Enhancing Near-Duplicate Image Detection Using Spatial Transformer Networks in Deep Learning Models

In recent years, detecting near copy images has been believed to be the most important issue in many application fields, for example, digital media management, surveillance, and image retrieval. Traditional methods do not work well compared between near duplicate images which experience transformations from rotation, scaling and cropping. To improve the performance of near duplicate image detection, we introduce this research by introducing an innovative method that we integrate spatial transformer networks (STN) into the deep learning framework (e.g. convolutional neural network (CNNs) and vision transformer (ViTs)) to detect near duplicate images. Learning of spatial transformations in STNs allows a dynamic capture of spatial transformations for image alignment and normalization, leading to better features learned. On large scale datasets experiment results show a substantial improvement on detection accuracy compared with existing methods using the proposed STN based models. We demonstrate that this method yields a scalable and robust solution for the high dimensional transformation of large numbers of near duplicate images, resulting in a powerful tool in the fields of image retrieval and similarity analysis.