基于卷积神经网络的多光谱图像多标签场景分类

CNN based multilabel scene classification method for multispectral image

基于实现多光谱图像的多标签场景分类为目的,采用卷积神经网络的方法,通过计算数据集中所有样本标签的共现矩阵,利用共现矩阵为每个标签分配不同的权重,提出了一种新的计算损失函数的方法。所设计的卷积神经网络能够充分利用除了红绿蓝三通道之外的光谱信息,同时也能够利用已有的预训练的卷积神经网络权重进行参数的初始化,使得网络能够快速收敛。所提出的算法在Planet Amazon数据集上取得了最高的F值,从而得出了该算法具有高准确率和高可行性的结论。

Based on the purpose of multilabel scene classification of multispectral image, In this paper, the convolutional neural network （CNN）is used to do feature extraction and classification, the co-occurrence matrix of the labels of all samples are obtained, then different labels are assigned corresponding different weights, and the final loss function can be computed by these weights. The new designed convolutional neural network can not only make use of the extra spectra information except for red, green and blue channel, but also can it leverage the existed pretrained weights on imagenet, which can help our CNN converging much faster. The proposed algrithm achieves the highest F score on Planet Amazon dataset, which verifies its high accuracy and effectiveness.