一种注意力机制特征融合的小目标检测模型

Small object detection model based on feature fusion of attention mechanism

针对图像中小目标的特征难以有效提取,从而对小目标的检测不利的问题,提出了一种通道-空间注意力机制特征融合的小目标检测模型.该模型以Faster R-CNN作为基础检测模型,首先设计了一种基于通道-空间注意力机制的特征融合方法,用于降低特征融合过程中引起的混叠效应;然后设计了一种跳跃残差连接模块用于降低特征融合过程中高层特征信息的丢失;最后基于ResNet101深层特征提取能力强的特点,使用其提取特征,将提取的特征采用通道-空间注意力机制特征融合方法融合生成特征金字塔网络,并将生成的特征金字塔网络作为Faster R-CNN的主干网络.在NWPU VHR-10数据集上对小目标检测的实验结果表明:本文模型的平均检测精度为82.5%,高于DSSD(55.4%)、FSSD(77.3%)、TDFSSD(76.8%)、Faster R-CNN(44.2%)和FPN(68.9%)的平均检测精度.

Aiming at the disadvantage of small object detection caused by the difficulty of feature extraction in the image,a small object detection model was proposed based on the feature fusion of channel-space attention mechanism. In this model, Faster RCNN was used as the basic detection model. First, a feature fusion method was designed based on channel-space attention mechanism to reduce aliasing effect caused by feature fusion. Secondly, a jump residual connection module was designed to reduce the loss of high-level feature information in the process of feature fusion. Finally, based on the strong deep feature extraction ability of ResNet101,it was used to extract features in fast Faster R-CNN.The extracted features were fused using the feature fusion method of channel spatial attention mechanism proposed to generate the feature pyramid network,and the generated feature pyramid network was used as the backbone network of Faster R-CNN.The experimental results of small object detection on NWPU VHR-10 data set show that, the average detection accuracy of this model is 82.5%, which is better than deconvolutional single shot detector’s 55.4%(DSSD),feature fusion single shot multibox detector’s 77.3%(FSSD),top-down feature fusion single shot multiBox detector’s 76.8%(TDFSSD),Faster R-CNN’s 44.2% and feature pyramid network’s 68.9%(FPN).