Object segmentation in still images using topic modelling method

Abstract

One of the key components towards achieving high performance automated visual-based object recognition is the quasi-error free object segmentation process. Being an important integral part of many machine vision as well as computer vision systems, a tremendous amount of effort in object segmentation has been proposed in the literature. One of these approaches is the work that implements Probabilistic Graph Modelling (PGM) techniques. PGM is a rich framework for calculating probability and statistics in large given data sets and fields. One of the comprehensive methods in PGM is the Topic Modelling (TM) method introduced in the early 2000. TM has shown to be successful in classifying humongous information related to text and documents and has been implemented in many online search engines. Since image contains huge amount of information (in terms of pixels), segmentation of this information into meaningful region of interest (in this case objects) does fit into the framework of TM. The objectives of this project are to implement and analyze the capability and efficiency of TM in recognizing objects found in stationary images. TM is a process where it uses approximation technique to discover important segment or structure based on object classification. However, to proceed with object classification, object segmentation is firstly executed, making object segmentation as the most important part in the system. Through TM, the classification can be done by grouping the pixels (superpixels) accordingly in order to clearly represent the object of interest. In achieving this goals, Open Computer Vision (OpenCV) library will be fully utilized. It is expected that the proposed method will be able to perform object segmentation with high confident similar to state-of-the-art methods.