基于卷积神经网络的遥感图像变化检测

Remote Sensing Image Change Detection Based on Convolutional Neural Network

主要研究了针对目标区域的基于卷积神经网络的变化检测方法,对比了两种卷积神经网络的方法,包括使用以VGG16为网络结构,将图像裁剪为16×16的图块,以这些图块为单元,为每个图块做变化或不变化的标签,最终生成基于图块的变化图像。基于这种方法适合用于大面积变化明显的区域,对于变化范围较小的图像则不能很好地判断是否变化,其次是由于设置的图像块比较大,因此在提高训练速度的同时,难以保证良好的视觉效果。第二种是以Siamese Network为网络框架,基于编码解码的网络结构对其进行改进,通过两个通道分别输入前后时相的图像,每一次卷积后相减并利用跳跃连接与解码端的图像叠加提取特征,获得了良好的测试结果。

This paper mainly studies the variation detection method based on convolutional neural network for the target region,and compares the methods of two convolutional neural networks,including using VGG16 as the network structure,and cutting the image into 16×16 tiles.The tile is a unit,and the label is changed and unchanged for each tile,and finally a tile-based change image is generated.Based on this method,it is suitable for areas with large changes in large areas.For images with small variation range,it is not easy to judge whether the change is made.Secondly,because the set image blocks are relatively large,it is difficult to guarantee the training speed.Good visual effect.The second is to use Siamese Network as the network framework,based on the VGG16 network structure to improve it,input the images of the front and back phases through two channels,subtract each time after convolution and superimpose the image with the deconvolution process.Features and obtained good test results.