基于预测和矢量量化的高光谱图像压缩算法

Hyperspectral image compression based on prediction and vector quantization

针对高光谱图像波段数目多,存储空间大,为后续研究带来极大不便的特点,提出了一种基于广义回归神经网络预测与矢量量化算法结合的高光谱图像压缩算法。该算法通过预测前一波段的像素数据,将符合要求的预测结果继续预测下一波段,通过设置合理的参数,在预测步骤中95%以上的波段可通过预测结果来预测波段数据。利用矢量量化算法对预测数据进行压缩。该算法只需要已知图像的前2个波段数据,即可预测整个图像波段的数据。广义回归神经网络具有很强的非线性映射能力和学习速度,预测效果好。通过预测得到的数据在不同的压缩比下进行实验,结果表明,在保证图像质量的前提下,该算法与对比算法相比,有效降低了运算复杂度,节约了时间,提高了峰值信噪比。

Aiming at the large number of bands in hyperspectral images and the large storage space, it poses a great inconvenience for follow-up research. A hyperspectral image compression algorithm based on generalized regression neural network prediction and vector quantization algorithm is proposed. By predicting the pixel data of the previous band, the algorithm continues to predict the desired prediction result for the next band. By setting reasonable parameters, more than 95% of the bands in the prediction step can predict the band data through the prediction results. Finally, the prediction data is compressed using a vector quantization algorithm. The algorithm only requires the first two bands of the known image to predict the entire image band. The generalized regression neural network has strong nonlinear mapping ability and learning speed, and has good prediction effect. The data obtained by prediction is tested under different compression ratios. The results show that under the premise of ensuring the image quality, the algorithm effectively reduces the computational complexity compared with the comparison algorithm, saves time, and improves the peak signal-to-noise ratio.