Visual information retrieval in HEVC compressed domain

In this paper, a texture based retrieval method is proposed in HEVC compressed domain. The proposed method can be used for either video retrieval of HEVC coded videos or image retrieval of HEVC I-frame coded images. In I-frame coding various prediction modes are utilized to spatially predict pixels of a block. The selected prediction mode and the size of prediction unit for coding of a block indicate the way, in which the pixels of the block are related to their neighboring pixels and they can considered as texture features. Thus, we measured the texture similarity of coded I-frames in HEVC based on the similarity of the prediction modes and block size histograms. This retrieval method achieves 0.34 ANMRR in videos with high resolution, which is better compared to the other compressed domain retrieval method based on previous standard H.264/AVC that have achieved 0.45 ANMRR in the same image database.