Latent Semantic Analysis (LSA) based object recognition and clustering

Abstract

Object Recognition and clustering are major techniques in Pattern Recognition, Computer Vision, Artificial Intelligence and Robotics. Conventionally these techniques are implemented in Visual-Feature based methods and Cosine Similarity method or Vector Space method which uses semantic similarity among the objects to solve these kinds of problem, but this method has two problems synonymy and polysemy. In synonymy many ways to refer to the same objects for e.g. car and automobile, which leads to poor recall. Similarly, in polysemy most words have more than one meaning, e.g. model, python and chip, which leads to poor precision. In this paper, we propose a method which will overcome the problems like synonymy and polysemy. If the text printed on the object the semantic feature of that object is extracted and clustered according to Latent Semantic Analysis (LSA). The Proposed method is based on semantic information so an experiment is conducted with the dataset of images which contains the packing cases of commercial products (e.g. Mobile, Laptop etc). Semantic information in the dataset is retrieved using text extraction modules and then the results of text extraction are passed through an Internet search module. Finally, objects are recognized and clustered using the LSA module. The clustering results are more accurate than cosine similarity or vector space method.