Residual-Transferred RNN Used for Object Detection in Optical Remote Sensing Images

Remote sensing technology provides high-precision, wide-coverage real-time data support for applications in traffic management and aviation safety. However, automated recognition based on remote sensing images encounters significant technical challenges in complex urban environments, including scene complexity, target diversity, and environmental interference, which hinder accurate feature extraction using conventional recognition methods. To address these challenges, this study proposes a Residual Transfer Recurrent Neural Network (RT-RNN) model based on deep transfer learning. The model integrates the spatial feature extraction capability of ResNet50 with the sequential information processing capacity of RNN to achieve efficient recognition of traffic objects and aircraft movements in remote sensing imagery. The residual modules in ResNet50 effectively mitigate gradient vanishing and explosion issues during deep network training through skip connections, significantly improving feature extraction efficiency Experimental results on the Dataset for DOTA demonstrate that the proposed model achieves 89% recognition accuracy, enabling rapid and efficient precise recognition of remote sensing images. This study provides an effective solution for target recognition tasks in complex urban environments.