In this paper, we introduce an efficient method to substantially increase the recognition performance of object recognition by employing feature selection method using bag-of-visual-word representation. The proposed method generates visual vocabulary from a large set of images using visual vocabulary tree. Images are represented by a vector of weighted word frequencies. We have introduced on-line feature selection method, which for a given query image selects the relevant features from a large weighted word vector. The learned database image vectors are also reduced using the selected features. This will improve the classification accuracy and also reduce the overall computational complexity by dimensionality reduction of the classification problem. In addition, it will help us in discarding the irrelevant features, which if selected will deteriorate the classification results. We have demonstrated the efficiency our method on the Caltech dataset.