Confidence-driven feature selection and ensemble classification using ResNet, AlexNet, and DenseNet for scene recognition

Abstract

Indoor scene recognition is a significant problem because of its structural complexity, illumination variations, and visual similarities between the various scene classes. Recent approaches use convolutional neural networks (CNNs) and transfer learning to extract features for classification. In these methods, all the features were considered for classification, which leads to an increase in computational complexity and training time. This leads to more time consumption for processing. To address these challenges, this study proposed a hybrid indoor scene recognition framework that integrates illumination-aware pre-processing, multi-model deep feature extraction, and confidence-based feature fusion. The pre-processing technique uses contrast-limited adaptive histogram equalization (CLAHE) and Retinex enhancement. Deep features are extracted using Alex Krizhevsky network (AlexNet), densely connected convolutional network (DenseNet), and residual network 101 layers (ResNet101) to capture spatial and semantic information. The features extracted are fed to confidence-driven world cup optimization (CD-WCO) to extract the significant features. The final representation is classified using a deep liquid state machine (DLSM) classifier to enhance class discrimination. Our experiments, conducted on the NYU indoor scene dataset with 20 scene categories and the MIT Indoor-15 class subset results indicate the effectiveness of the proposed framework, achieving an overall classification accuracy of 96.5% and 95.2%, respectively.