Small-Object Detection Model for Optical Remote Sensing Images Based on Tri-Decoupling++Head

The accurate detection of small object in optical remote sensing images (ORSIs) presents a significant challenge due to the background interference and the small size of the objects, making them susceptible to misdetection and omission. In this article, we construct a high-precision detection model, Tri-Decoupling++Net, to solve the problem of misdetection and omission of small objects in ORSIs. Specifically, we adopt Mosaic9 data augmentation in data preprocessing to indirectly increase the batch size and optimize the model's ability to learn small-object features in ORSIs. On this basis, we design the Tri-Decoupling++Head to optimize the image's three tasks of category prediction, predicted box localization, and confidence calculation independently and in parallel, perceiving the position and structure of the small objects more comprehensively. Moreover, to further improve the feature extraction capability of the model, we construct a layer-by-layer dense residual connection module and embed it into the neck to obtain deeper semantic information, helping the model to better distinguish between small objects and backgrounds in ORSIs. Finally, we design the adaptive Focal-EIoU loss function that can adaptively adjust the weight of centroid distance loss as well as width and height losses to better handle the small-object localization and scale estimation problems, further improving the accuracy and robustness of small-object detection in ORSIs. Extensive experiments on two publicly available datasets demonstrate that Tri-Decoupling++Net outperforms state-of-the-art ones.