A FUSION OF HAND-CRAFTED FEATURES AND DEEP NEURAL NETWORK FOR INDOOR SCENE CLASSIFICATION

Abstract

Convolutional neural networks (CNN) have proved to be the best choice left for image classification tasks. However, hand-crafted features cannot be ignored as these are the basic to conventional image processing. Hand-crafted features provide a priori information that often acts as the contemporary solution to CNN in image classification, and hence an attempt is made to fuse the two. This paper gives a feature fusion approach to combine CNN and hand-crafted features. The proposed methodology uses two stages, where the first stage comprises feature encoder that encodes non-normalized features of CNN, which utilizes edge, texture, and local features. The fusion of handcrafted features with CNN features is carried out in the second Hand-crafted crafted features are validated that helped CNN to perform better. Experimental results reveal that the proposed methodology improves over the original Efficient-Net(E) on the MIT-67 dataset and achieved an average accuracy of 93.87%. The results are compared with state-of-the-art methods.