A Compressed-domain Robust Descriptor for Near Duplicate Video Copy Detection

This paper introduces a global descriptor from the compressed video domain (H.264) for near duplicate video copy detection tasks. The proposed descriptor uses a spatial-temporal feature structure in an ordinal pattern distribution format. The proposed descriptor is constructed from Intra-Prediction Modes (IPM) of key frames (IDR & I slices) and extracted from the compressed video files, using the MPEG4/AVC (H.264) codec. Intra-prediction is the compression technique used in the key frames of the H.264 codec. As the proposed feature describes pictures globally, this research compares the feature with the two other well-known global image descriptors, ordinal intensity/colour Histograms and ordinal Auto-correlograms, as baselines. Our experiments show how the proposed feature outperforms the baseline features in non-geometric transformations T3, T4 and T5 in effectiveness as well as efficiency. It is due to better representation of the image content and smaller feature vector size. The core competency of the proposed feature is in non-linear brightness and contrast changes (Gamma expansion and compression) in which the intensity/colour Histograms and Auto-correlograms are deficient.