基于卷积神经网络的高分六号卫星多光谱图像压缩

GF-6 Multispectral Image Compression Based on Convolutional Neural Network

高分六号多光谱图像的空间冗余和谱间冗余较高,为了降低高分六号多光谱图像所占用的存储空间,提高国产高分多光谱图像的压缩效率,提出一种基于卷积神经网络的端到端多光谱图像压缩模型SMIC。SMIC模型由自编码器、量化结构、熵编码3个部分组成。自编码器通过卷积下采样提取图像的特征,降低数据的空间维度。量化结构采用多进制量化将特征矩阵离散化,减少图像压缩过程中的信息损失。熵编码采用高斯混合模型进行编码,降低码流,减少图像所占用的存储空间。实验结果表明:在相同码率下SMIC模型的高分六号多光谱图像压缩效果明显优于传统图像压缩算法JPEG,重建图像的质量明显提高,图像的峰值信噪比较JPEG高约2 dB,且SMIC重建图像的误差值主要集中在[-100,100]范围内,区间占比达到80%以上;SMIC模型重建图像的NDVI与原始图像NDVI的决定系数R2为0.93;SMIC模型的冬小麦提取准确率为87.16%,误检率为4.47%,冬小麦提取结果验证了SMIC模型能够满足部分定量遥感的应用需求。

The spatial and inter-spectral redundancy of the multispectral images of Gaofen No.6(GF-6)are high.To reduce the storage space occupied by these images and improve their compression efficiency,an end-to-end multispectral image compression model SMIC based on a convolutional neural network is proposed.The SMIC model consists of three parts:self encoder,quantization structure,and entropy coding.The self encoder extracts image features through convolutional down sampling to reduce the spatial dimension of the data.The quantization structure uses the multi-band quantization structure to discretize the feature matrix,reducing information loss in the image compression process.The entropy coding adopts the Gaussian mixture model to reduce the code stream and storage space occupied by the image.Experimental results at the same bit rate show that the compression effect of the SMIC model's GF-6 multispectral images is significantly better than that of the traditional JPEG image compression algorithm,and the quality of the reconstructed image is significantly improved.The Peak Signal-to-Noise Ratio(PSNR)of the image is approximately 2 dB higher than that of JPEG,and the error value of the reconstructed image is mostly within[-100,100],accounting for more than 80%.The determination coefficient R2 between the Normalized Difference Vegetation Index(NDVI)extracted from the reconstructed image and NDVI of the original image is 0.93.The quantitative remote sensing results of a winter wheat extraction experiment show an SMIC model accuracy of 87.16%,and the false detection rate is 4.47%.The results of winter wheat extraction show that the SMIC model can meet the application requirements of partial quantitative remote sensing.