结合改进混合卷积模型的遥感影像变化检测  

Change Detection in Remote Sensing Image Based on Deep Convolutional Neural Network with Improved Hybrid Convolution

作　　者：代云锋 刘丽娜 DAI Yunfeng;LIU Lina(Ningbo Metallurgical Survey and Design Research Co.Ltd.,Ningbo,Zhejiang 315100,China;Ningbo Yinzhou District Bureau of Surveying and Mapping,Ningbo,Zhejiang 315100,China)

机构地区：[1]宁波冶金勘察设计研究股份有限公司,浙江宁波315100 [2]宁波市鄞州区测绘院,浙江宁波315100

出　　处：《遥感信息》2022年第6期53-59,共7页Remote Sensing Information

摘　　要：针对基于深度学习的变化检测模型搭建中提高变化检测精度这一难点,在综合考虑面向像元和面向对象变化检测算法的基础上,设计了一种基于改进混合卷积特征提取模块的变化检测模型。该模型结合多切片思想和并行神经网络结构,融合不同尺寸的卷积核获取丰富的多尺度特征。首先,利用超像素分割算法将测试影像分割成无重叠的同质性区域;然后,选取一定数量的样本对模型进行训练,得到测试影像的像素级变化检测结果;最后,利用投票法,将网络得到的像素级结果与分割对象相结合,得到最终的变化检测结果。实验结果表明,基于该方法的网络模型性能较好,该模型可以有效学习多时相影像中的空间信息及差异特征,同时结合分割算法能够降低虚检率和漏检率,有效提高了变化检测精度。Aiming at the difficulty of improvement the accuracy of change detection in the construction of change detection model based on deep learning, a change detection model based on improved hybrid convolution feature extraction module is designed on the basis of comprehensively considering pixel-oriented and object-oriented change detection algorithms. This model combines the technique of multi-slice and siamese neural network structure, and fuses different convolution kennels to obtain abundant multi-scale features. Firstly, the segmented object is obtained by superpixels segmentation and training samples are acquired by calculating the characteristics of segmentation object. Then, the neighborhood information of the training sample is taken as the input for training, and the pixel-level change detection result is obtained. Finally, a change detection map is obtained via an uncertainty analysis, which combines the superpixels segmentation with the output from the model. Experimental results demonstrate that the proposed method achieves comparable and better performance than that of the mainstream methods. This model can effectively learn the spatial information and difference features of multi temporal images. At the same time, combined with the segmentation algorithm, it can reduce the omission and commission rate, and improve the change detection accuracy.

关 键 词：改进混合卷积 多特征提取 多切片 深度特征融合 变化检测 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]