Impact of Advertial Data Augmentation on Robust Indoor Scene Recognition: An Evaluation Study

Indoor scene recognition is an emerging technology with significant potential for smart homes, robotics, and virtual/augmented reality applications. However, the robustness of indoor scene recognition algorithms against adversarial attacks is a significant concern for their practical deployment. This research project aims to investigate the impact of adversarial data augmentation on the robustness of an indoor scene recognition algorithm. We develop an indoor scene recognition model based on image captioning and comprehensively analyse its accuracy in classifying indoor scenes. Subsequently, we employ data augmentation techniques, specifically image superimposition on both the test and training datasets. By superimposing five images featuring diverse objects to our dataset, such as a Christmas tree, airplane, monkey, train, and palm tree, we aim to evaluate the model’s reaction to noisy input and assess its ability to generalize in the presence of unexpected objects within indoor scenes. Moreover, by training a model on superimposed and standard images, we aim to evaluate whether the new model has enhanced regularisation, transfer learning capabilities, and noise tolerance.