基于改进RFBNet算法的遥感图像目标检测

针对遥感图像中的小目标存在信息少、易受背景干扰、特征表达较弱等缺陷, 导致目前通用目标检测算法在对这类小目标进行检测时效果不理想的问题, 为提高对遥感图像中小目标的检测能力, 提出一种基于RFBNet的改进算法. 该算法以RFBNet为框架, 首先利用自校正卷积取代特征提取网络中的常规卷积, 以扩展感受野丰富输出, 进而强化对弱特征的提取能力; 然后设计多尺度特征融合模块, 丰富浅层特征图的抽象信息； 最后设计稠密预测模块, 提前在较浅层整合上下文信息, 使最后阶段的每层输出都含有丰富且联系紧密的多尺度特征信息. 将该算法在数据集UCAS_AOD和NWPU VHR-10上进行实验, 平均检测精度分别达83.4%和94.8%. 实验结果表明, 该算法有效提高了遥感图像中目标检测的精度, 且针对遥感图像中的小尺度目标检测问题改善明显.

Target Detection for Remote Sensing Image Based on Improved RFBNet Algorithm

Aiming at the problem that small targets in remote sensing images had some defects, such as less information, easy to be interfered by the background, weak feature expression and so on, which led to the current general target detection algorithm was not ideal in the detection of such small targets. In order to improve the detection ability of small targets in remote sensing images, we proposed an improved algorithm based on RFBNet. The algorithm was based on the framework of RFBNet. Firstly, the self-correcting convolution was used to replace the conventional convolution in the feature extraction network to expand the receptive field and enrich the output, so as to enhance the ability of weak feature extraction. Secondly, a multi-scale feature fusion module was designed to enrich the abstract information of the shallow feature map. Finally, a dense prediction module was designed, and contextual information was integrated in the shallow layer in advance, so that the output of each layer in the final stage contained rich and closely related multi-scale feature information. The proposed algorithm was tested on UCAS_AOD and NWPU VHR-10 datasets, and the average detection accuracy reached 83.4% and 94.8%, respectively. The experimental results show that the proposed algorithm can effectively improve the accuracy of target detection in remote sensing images, and has a significant improvement for the problem of small-scale target detection in remote sensing images.