PAMA-DETR: a lightweight attention and multi-kernel feature fusion detection model for UAV images
摘要
Due to small object sizes, dense distributions, and complex backgrounds in UAV imagery, existing object detection algorithms face significant challenges. This study introduces PAMA-DETR, an advanced object detection algorithm based on RT-DETR, designed for high precision and robustness. The designed efficient and lightweight Partial Attention Gate Network (PAGNet) improves extraction capabilities for small targets in complicated scenarios. PAGNet adopts an innovative single-head self-attention mechanism combined with convolutional gated linear unit (CGLU) to augment the capacity to distinguish small targets from the background. Recognizing the difficulty of fusing multi-scale features within dense scenes, we designed the multi-kernel adaptive feature fusion (MAFI) module. MAFI employs a multi-branch parallel architecture integrating diverse convolutional variants to establish multi-scale receptive fields, accommodating objects of various sizes. Additionally, we optimize the loss function by replacing GIoU with Inner-SIoU and further improve localization accuracy by constructing auxiliary bounding boxes. Experimental on VisDrone2019 and UAVDT datasets show that PAMA-DETR achieves superior performance compared to RT-DETR, with improvements of