Abstract: To address the problem of low accuracy of DETR(DEtection Transformer) in small object detection, an improved DETR for small object detection was proposed. Firstly, an improved MetaFormer combined with a multi-scale attention mechanism was adopted as the backbone network, aiming to solve the problems of weak extraction ability, low efficiency, and detail loss in small object feature extraction of ResNet-50, thereby enhancing the representation capability for small objects. Secondly, a deformable attention decoder was introduced to address the problems of slow convergence and limited feature space resolution in the Transformer attention module when processing image feature maps. This enabled the model to focus on key sampling regions around reference points, accelerating convergence and improving detection accuracy for small objects. Finally, the Wise-IoU (WIoU) v3 loss function was incorporated to overcome the limitation of the GIoU loss function in evaluating prediction box quality. Differentiated gradient gains were assigned to prediction boxes of varying quality, guiding the model to converge towards higher accuracy. Experimental results on the COCO2017 object detection dataset showed that, compared with DETR, the proposed method improved the average precision for small objects by 7.6 percentage points and the overall average precision by 4.7 percentage points, demonstrating superior detection performance.

Key words: DEtection Transformer (DETR), small object, deformable attention, multi-scale attention, WIoU (Wise-IoU) v3

摘要： 针对DETR（DEtection Transformer）在小目标检测方面精度较低的问题，提出了一种基于改进DETR算法的小目标检测方法。首先，针对骨干网络ResNet-50在小目标特征提取方面提取能力弱、效率低、易丢失细节等问题，提出了一种结合多尺度注意力机制的改进MetaFormer作为DETR的骨干网络，增强模型对小目标的表征能力。其次，针对Transformer注意力模块在处理图像特征映射时存在收敛慢、特征空间分辨率受限等问题，引入了可变形注意力解码器，使模型能够聚焦于参考点周围的关键采样区域，从而加快模型收敛并提升小目标检测精度。最后，针对GIoU损失函数无法衡量预测框质量的问题，引入了WIoU(Wise-IoU) v3损失函数，为不同质量的预测框赋予差异化的梯度增益，引导模型收敛到更高的精度。在COCO2017目标检测数据集上的实验结果表明，相较于DETR，所提方法对小目标的平均检测精度提升了7.6个百分点，整体的平均检测精度提升了4.7个百分点，表明所提方法具有更高的检测精度。

关键词: DETR, 小目标, 可变形注意力, 多尺度注意力, WIoU v3