Tucker Bilinear Attention Network for Multi-scale Remote Sensing Object Detection

Object detection on very high resolution (VHR) remote sensing images plays an indispensable role in many applications. However, the large-scale target variation of remote sensing presents a significant challenge to high-precision VHR remote sensing target detection. Although existing methods attempt to enhance the feature pyramid structure and utilize various attention modules to improve the accuracy of high-resolution remote sensing object detection, small objects can still be overlooked due to the loss of critical detail features. There remain ample opportunities for improvement in multi-scale feature fusion and balance. To address this issue, this paper proposes two novel cross-layer modules: Guided Attention and Tucker Bilinear Attention. The former effectively preserves crucial detailed features, while the latter further enhances features by reasoning about semantic-level correlations. Based on these two modules, a new multi-scale remote sensing object detection framework is introduced. Compared to state-of-the-art methods on DOTA, DIOR, and NWPU VHR-10 datasets, the framework achieves comprehensive average detection accuracy for full-scale objects, and significantly enhances the detection accuracy for small objects. Code and models are available at https://github.com/Shinichict/GTNet.