基于改进ConvNeXt的遥感图像目标检测算法

Remote Sensing Image Target Detection Algorithm Basedon Improved ConvNeXt

针对遥感图像中目标排列紧密、背景信息复杂、小目标众多导致的目标检测精度低的问题,结合YOLOv5s,提出了一种基于改进ConvNeXt的遥感图像目标检测算法。首先,在特征提取网络的底端引入了改进的ConvNeXt Block,通过大核卷积与自注意力交互扩宽感受野、丰富语义信息;其次,在多尺度特征融合部分加入了一组自底向上的金字塔结构,以放大浅层特征图的作用,弥补遥感图像中小目标因为深度卷积而损失的位置信息;最后,引入SIoU损失函数,重新定义惩罚指标,并加快整体网络的收敛速度。将所提出的检测算法在RSOD数据集上进行了消融实验,平均精准率均值为92.27%,实验结果表明,所提算法能够实现对遥感图像目标的准确检测。

To solve the problem of low target detection accuracy caused by closely arranged targets complex background information and numerous small targets in remote sensing images a remote sensing image target detection algorithm based on the improved ConvNeXt is proposed by using YOLOv5s.Firstly an improved ConvNeXt Block is introduced at the bottom of the feature extraction network to widen the perceptual field and enrich the semantic information through the interaction between large kernel convolution and self-attention.Secondly a set of bottom-up pyramidal structures is added to the part of multi-scale feature fusion to amplify the role of shallow feature maps and compensate for the position information of small targets in remote sensing images which is lost due to deep convolution.Finally the SIoU loss function is introduced to redefine the penalty index and accelerate the convergence of the overall network.The proposed detection algorithm is ablated on the RSOD dataset with a mean Average Precision(mAP)of 92.27%and the experimental results show that the proposed algorithm can realize accurate detection of remote sensing image targets.