A robust multilevel segment description for multi-class object recognition

Abstract

We present an attempt to improve the performance of multi-class image segmentation systems based on a multilevel description of segments. The multi-class image segmentation system used in this paper marks the segments in an image, describes the segments via multilevel feature vectors and passes the vectors to a multi-class object classifier. The focus of this paper is on the segment description section. We first propose a robust, scale-invariant texture feature set, named directional differences (DDs). This feature is designed by investigating the flaws of conventional texture features. The advantages of DDs are justified both analytically and experimentally. We have conducted several experiments on the performance of our multi-class image segmentation system to compare DDs with some well-known texture features. Experimental results show that DDs present about 8 % higher classification accuracy. Feature reduction experiments also show that in a combined feature space, DDs remain in the list of most effective features even for small feature vector sizes. To describe a segment fully, we introduce a multilevel strategy called different levels of feature extraction (DLFE) that enables the system to include the semantic relations and contextual information in the features. This information is very effective especially for highly occluded objects. DLFE concatenates the features related to different views of every segment. Experimental results that show more than 4 % improvement in multi-class image segmentation accuracy is achieved. Using the semantic information in the classifier section adds another 2 % improvement to the accuracy of the system.