Multi-level local descriptor quantization for bag-of-visterms image representation