摘要
针对空间遥感图像数据量剧增的问题,提出一种改进的BP神经网络图像压缩方法。该算法利用Levenberg-Marquart算法提高神经网络的收敛速度,利用α算法提高神经网络的泛化能力。比较分析了新算法和标准BP算法对同一幅图像进行压缩的结果和性能误差函数。实验结果表明,实验结果表明,标准BP算法在图像压缩比为1/2时,均方误差(MSE)为343.3750;改进后的BP算法在图像压缩比为1/16时,MSE为69.5796,图像压缩比为1/8时,MSE为20.9561,图像压缩比为1/4时,MSE为5.5123。并且利用改进后的算法压缩图像的峰值信噪比均在30dB^40dB之间。改进算法已用于实际工程中,满足实际需求。
To overcome the dramatically increasing data amount of space remote sensing image, an improved back-propagation（BP） neural network was put forward to compress it. The algo- rithm used the Levenberg-Marquart algorithm to improve the convergence speed of neural net- work and used a algorithm to improve the generalization ability of neural network. We com- pared and analyzed the compression result and error performance function of the improved algo- rithm and the standard BP algorithm to the same image. The experimental results show that, when the image compression ratio is 1/2, the mean square error （MSE） of standard BP algo- rithm is 343. 3750; for improved BP algorithm, the MSE is 69. 5796 when the image compres- sion ratio is 1/16, the MSE is 20. 9561 when the image compression ratio is 1/8, and the MSE is 5. 5123 when the ratio is 1/4. Moreover, the peak signal-to-noise ratio （PSNR） of the im- proved algorithm is always in the range from 30 dB to 40 dB. The improved algorithm has been applied in practical engineering, which meets the need of practical work.
出处
《应用光学》
CAS
CSCD
北大核心
2013年第6期974-979,共6页
Journal of Applied Optics
基金
中国科学院二期创新基本资助项目(C05T022)
