DISO-DETR: A Detail-Enhanced Infrared Small Object Detection Transformer Based on UAV Imagery

When it comes to detecting small objects in infrared images using unmanned aerial vehicles (UAVs), existing models really struggle with a few big obstacles. The tiny size of the targets, the cluttered backgrounds, and the need to run efficiently on edge devices all present major challenges. To address these issues, this paper proposes a novel detail-enhanced infrared small object detection model, DISO-DETR, based on the RT-DETR architecture. It introduces two key innovations: the Multiscale Detail Feature Backbone Network and the Wavelet-based Frequency Feather Aggregation module. These two modules effectively boost the model to capitalize on the infrared image details while maintaining a lightweight design. This makes it a great fit for real-time use on edge devices. Experimental validation shows that the mAP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5</inf> of DISO-DETR is raised by 6%, and the number of parameters and Gflops are reduced by 21.6% and 22.6%, respectively, over the base model on the HIT-UAV dataset. Moreover, when compared to other advanced models, DISO-DETR not only improves infrared small target detection performance but also demonstrates superior efficiency and robustness, proving its advantages in real-world applications.