改进YOLOX的遥感图像目标检测算法

Remote Sensing Image Object Detection Algorithm with Improved YOLOX

针对遥感图像目标检测算法复杂背景下目标检测精度低、小目标特征丢失的问题,提出一种改进YOLOX的遥感图像目标检测算法MYOLOX(modified YOLOX)。该算法在主干网络引入残差金字塔卷积模块(residual pyramid convolution module,RPCM)增强浅层特征图中的空间位置等细节信息,缓解下采样过程中的特征丢失。引入增强跨阶段局部块(improved cross stage partial block,ICSP)提取丰富的上下文信息并抑制噪声干扰,减少复杂背景及噪声干扰带来误检。将改进算法应用于使用DIOR数据集对NWPU VHR-10数据集扩充后数据集和SSDD数据集,MYOLOX算法检测平均精度均值(mean average precision,mAP)分别达到了80.8%和94.4%,较原算法提升了4.1和4.5个百分点。实验结果证明,改进后的算法能够明显提高遥感图像目标检测精度。

Aiming at the problems of low target detection accuracy under complex background and small target feature loss in remote sensing image object detection algorithm,this paper proposes a remote sensing image object detection algorithm MYOLOX(modified YOLOX).The residual pyramid convolution module(RPCM)is introduced into the backbone network to enhance the spatial location and other details in the shallow feature map,which alleviates the feature loss in the down-sampling process.The improved cross stage partial block(ICSP)is introduced to extract a wealth of contextual information and suppress the interference of noise,which can effectively reduce false detection problems caused by complex background and noise.The improved algorithm is applied to the augmented dataset of NWPU VHR-10 dataset and SSDD dataset using DIOR dataset.The mean average precision(mAP)of MYOLOX algorithm detection reaches 80.8%and 94.4%,which is 4.1 and 4.5 percentage points higher than that of the original algorithm.Experimental results show that the improved algorithm can significantly improve the accuracy of remote sensing image target detection.