随着技术的不断发展, 遥感技术被广泛应用于地图绘制、资源勘探以及灾害预警等领域。遥感目标检测是进行遥感图像解译的关键步骤。传统的目标检测算法在对遥感目标进行检测的过程中存在目标漏检、检测精度低以及无法解决小目标检测等问题。提出一种基于多尺度特征增强卷积神经网络 (MSFE-CNNs) 的遥感目标检测算法, 通过对不同卷积层特征进行增强和融合, 使得模型具有更快的训练速度和更高的检测精度。所提算法结合特征提取模块、特征增强模块、自注意力机制和金字塔特征注意力机制。特征提取模块对输入的海量遥感数据进行特征提取, 获取不同类别目标的多尺度特征; 特征增强模块用于增强不同卷积层特征相关性, 强化模型的学习能力和特征之间的非线性关系; 自注意力机制和金字塔特征注意力机制主要解决传统卷积神经网络无法获取小尺度目标特征的问题。为了验证所提算法的有效性, 在DOTA数据集上进行不同方法对比, 实验结果表明所提算法在检测精度和训练速度上均优于现有基于深度学习的目标检测算法。

With the continuous development of remote sensing technology,it is widely used in the fields of map drawing,resource exploration and disaster early-warning.Remote sensing target detection is the key step of remote sensing image interpretation.In the process of detecting remote sensing targets,the traditional detection algorithm has some deficiencies,such as target missing,low detection accuracy and inability to detect small target.A remote sensing target detection algorithm based on Multi-Scale Feature Enhancement Convolution Neural Networks (MSFE-CNNs) is proposed.By enhancing and fusing the features of different convolution layers,the model has faster training speed and higher detection accuracy.The proposed algorithm combines feature extraction module,feature enhancement module,self-attention mechanism and pyramid feature attention mechanism.The feature extraction module extracts features from the input of massive remote sensing data to obtain multi-scale features of different types of targets.The feature enhancement module is used for enhancing the correlation of features of different convolution layers and strengthening the learning ability of the model and the nonlinear relationship between features.Self-attention mechanism and pyramid feature attention mechanism mainly solve the problem that traditional convolutional neural network can not obtain the features of small-scale targets.To verify the effectiveness of the proposed algorithm,different algorithms are compared on DOTA data sets.Experimental results show that the proposed algorithm is superior to the existing target detection algorithms based on deep learning in both detection accuracy and training speed.