Canonical Correlation-Based Feature Fusion Approach for Scene Classification

Abstract

Vision-based scene recognition and analysis is an emerging field and actively conceded in computer vision and robotics area. Classifying the complex scenes in a real-time environment is a challenging task to solve. In this paper, an indoor and outdoor scene recognition approach by linear combination (fusion) of global descriptor (GIST) and Local Energy based Shape Histogram (LESH) descriptor with Canonical Correlation Analysis (CCA) is proposed. The experiments have been carried out using publicly available 15-dataset and the fused features are modeled by Random forest and K-Nearest Neighbor for classification. In the experimental results, K-NN exhibits the good performance in our proposed approach with an average accuracy rate of 81.62%, which outperforms the random forest classifier.