目的 遥感目标检测技术在遥感测绘、智慧城市、乡村振兴、国防军事等领域具有广泛应用，但由于遥感图像存在背景信息复杂、待检目标多且小等特点，导致目标特征随着网络加深淹没在背景信息中，不利于后续检测任务。方法 本文构建了一种局部无参注意力YOLO网络。首先提出一种局部无参注意力，能够根据当前特征提高局部区域内目标关注度，而不引入训练参数，以此构建无参注意力残差块，嵌入到骨干网络的不同阶段中，实现多尺度目标特征增强和背景信息抑制。在此基础上，利用最优传输距离度量边界框的相似性，构造了名为W-CIoU的联合度量方法和相应损失函数，以缓解锚框与真实框差异过大导致的标签误分配问题，降低小目标漏检率。结果 在RSOD和RSSOD数据集上的实验结果表明，该网络在保证模型参数量和复杂度基本不变的前提下，其平均精度均值(mAP)分别达到98.2%以及87.4%。结论 本文所提出的算法能够抑制背景信息并增强目标特征，相较对比算法能更好满足复杂场景下小目标检测需要。

Objective The technology of remote sensing object detection has found extensive use in many fields, including remote sensing mapping, smart cities, rural revitalization, national security and military affairs. However, due to the complexity of background information and the many small objects to be detected in remote sensing images, the object features are submerged in the background information with the deepening of the network, which is not conducive to subsequent detection task. Method constructs a local parameter free attention YOLO network. First, we propose a local parameter free attention, which can improve the object attention in a local region based on current features, without any trainable parameters, and constructs the parameter free attention residual block based on it, which is embedded in the different stages of the backbone network to achieve multi-scale object feature enhancement and background information suppression. On this basis, this paper uses Wasserstein distance to measure the similarity of bounding boxes, and constructs a combined measurement method named W-CIoU and relevant loss function, which alleviates the problem of label misallocation caused by significant differences between anchor boxes and GT boxes, and reduces the missed detection rate of small objects. Result The experimental results on RSOD and RSSOD datasets show that the mean accuracy(mAP) of the network is 98.2% and 87.4% respectively, while the model parameters and complexity are basically unchanged. Conclusion The algorithm proposed in this paper can suppress background information and enhance the object features. Compared with the comparative algorithms, it can better meet the needs of small object detection in complex scenes.