Deep Neural Network Based Old Movie Restoration

Movie is a kind of electronic historical and cultural documents which can represent them vividly.The materials of old films have precious value in recording history and propagating culture.Old movies are stored in the form of films.Owing to the limitation of storage conditions and the long-playing consumption,there are various kinds of damage to old cine films.When they are reflected in the quality of the moving pictures,there are different forms of degradations,such as speckle noise,brightness flicker,scratches,color fading,and even destruction.Consequently,the protection and the salvageable restoration of the materials of old films are of no delay time.With the development of digital multimedia technology,it has become one of the significant ways to protect the image materials of old films that the content of old films is transcribed from film to digital format,and then digitally restored.Old film restoration has significant research value in the fields of computer vision,digital video processing and so on.In recent years,the tremendous development of deep learning provides a new train of thought for the old film restoration based on computer technology.Focusing on the digital old film restoration,this thesis studies on the detection of speckle noise,the restoration of speckle noise and the restoration of brightness flicker in old films to strengthen restoration quality of old films.The main contributions of this thesis are list as follow.(1)An adaptive spatial-temporal based speckle noise detection for old films.For the speckle noise detection in old films,the method takes advantage of the distribution characteristics of speckle noise in time series and space and simply detect and carefully select speckle noise from time series and space-time constraints respectively.First of all,the motion estimation is in progress to the image of the detected frame and the motion vectors of the corresponding positions of each pixel in the front and next frames of the current frame image are obtained;After that,based on the consequence of motion estimation,use the improved algorithms of speckle detection index and algorithms of enhanced hierarchical differential detector to detect all the speckles and adjusts the threshold through self-adaption and detect all the areas that are possible to become speckle;Finally,use space-time constraints to further filter the speckle area and obtain accurate the results of speckle position detection.The experimental results show that,compared with the current algorithms of noise detection,the algorithm that is proposed in this paper can achieve higher accuracy of speckle noise detection and effectively reduce the false detection rate.(2)A auto-encode-decode network based speckle noise restoration for old movies.For the issue of speckle noise restoration,draw lessons from the depth image priori,use the characteristics that self-coding and decoding network has low impedance to natural image coding and decoding,but high impedance to noise coding and decoding and actualize unsupervised speckle noise restoration algorithm based on deep self-coding and decoding network.The basic framework of the method is the self-coding and decoding network and it constructs a depth coding network to capture the prior information of image depth and inputs the detection template of noisy image and noisy location into the self-coding network.At the same time,the noisy image is used as the output target to realize the network training of unsupervised learning.Through the iteration of specific rounds,finally output the restored image without noise.The experimental results show that the proposed algorithm can achieve unsupervised speckle noise restoration,the effect of speckle noise restoration in old films is good,and the subjective visual quality is effectively improved.(3)An attention-aware deep recirculating network based flicker restoration algorithm for moving pictures.For the issue of brightness flicker in old films,use mechanism of attention to determine the input video sequence,contributes weight to the prediction of the result,and achieve brightness flicker restoration through deep cyclic structure network.The experimental results show that the method can effectively repair brightness flicker in old films,and the subjective and objective results are both better than the other comparison method.