ABSTRACT This paper addresses the problem of recognizing semantic content from images and video for content based retrieval purposes. Semantic features are derived from a collection of low-level features based on color, texture and shape combined together to form composite feature vectors. Both Manhattan distance and Neural Networks are used as classifiers for recognition purposes. Discrimination is done using five semantic classes viz. mountains, forests, flowers, highways and buildings. The composite feature is represented by a 26-element vector comprising of 18 color components, 2 texture components and 6 shape components. General Terms Semantic content recognition Keywords Color, Texture, Shape. 1. INTRODUCTION Automated schemes for content based image retrieval (CBIR) are usually based on low-level features like color, texture and shape. Recognition of such low-level content has been extensively studied over the last couple of decades. However humans are more interested in image retrieval based on semantic or high-level content i.e. images or video frames need to be retrieved which are semantically similar to some given image, rather than visually. This is a more difficult proposition firstly because automated systems are not capable to directly recognizing semantic concepts, and secondly because semantic concepts are human abstractions and they have no fixed definitions e.g. a house or a car can have all types of colors, textures and shapes. For semantic content recognition therefore systems are first designed to retrieve a set of low-level features and then at a higher level a mapping is done to associate a set of low-level features with one or more high-level concept. This paper addresses the problem of recognizing high level concepts from images and video by utilizing a number of low-level features. The paper is organized as follows: section 2 provides an overview of related work, section 3 outlines the proposed methodology, section 4 provides details of the dataset and experimental results obtained, and section 5 provides the overall conclusions and the scope for future research.
Read full abstract