A Novel Collaborative Heterogeneous Supervision Network for Small Object Detection Method Based on Panchromatic and Hyperspectral Images

With the increasing availability of simultaneous panchromatic and hyperspectral images, object detection methods based on them have demonstrated significant application advantages. However, they still face several challenges that limit detection performance: 1) the sizes of small objects remain very small even in the fused images, insufficient texture information and spectral information that is easily confused, leading to lower accuracy in object detection; 2) etection-by-Preprocess (DBP) methods often suffer from spectral and spatial detail distortions, compromising target features; 3) Preprocess-free detection (PFD) methods extract panchromatic and hyperspectral features directly through networks, but the black-box nature of deep network makes it difficult to ensure precise alignment of these two types of features, thereby hindering further improvements in detection accuracy. Therefore, this paper proposed a novel Collaborative Heterogeneous Supervision Network (CHS-Net) for small object detection on panchromatic and hyperspectral images. First, integrating fusion and detection components into a heterogeneous supervision network enhances learning capabilities by incorporating more empirical knowledge. Second, a unified joint regulation strategy was introduced to enhance integrated learning capabilities using optimized feedback loss functions. This approach enhanced the attention of different components to target features, effectively improving weak small target detection performance. Finally, comparative experiments based on EO-1 dataset demonstrate that the proposed method outperforms many start-of-the-art approaches.