Recognition of partially occluded objects in 2D images

Abstract

Object recognition is spread across too many fields such as industrial, image retrieval and medical models. A human being can identify the objects with high performance and professionalism; on the other hand, a machine facing difficultly and effort to identify the object. In order to facilitate the process of identifying and analyzing objects easily using the machine, researchers worked hard to create new technologies and develop technologies that already exist for this purpose. The boundaries of computer vision are especially challenged by partial occluded object recognition. The aim of our research is to propose an algorithm using to recognize the partially occluded objects in two different cases: an object missing some part and objects are overlapping each other. The dataset that used in this research is silhouette images; we chose 60 images to represent the occluded object which missing part of the object. These images divided into three categories according to the percentages of the occlusion, and overlapping objects contains 36 images (each scene contains two objects). We collected those images from the MPEG-7 dataset downloaded it from the Internet. Adaptive Window is the purpose technique for extracting the features. Dynamic Time Warping (DTW) works for matching between objects. Orientation field is used to calculate the angle of a window's fragment. Algorithm goes through multiple stages, starting with the pre-processing through extract features from images and ends by comparing the images that enable us to obtain results of the matching, performance and efficiency of this algorithm. The experiments results demonstrate that the proposed algorithm is robust to identify missing and overlapping objects and it can work with strength occlusion.