面向遥感图像的多阶段特征融合目标检测方法

Multi-Stage Feature Fusion Object Detection Method for Remote Sensing Image

遥感图像目标具有多尺度、大横纵比、多角度等特性,给传统的目标检测方法带来了新的挑战.针对现有方法应用于目标尺度小、横纵比例不均衡的遥感图像时存在的精度下降问题,提出一种基于多阶段特征融合的目标检测方法MF2M(Multi-stage Feature Fusion Method).该方法在一阶段对特征图通道进行组合拆分,再采用卷积拼接的融合方式聚合通道维度的特征,从而强化输出的目标空间轮廓信息;二阶段设计多比例的非对称卷积块,增强大横纵比目标的高维全局特征,改善目标与检测框匹配粗糙的问题,同时利用串并行相结合的处理方式减少冗余卷积参数,加速网络收敛.在DOTA(Dataset for Object deTection in Aerial images)数据集上的实验结果表明,基准方法引入MF2M后,在保证检测速度的前提下精度指标mAP提高至76.44%,结果验证了所提算法的有效性与可靠性.

The remote sensing image objects has the characteristics of multi-scale,large aspect ratio,multi-angle and so on,which brings new challenges to traditional object detection methods.To solve the problem of loss of accuracy when existing methods are applied to remote sensing images with small object scales and unbalanced aspect ratios,an object detection method based on dual-stage feature fusion—MF2M(Multi-stage Feature Fusion Method)is proposed.This method combines and splits the feature map channels in first stage,and then adopts the fusion method of convolution splicing to aggregate the characteristics of the channel dimensions,thereby enhancing the output object spatial contour information;in the second stage,we design a multi-scale asymmetric convolution blocks,enhancing the high-dimensional global features of large aspect ratio targets,improving the problem of rough matching between the target and the detection frame,and using a combination of serial and parallel processing to reduce redundant convolution parameters.Finally,we achieve the effect of accelerating network convergence.The experimental results on the DOTA(Dataset for Object deTection in Aerial images)dataset show that after the benchmark method is introduced into MF2M,the accuracy index mAP is increased to 76.44%under the premise of ensuring the detection speed.The results verify the effectiveness and reliability of the algorithm.