基于密集连接与特征增强的遥感图像检测

Remote Sensing Images Detection Based on Dense Connection and Feature Enhancement

遥感技术的快速发展使得遥感图像检测技术广泛应用于军事、农业、交通、城市规划等多个领域。随着遥感分辨率和数据体量的不断提升,通过人工处理数据的方法已经无法满足实时性需求,因此,实现高效、精准的自动化数据处理方式成为该领域的研究热点。针对遥感图像分辨率高、背景复杂、目标尺度小等特点,提出一种改进的YOLOv3算法,用以提升遥感图像的检测效果。在原始YOLOv3算法的基础上,使用改进的密集连接网络替换原有的DarkNet53作为基础网络,以提升网络输入和预测结果的尺度多样性。通过阀杆模块降低输入损失,同时在浅层特征图中加入特征增强模块,从而丰富特征图的感受野,强化网络对浅层特征信息的提取,在保证整体检测性能的同时使网络对遥感图像中、小目标的检测精度和鲁棒性均有所提升。在遥感图像数据集上进行多组对比实验,结果表明,相比原始YOLOv3算法,该算法的平均准确率提高9.45个百分点,在小尺度目标上的检测准确率提升更显著,达到11.03个百分点,且模型参数量得到有效缩减。

With the rapid development of remote sensing technology,remote sensing image detection is being widely used in military,agriculture,transportation,urban planning,and other fields.With the continuous improvements of the remote sensing resolution and data volume,a manual data processing method has been unable to meet the real-time requirements.Therefore,the realization of an efficient and accurate automatic data processing approach has become a research hotspot in this field.Aiming at the characteristics of a high resolution,complex background,and small target scale of a remote sensing image,this paper proposes an improved YOLOv3 algorithm to improve the detection effect of such an image.Based on the original YOLOv3 algorithm,the improved dense connection network replaces the original DarkNet53 as the basic network,thereby improving the scale diversity of network input and prediction results.This study reduces the input loss through the valve stem module,and adds a feature enhancement module to the shallow feature map,thereby enriching the receptive field of the feature map and strengthening the extraction of the shallow feature information of the network,while ensuring the overall detection performance,this study improves the detection accuracy and robustness of the network for medium and small targets in remote sensing images.This study also carried out several groups of comparative experiments on remote sensing image datasets.The results show that,compared with the original YOLOv3 algorithm,the average accuracy of the algorithm is improved by 9.45 percentage points,and the detection accuracy of small-scale targets is increased even more significantly,reaching 11.03 percentage points.In addition,the number of model parameters is effectively reduced.