Image matching based on a structured deep coupled metric learning framework

The general similarity metric method is difficult to calculate the similarity of images under different resolutions, occlusion, and deformation. Therefore, this paper proposes a structured deep coupled metric learning framework (SDCML), which provides a solution for image matching under the above special interference. For the input image pairs, the structured region segmentation is performed to derive global and local sub-region images. After that, a stacked sparse auto-encoder network is used to implement deep feature extraction on these derived images. Then, the features with different dimensions of derived image pairs are transformed to coupled space through multi-supervised coupled mapping, and the similarity metric of feature pairs is completed in this space. Finally, all calculated similarities of derived image pairs are fused to obtain the final matching result. The effectiveness of the proposed method is verified by image matching experiments under non-ideal conditions; the optimal recognition rate was 91.98% on Yale-B dataset and 72.80% on Market1501 dataset, respectively.