Machine Learning-Based Low-Complexity Image Feature Descriptor for MPEG-CDVS Standard

Compact descriptors for visual search (CDVS) is a completed standard from the ISO/IEC moving pictures experts group (MPEG). CDVS has a low complexity and bitrate efficiency on image matching and retrieval. The MPEG CDVS framework detects key feature points in the image and extracts descriptors to obtain bitstreams to match images. We proposed a low-complexity image descriptor for visual searching within the MPEG CDVS framework. The proposed algorithm integrated traditional image processing methods with a machince learning (ML) network, resulting in reduced compression time and the size of the CDVS standard bitstream. To evaluate the effectiveness of the proposed algorithm, we tested the proposed algorithm with the CDVS standard software to ensure its compatibility. On the CDVS-Benchmark, the experimental results demonstrated that the proposed algorithm reduced matching time by 31.7%, bitstream size by 31.8%, and extracting time by 12.8% compared to the CDVS standard test model. Similarly, on the HPatches-Benchmark, the proposed algorithm reduced matching time by 5.3%, bitstream size by 25.2%, and extracting time by 32.0% compared to the CDVS standard test model.