Semi Automatic Digital Motion Picture Restoration System with Learning Capabilities

We have developed a semi automatic film restoration system, DIMORF (DIgital MOtion Picture Restoration System for Film Archives), working with a minimum effort of parameter settings and training, with adaptive learning abilities, at relatively high speed. The system is being applied for the restoration of old films, characterized by strong color distortions, scratches, dirt, vibration, and flickering. Since we handle large amounts of visual information, video indexing and retrieval components are also very important. The system we present herein contains novel features very important for state-of-the-art application. Some of these are an automatic self-teaching stabilization algorithm, a neural network based learning blotch detection method, a novel automatic shot detection and indexing framework. Data regarding the restoration process, shot data, filtering data, etc. are all stored in an XML-based portable and versatile database backend, providing support for the learning algorithms, for browsing and retrieval purposes and also for logging. The topics we will concentrate in this paper are the automatisms and the learning possibilities, which we use during the restoration process in filters and in video indexing.