结合深度学习与条件随机场的遥感图像分类

发布时间：2018-04-26 19:17

本文选题：遥感图像分类 + 深度学习　；参考：《中国图象图形学报》2017年09期

【摘要】：目的为进一步提高遥感影像的分类精度,将卷积神经网络(CNN)与条件随机场(CRF)两个模型结合,提出一种新的分类方法。方法首先采用CNN对遥感图像进行预分类,并将其类成员概率定义为CRF模型的一阶势函数;然后利用高斯核函数的线性组合定义CRF模型的二阶势函数,用全连接的邻域结构代替常见的4邻域或8邻域;接着加入区域约束,使用Mean-shift分割方法得到超像素,通过计算超像素的后验概率均值修正各像素的分类结果,鼓励连通区域结果的一致性;最后采用平均场近似算法实现整个模型的推断。结果选用3组高分辨率遥感图像进行地物分类实验。本文方法不仅能抑制更多的分类噪声,同时还可以改善过平滑现象,保护各类地物的边缘信息。实验采用类精度、总体分类精度OA、平均分类精度AA,以及Kappa系数4个指标进行定量分析,与支持向量机(SVM)、CNN和全连接CRF相比,最终获得的各项精度均得到显著提升,其中,AA提高3.28个百分点,OA提高3.22个百分点,Kappa提高5.07个百分点。结论将CNN与CRF两种模型融合,不仅可以获得像元本质化的特征,而且同时还考虑了图像的空间上下文信息,使分类更加准确,后加入的约束条件还能进一步保留地物目标的局部信息。本文方法适用于遥感图像分类领域,是一种精确有效的分类方法。
[Abstract]:Objective to further improve the classification accuracy of remote sensing images, a new classification method is proposed by combining the two models of convolution neural network (CNN) and conditional random field (CRF). Methods the remote sensing images were preclassified by CNN, and the probability of their class members was defined as the first-order potential function of the CRF model, and then the second-order potential function of the CRF model was defined by the linear combination of Gao Si kernel function. Then the region constraint is added to obtain the super-pixel by using Mean-shift segmentation method, and the classification results of each pixel are corrected by calculating the posterior probability mean of the super-pixel. Finally, the mean field approximation algorithm is used to infer the whole model. Results three groups of high resolution remote sensing images were used to classify ground objects. This method can not only suppress more classification noise, but also improve the over-smoothing phenomenon and protect the edge information of all kinds of ground objects. Compared with support vector machine (SVM) and fully connected CRF (CRF), the experimental results show that the accuracy obtained is significantly improved compared with the four indexes of class accuracy, total classification accuracy, average classification accuracy and Kappa coefficient. Among them, AA increased 3.28%, OA increased 3.22% and Kappa increased 5.07%. Conclusion the fusion of CNN and CRF models can not only obtain the essential feature of pixel, but also consider the spatial context information of image, so that the classification can be more accurate. The constraint condition can further preserve the local information of the object. This method is applicable to the field of remote sensing image classification and is an accurate and effective classification method.
【作者单位】：南京理工大学计算机科学与工程学院;公安部第三研究所;
【基金】：国家自然科学基金项目(61371168) 江苏省科技支撑计划基金项目(BE2014646) 南京市科技计划基金项目(201505026)~~
【分类号】：TP751

【相似文献】