Image Classification Based on Mid-Level Feature Fusion

Abstract

Image classification and analysis aim to classify the images according to the nature of visual contents. Due to recent increase in the number of multi-media contents, the classification of images is considered as a challenging and complex problem. A list of low-level and mid-level features is available in the literature that aims to represent the images in the form of feature vectors to be used as an input for classification-based problems. The main problem with these features is the domain and application specific nature and the use of features in one domain may not show the same result when applied in a different domain. The feature fusion-based approaches aim to enhance the performance of image classification models as single feature-based approach is not robust to handle image transformations such as translation, rotation, scaling, etc. In this research, we aim to investigate the image classification-based performance of mid-level features. We applied Bag of Visual Words (BoVW) model with image classification framework. This proposed late feature fusion result in a higher classification accuracy. The features are retrieved from the images by using two well-known feature extraction techniques that are Scale-Invariant Feature Transform (SIFT) and Histogram of Oriented Gradients (HOG). The experiments are performed while using two data sets that are Corel-1k and Corel-1.5k. The result shows that the classification accuracy increases when SIFT and HOG are used in fusion and the proposed results outperforms the standard BoVw model when SIFT and HOG is used separately.