结合像素注意力与分类器解耦的小样本遥感图像目标检测

Combine pixel attention and decoupled classifier for few-shot object detection in remote sensing images

针对小样本遥感图像中实例标注缺失导致的前景类与背景类混淆的问题,提出了结合像素注意力与分类器解耦的小样本遥感图像目标检测方法。本方法设计一种新颖的像素注意力特征金字塔结构以捕获更重要的空间和通道语义信息,从而在抑制背景噪声的同时突出目标的关键特征。此外,将标准分类器解耦为两个并行检测头,分别处理前景类和含噪的背景类,以缓解分类器的偏置分类问题。提出的方法在两个公共遥感数据集上进行实验,结果表明,与目前最新的方法相比,所提方法在DIOR数据集上平均精度提升了4%~7%,在NWPU VHR-10数据集上平均精度提升了11%~17%,检测性能良好。

Aiming at the problem of confusion between foreground class and background class caused by the lack of example labeling in few-shot remote sensing images,a object detection method for few-shot remote sensing images combined with pixel attention and decoupled classifier is proposed.In this method,a novel pixel attention feature pyra-mid structure is designed to capture more important spatial and channel semantic information,so as to highlight the key features of the objects while suppressing background noise.In addition,the standard classifier is decoupled into two parallel detection heads to process the foreground classes and the noisy background classes respectively to alleviate the bias classification problem of the classifier.The proposed method is experimentally carried out on two public remote sensing datasets,and the results show that compared with the current new methods,the average accuracy of the pro-posed method is improved by 4%-7%on the DIOR dataset and 11%-17%on the NWPU VHR-10 dataset,and the detection performance is good.