A Novel Method For Extracting Object-of-Interest From Natural Image By Integrating Prior Knowledge

The detection and segmentation of Object-of-Interest (OOI extraction) are always among the key issue of several fields, such as computer vision, image understanding and pattern recognition etc. It has lots of important applications in industry, e.g. object recognition, content based image retrieval, content based image and video coding and compression, video surveillance, etc. The former methods of OOI extraction were often based on weak assumption that the OOI was in the central part of the image or was the most salient object in the image. Actually, this is not often in the case, so they were lack of application range. Inspired by how the human brain performs in this problem, we proposed a learning stage to obtain the prior knowledge about what we are interest indeed, thus leaded to the whole solution proposed here.This paper proposed a novel framework for extracting OOI from natural images. It mainly includes two stages:(1) Learning stage, which learns prior knowledge from positive samples of OOI, including shape and appearance prior knowledge. At this stage, we learned active shape sketch from several roughly aligned positive samples. Meanwhile, we computed the appearance feature histograms of OOI. Thus, both the shape and appearance prior knowledge were obtained.(2) Extracting stage. By integrating the learnt prior knowledge, we adopt the stochastic inference to do the segmentation and detection for extracting the OOI. At this stage, we firstly utilized the learnt active basis model to propose the initial detection candidate areas, then Markov Chain Monte Carlo based segmentation method was adopted. The best result is selected as the one with the highest posterior ratio.The two stages above are united together as a whole solution: the prior knowledge obtained in the learning stage can make the target pattern (Object-of-Interest) clearer, and can also guide the inference process in the extraction stage. Therefore, this method can obtain relatively accurate segmentation when extracting the OOI. The testing experiment in some public datasets has shown that our method outperforms the state-of-the-art.As a summary, the main contributions of this paper are addressed below: 1. A novel framework was proposed for solving the OOI extraction. Unlike the weak assumption based methods, we introduce a learning stage to obtain the prior knowledge of OOI, which makes the target pattern clearer and improves the accuracy of extraction. Besides, it matches the human's mechanism.2. A new method of constraining the boundary of the area with discrete shape sketch was proposed in this paper. This brings the possibility that we can measure the similarity between active basis template and a continuous contour.3. The method of embedding the prior knowledge to the prior model and likelihood model was original as we known, which improves both the computation efficiency and the accuracy of OOI extraction.