Loading...

Multi-Class Object Locating and Recognition

Mostajabi, Mohammad Reza | 2012

416 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 43810 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Gholampour, Iman
  7. Abstract:
  8. Environment Identification and recognizing surrounding objects is an exigent need in future applications. For example one of the emerging technologies in car industry is driverless cars. In driverless cars, navigation system should be able to detect and recognize pedestrians, traffic signs, roads, surrounding cars and so on. Therefore, conventional single-object recognition systems are not capable of handling the needs of advanced machine vision based applications. In recent years, designing and analyzing multi-class object detection and recognition systems have become a big challenge in machine vision. In this thesis our goal is to identify and analyze the existing problems in designing multi-class object detection and recognition systems. For designing a multi-class object detection and recognition system, three main stages should be considered: 1- Creating descriptive models for images 2- Feature extraction 3- Classification. Most of the recent scientific works focused on the third stage, while the other stages are playing key roles in multi-class object detection and recognition systems. Therefore, in this thesis, in spite of considering different classifiers and their characteristics, our focus is on the two first stages, namely: creating descriptive model of images and feature extraction. In creating descriptive model of image, we proposed a framework based on affine invariant regions to improve the segmentation accuracy of unsupervised image segmentation methods. The main advantage of the proposed framework is its independency to the exploited unsupervised image segmentation method. As a matter of fact, the proposed framework is able to improve the segmentation accuracy of any unsupervised image segmentation method. In the feature extraction stage we proposed a scale-invariant texture feature set, named ‘Directional Differences (DDs)’, for multi-class object detection and recognition applications. Investigating the flaws of conventional texture features in discriminating different classes of objects, led us to Directional Differences which are much more robust to the scale changes of objects in different images. Our experimental results show that new feature presents about 8% higher accuracy, comparing to Gradient based features in a multi-class image segmentation evaluation. In this thesis, two methods have been proposed for semantic information extraction: different levels of feature extraction and semantic information extraction based on image context. Experimental results highlight the importance of these two methods in multi-class object detection and recognition system. In comparison with other systems, the proposed system achieved the best segmentation accuracy with the following specifications: processing each image of size 320×213 in 0.6 second on a personal computer with 2.9 GHz Intel Core-i5 CPU and gaining 79% accuracy in image segmentation
  9. Keywords:
  10. Feature Extraction ; Classification ; Unsupervised Image Segmentation ; Object Recognition

 Digital Object List

 Bookmark

No TOC