Abstract: In view of the correlation of neighboring pixels and characteristic of salt-pepper noise in nature images，a SVM（support vector machine）-based method is proposed for restore images corrupted by salt-pepper noise.Firstly，gray surface of image is optimally fitted by the learning mechanism of SVM.Then the support vectors are extracted from the training samples and decision function is built up as a training result.Accordingly，original intensity values of noise pixels are predicted using well-fitted gray surface.Compared with traditional median filters and average filters，the approach can remove noise efficiently while preserving the more detail information，especially for those images with high noise ratio.

Key words: salt-pepper noise, median filter, support vector machine, regression

摘要： 针对自然图像中相邻像素的相关性及其椒盐噪声的特点，提出了一种基于支持向量机的椒盐噪声消除方法。该方法应用支持向量机的学习机制对图像灰度曲面进行最佳拟合，并从训练样本中提取支持向量与相应的决策函数，最后根据决策函数在拟合曲面上进行噪声像素点的灰度值预测，从而恢复噪声点的原始信号。通过与传统的中值滤波和均值滤波进行实验对比，提出的方法可有效地去除椒盐噪声，同时最大限度地保留图像的细节信息，尤其对高密度椒盐噪声图像的处理效果更为理想。

关键词: 椒盐噪声, 中值滤波, 支持向量机, 回归