Identification of Visual Correlations Between Non-Consecutive Shots in Digital Image Sequences

A number of automated methods for indexing audio-visual sequences have been developed. Typically, processing starts with a low level segmentation of a sequence of images so as to identify a series of shots (i.e. continuous camera records). To reach a higher level of description, patterns must be identified in the flow of consecutive shots. In this work, three different techniques for measuring visual correlations among non consecutive shots are proposed and compared. Two methods measure the visual correlation among shots by analysing the respective K-frames. In particular, they compare K-frames based either on a low resolution DC JPEG representation or on color and spatial organisation of the spatial information. The third technique measures the similarity between shots by comparing their associated codebooks, which are obtained using the Learning Vector Quantisation approach. Simulations have shown that the Learning Vector Quantisation approach leads to the best performance.