Multi-Source Ensemble Model for Scene Recognition

Scene recognition is a challenging task in computer vision due to diverse scene layouts and high inter-class similarities arising from shared objects across various categories. This complexity limits a single Deep Neural Network (DNN) from producing the highly discriminative features needed for effective performance. While many existing methods attempt to combine features from multiple pre-trained models, they often introduce several hyperparameters and require multiple training stages, complicating model development and increasing the risk of overfitting. In this paper, we present a novel ensemble model that integrates three distinct pre-trained models, specialized in object recognition, scene recognition, and semantic segmentation. This integration allows our ensemble model to leverage complementary information at the object, scene, and contextual levels, enhancing its ability to accurately interpret complex scenes. Additionally, during the testing phase, for every sample we assess the performance of each base model in our ensemble and utilize Hard Sample Review (HSR) to enhance the performance of models that exhibit weak results. HSR is an adaptive Test Time Augmentation (TTA) technique that reduces computational costs while delivering superior performance compared to traditional TTA. Our method achieves an accuracy of 90.15% on the MIT-67 dataset, highlighting its effectiveness over prior studies while maintaining simplicity.