一种具有双层信息损失优化结构的遥感图像检索方法

Method of remote sensing image retrieval with double-layer information loss optimization structure

当前主流的图像检索方法在处理遥感图像时不能针对遥感图像信息丰富、特征维度高的特点,并且通过传统的特征提取方法得到的图像特征表达能力弱、信息损失严重,因此不能取得较高精度的检索结果。针对上述问题,提出具有双层信息损失优化结构的哈希编码方法用于遥感图像检索。首先,将经过傅里叶变换滤波降噪处理后的遥感图像数据输入卷积网络(convolutional neural network,CNN),通过多层卷积得到表达图像的深层特征向量;然后利用K-means算法对图像特征聚类,再在每个聚类内寻找最优的哈希函数,进而得到图像对应的二进制哈希码;最后利用汉明距离对图像进行相似性比较,完成对图像数据的有效检索。实验结果表明,对比于其他算法,该方法提高了检索的查准率、查全率以及平均检索精度,对于遥感图像有较好的适用性。

At present,the main image retrieval methods can not get a good retrieval result when processing remote sensing image because it can＇t consider characteristics of remote sensing images and the poor image expression ability and serious loss of information based on the traditional feature extraction methods. Regarding the issue above,this paper proposed a hash coding method with double-layer information loss optimization structure. This method firstly entered the remote sensing image data into CNN which had processed by Fourier transformation filter to noise reduction,and used the multi-layer convolution neural network to extract the deep features of the image. Then it clustered the feature of data with K-means algorithm. So in each cluster,it could get an accurate hash function and the corresponding binary hash code. Finally,it used the Hamming distance to compare the image similarity,and completed the effective retrieval of the image data. Experimental results show that,compared with the traditional method,this method improves the retrieval precision,recall and mean average precision. The proposed retrieval method has good applicability to remote sensing image.