Indexing Audio-Visual Sequences by Joint Audio and Video Processing

The focus of this work is oriented to the creation of a content-based hierarchical organisation of audio-visual data (a description scheme) and to the creation of meta-data (descriptors) to associate with audio and/or visual signals. The generation of efficient indices to access audio-visual databases is strictly connected to the generation of content descriptors and to the hierarchical representation of audio-visual material. Once a hierarchy can be extracted from the data analysis, a nested indexing structure can be created to access relevant information at a specific level of detail. Accordingly, a query can be made very specific in relationship to the level of detail that is required by the user. In order to construct the hierarchy, we describe how to extract information content from audio-visual sequences so as to have different hierarchical indicators (or descriptors), which can be associated to each media (audio, video). At this stage, video and audio signals can be separated into temporally consistent elements. At the lowest level, information is organised in frames (groups of pixels for visual information, groups of consecutive samples for audio information). At a higher level, low-level consistent temporal entities are identified: in case of digital image sequences, these consist of shots (or continuous camera records) which can be obtained by detecting cuts or special effects such as dissolves, fade in and fade out; in case of audio information, these represent consistent audio segments belonging to one specific audio type (such as speech, music, silence, ...). One more level up, patterns of video shots or audio segments can be recognised so as to reflect more meaningful structures such as dialogues, actions, ... At the highest level, information is organised so as to establish correlation beyond the temporal organisation of information, allowing to reflect classes of visual or audio types: we call these classes idioms. The paper ends with a description of possible solutions to allow a cross-modal analysis of audio and video information, which may validate or invalidate the proposed hierarchy, and in some cases enable more sophisticated levels of representation of information content.