Abstract: To address the problems of difficult feature extraction and insufficient detection accuracy when detecting small targets in infrared images obtained using unmanned aerial vehicles (UAVs) owing to small target sizes, low signal-to-noise ratios, and complex background interference, this study investigated an improved RT-DETR algorithm, denoted as MPH-DETR. First, a multiscale adaptive feature extraction module was designed, which simultaneously enhanced the ability to perceive weak targets and reduced the model parameters by means of cross-layer feature fusion and an adaptive attention mechanism. Second, in the neck network, a Pola-SEFN structure was proposed to improve the detection accuracy of the model, which decoupled the target and background signals and optimized the feature transmission. Finally, in the feature-fusion stage, a hyper-MFM structure was designed to address the problems of cross-scale semantic fragmentation and feature loss. Experimental results showed that compared with the RT-DETR baseline, the accuracy, mAP@50, and mAP@50:95 metrics were improved by 8.4%, 5.8%, and 4.8%, respectively, when using the MPH-DETR algorithm on the HIT-UAV infrared dataset, which effectively alleviated the problems of false detection and missed detection in complex backgrounds.