基于收缩卷积神经网络的高分辨率影像分类方法

High resolution image classification method based on contractive convolution neural network

针对高分辨率遥感影像上存在同类地物差异性大、类间差异性小等问题,提出一种端到端收缩卷积神经网络的高分辨率遥感影像分类方法。首先,将卷积神经网络的最大池化层设置在卷积层之后,确保提取深度特征具有空间不变特征;然后,针对SoftMax分类器缺少对测试数据可分性显示建模,依据SoftMax分类器原理提出聚类损失函数与分类器损失函数融合获取影像分类目标函数,确保训练数据与类别中心接近,且不同类中心彼此分离;最后,采用实验数据验证本文方法的可行性。结果表明,本文方法能够提高影像特征的可分性,有效解决高分辨率遥感影像相似场景的错分问题。

Aiming at the problems of large differences between similar ground objects and small differences between classes in high-resolution remote sensing images,an end-to-end shrinkage convolution neural network classification method for high-resolution remote sensing images is proposed.Firstly,the maximum pooling layer of convolutional neural network is set after the convolution layer to ensure that the extracted depth features have space invariant features;Then,aiming at the lack of modeling for the separability display of test data in SoftMax classifier,this paper proposes the fusion of clustering loss function and classifier loss function to obtain the image classification objective function,so as to ensure that the training data is close to the category center and different centers are separated from each other;Finally,experimental data are used to verify the feasibility of this method.Experimental results show that this method can improve the separability of image features and effectively solve the problem of misclassification of similar scenes in high-resolution remote sensing images.