Optimizing JPEG Quantization Table for Low Bit Rate Mobile Visual Search

Ling-Yu Duan,Xiangkai Liu,Jie Chen,Tiejun Huang,Wen Gao

DOI: https://doi.org/10.1109/vcip.2012.6410738

2012-01-01

Abstract:Smart phones is bringing about emerging potentials in mobile visual search. Extensive research efforts have been made in compact visual descriptors. However, directly extracting visual descriptors on a mobile device is computationally intensive and time consuming. Towards low bit rate visual search, we propose to deeply compress query images by learning a customized JPEG quantization table in the context of visual search. Distinct from traditional image compression, by incorporating pair-wise image matching precision into distortion measure, we optimize quantization table to seek a better trade-off between image compression rate and visual search performance. An evolutionary algorithm is employed to learn an optimal quantization table. Under MPEG CDVS evaluation framework, extensive evaluation has been done including image retrieval and pair-wise matching over 1 million database images. Experimental results have demonstrated that our optimized quantization table works much better than JPEG default one in terms of retrieval/matching performance vs. a set of different operating points. The proposed low bit rate solution may be easily deployed to smart phones without hardware support, as a useful complement to the ongoing MPEG CDVS standardization efforts.