一种基于CNN与梯度分水岭算法的卫星图像区域分割识别方法

Method of Satellite Images Region Segmentation and Recognition Based on CNN and Gradient Watershed Algorithm

卫星图像的准确分割与识别在军事、环境、民生方面都有着重要的研究意义与价值。传统的区域分割算法如分水岭算法、k-means算法等在错综复杂的卫星图像中表现不佳,且不能同时给出区域的类别。为解决上述问题,本文提出一种结合CNN与分水岭算法的图像区域分割方法。该方法首先使用人工标记的区域图像训练CNN(卷积神经网络)分类器,且使其具有旋转不变性及平移不变性,从而能适应不同状态下的图像分类。然后用分水岭算法对图像进行区域粗粒度的聚类,针对分割出的每一个候选区域,使用CNN分类器对其迭代打分,最后得到分割区域并给出识别结果。实验结果表明,该方法较传统方法有更好效果。

The accurate segmentation and recognition of satellite images is very important in military and environmental matters and for people＇s livelihoods. Traditional region segmentation algorithms, such as the watershed algorithm, k-means algorithm, etc., do not perform well on complex satellite images, and cannot simultaneously display the region category. To address this problem, a method of satellite images region segmentation is proposed based of a convolutional neural network （CNN） and the gradient watershed algorithm. Firstly, the artificial markers of regional images are used to train the CNN classifier to adapt to the different categories of image classification with rotation in variant and translation in variant. Then, the watershed algorithm is used for regional images＇ coarse-grained clustering. For each candidate region segmented, CNN classifiers were used to iterate and mark. The experimental region segmentation and the recognition results show that the proposed method is better than the traditional methods.