LAST-Net: Local Adaptivity Spatial Transformer Network for Multiobject Detection in UAV Remote Sensing Thermal Infrared Imagery

Object detection in unmanned aerial vehicle (UAV) infrared remote sensing is a critical technology in the fields of intelligent transportation and computer vision. However, this technology faces numerous challenges due to its inherent characteristics, such as complex background interference that hinders semantic scene understanding, low-contrast images that lead to the loss and blurring of feature information, and disparities in object scale and data distribution that result in mismatches in the feature map. Existing studies have focused on improving individual issues using convolutional neural networks (CNNs) or transformer architectures, but they have not achieved satisfactory results in UAV infrared scenarios. This paper proposes the local adaptivity spatial transformer network (LAST-Net) to address these challenges. (1) To address feature extraction in complex remote sensing scenes, the convolutional spatial transformer block (CST-Block) is proposed, which effectively integrates convolutional operations and transformer architecture through parallel computation to extract both local and global features, while suppressing background interference. (2) To address feature fusion for the low-contrast issue in infrared images, the dual-modality adaptive attention (DMAA) mechanism is proposed, which leverages the interaction and feedback between content and spatial matrices to enable coarse-to-fine queries, thereby effectively aggregating object information. (3) To handle object scale variation and sample imbalance, the coordinate crossing loss function (CCLF) is proposed, which dynamically adjusts weight factors to match different scales and multi-scene objects. Experiments on the DroneVehicle and HIT-UAV datasets show that LAST-Net achieves mean average precisions of 86.3% and 90.9%, surpassing state-of-the-art algorithms.