针对海底混响中的乘性斑点噪声使侧扫声呐图像中的目标无法准确识别的问题,提出了 1种利用同步码字优化字典学习法与相关正则化相结合的降噪方法.该方法利用侧扫声呐图像(side-scan sonar image,SSI)的稀疏性,同时更新任意一组码字和相应的稀疏系数,即同步码字优化(simultaneous codeword optimization,SimCO),得到合适的字典;并将乘性噪声对数变换成加性噪声,利用斑点噪声的伽马分布特性,构造出相应对数似然函数;最后利用正则化减少过拟合化特性,采用最大似然估计(maximum likelihood estimation,MLE)法估计出待恢复图像,实现声呐图像降噪.仿真结果表明,该方法降噪后图像可保持好的边缘信息,并且能有效降低降噪前后图像的平均绝对误差(mean absolute-deviation error,MAE),与传统MOD与K-SVD降噪法相比,等效视数(equivalent number of looks,ENL)可以提高40.17％,MAE值可以降低23.43％,降噪后声呐图像视觉效果有明显提升.
A sonar image noise reduction method that combines simultaneous codeword optimization and regularization
For multiplicative spot noise in the seafloor reverberation,the target in the side scan sonar image cannot be accurately identified.In this paper,a noise reduction method that combines the dictionary learning method of syn-chronous code word optimization with the correlation regularization is proposed.This method takes usage of the spar-sity of the side scan sonar(SSS)image,and updates any set of codewords and the corresponding sparse coeffi-cients,that is,Simultaneous Codeword Optimization(SimCO),to obtain a suitable dictionary;The logarithmic multi-plicative noise is converted into additive noise,and the gamma distribution characteristics of spotted noise are used to construct the corresponding log-likelihood function.Finally,regularization is used to reduce the overfitting char-acteristics,and the maximum likelihood estimation(MLE)method is used to estimate the image to be recovered,and the sonar image noise reduction is realized.Simulation results show that,the image can keep good edge informa-tion after noise reduction,and can effectively reduce the mean absolute error(MAE)of the image before and after noise reduction,compared with the traditional noise reduction method,the equivalent number of looks(ENL)can be increased by 40.17％,MAE value can be reduced by 23.43％,and the visual effect of sonar images after noise reduction is significantly improved.
side scan sonar imagemultiplying noisesynchronization code word optimizationregularizationimage noise reduction
万方数据
维普
