Fusion of Empirical Wavelet Features for Object Recognition
Abstract
In this paper, an approach to recognize object efficiently is presented based on empirical wavelet features. In many computer vision applications, object recognition is required and it is a challenging task due to size and orientation of objects in the image. The proposed approach uses Empirical Wavelet Transform (EWT) to extract the characteristic of objects in an image. From the components of EWT, energy and entropy features are extracted. Then K-nearest neighbor classifier is used to recognize the object in the given image. The results show that the fusion of energy and entropy features provides better classification accuracy of 99.81% where the energy and entropy features provide 98.42% and 98.97% respectively on the benchmark object database named Columbia Object Image Library Dataset (COIL-100).
Keywords
Download Options
Introduction
Humans can easily recognize objects of varying size, shape, and orientations. However, it is a challenging task for computers. To ease object recognition, many automated approaches are developed recently. Some of them are addressed in this section. Learning strategy that models membership functions of the fuzzy attributes of surfaces is employed using GA [1] for object recognition. The objective function aims at enhancing recognition performance in terms of maximizing the degree of discrimination among classes. It is composed of three stages: retrieval and feature extraction of number of local parts from each model object, modeling the objects by feature vectors and similarity measurement.
A group-sensitive multiple kernel learning technique is used for object recognition to accommodate the inter-class correlation and intra-class diversity in [2]. A midway representation between the individual images and object category is obtained. An optimization model to concurrently perform kernel dictionary learning and prototype selection is discussed in [3]. The representation matrix is implemented to ensure that only a few samples are actually used to reconstruct the dictionary. So a convergent algorithm is employed to resolve the formulated non-convex optimization problem.
Context model based object recognition is discussed in [4]. It gives an efficient model that captures the information for more than a hundred object categories using a tree structure. It improves the performance of the system and also a coherent interpretation of a scene is obtained. Data driven un-falsified control is implemented for solving the drawbacks in visual servoing for object recognition in [5]. It recognizes an object through matching image features. Supervisory visual servoing is implemented until an accord between the model and the selected features is achieved, so that model recognition and object tracking are done successfully.
Conclusion
In this paper, an approach for the recognition 100 objects in the COIL-100 database is presented using EWT and KNN. As EWT gives a better approximation of images than DWT, it produces an excellent performance for object recognition. From the EWT decomposed images, energy and entropy features are extracted. They are fused together and given to the KNN classifier for classification. Experimental results show that the proposed fusion approach produces 99.81% accuracy. Also, it is clearly observed that the fusion of energy and entropy features gives the highest accuracy than their individual counterpart.