For early filtering method, such as linear and nonlinear filtering, linear filtering includes Gaussian, mean and box filtering, nonlinear filtering includes median filtering and closed operation, which are determinant cycle operation at pixel level. According to the disadvantages of the traditional filtering methods mentioned of complicated calculation, data redundancy and image cannot be compressed effectively to perform digital transmission, a K-SVD filtering method based on principal component analysis (PCA) image fusion is proposed. And the disadvantages of single K-SVD such as without better filtering effect on salt and pepper noise is compensated effectively. N frames of images with noise are obtained through observing the source image many times, which have Gaussian and salt and pepper noise, both are additive noise. K-SVD filtering is performed after PCA extracting and fusing to N frames of images with noise. If K-SVD filtering is performed before PCA, the K-SVD filtering of multi-frame images is produced, which will lead to low efficiency and N times of calculation redundancy. So Gaussian noise interference is eliminated effectively and the disadvantage of being not sensitive to salt and pepper noise of K-SVD is resolved. And the research on data denosing at image feature level is finished.