PROJECT TITLE :

Robust Semantic Template Matching Using a Superpixel Region Binary Descriptor

ABSTRACT:

To compare the similarity between a template picture and a scene image, low-level image parameters like pixel intensity and pixel gradient are almost universally used in conventional template matching algorithms. There are many situations where these methods have been widely employed, however they cannot simultaneously handle all the issues of robustness. In this research, we provide a robust semantic template-matching (RSTM) approach in order to simultaneously handle the numerous problems. This paper proposes a new superpixel region binary descriptor (SRBD) to build a multilevel semantic fusion feature vector (RSTM) inspired by the local binary descriptor (LBD). Superpixels are extracted from the template image using a new kernel-distance-based simple linear iterative Clustering algorithm. For each superpixel's semantic characteristics, the dominant orientation difference vector, which is encoded as a rotation-invariant SRBD, may be derived from the average intensity difference between each superpixel region and its neighbours. Multi-level SRBD features and varying numbers of superpixels are combined into one feature vector in the offline matching phase. A marginal probability model is devised and used to determine the positions of template pictures in the scene image during the online matching phase. In addition, an image pyramid is used to speed up computations, as shown here. Experiments on a huge dataset from the MS COCO dataset are carried out to test this method's robustness. The findings of the experiments reveal that RSTM outperforms previous state-of-the-art template-matching algorithms by simultaneously addressing rotation changes, scale changes, noise, occlusions, blur, nonlinear lighting changes, and deformation.