old movies autonomously in order to compensate for the limitations of explicit modeling.For the restoration algorithm of old movies,an end-to-end convolutional neural network architecture is first proposed,which uses 3D convolution al layers to extract global features of multiple frames and controls the magnitude of information in adjacent frames by computing the Temporal Attention.Compared with two "deblotch + SR" combinations,the proposed method has significant advantages in several metrics and exceeds the comparison method in terms of visual clarity,naturalness,and satisfaction of human perception.Ablation experiments on the smoothing module demonstrate its role in complementing and equalizing global information.In addition,for the problem of temporal discontinuities such as flicker,this study proposes a meandering recurrent neural network architecture that can increase the temporal capacity of the network while achieving bidirectional information transfer.Meanwhile,to make full use of the temporal information,an unsupervised mask detection mechanism is used to assist in blotch removal.A multi-frame conditional discriminator is applied to supervise multiple temporal states of the generator network simultaneously.The experimental results show that the proposed meandering network is significantly useful in flicker removal and outperf orms the comparison method in spatial restoration.The effectiveness of mask detection is demonstrated by observing the mask features.