K Singular Value Decomposition (KSVD) dictionary learning has been successfully applied in the field of seismic data denoising. In order to make more efficient, full and accurate use of data, this paper proposes a Monte Carlo-filter dictionary learning (MC-FDL) method based on Monte Carlo segmentation for seismic data denoising, which uses Monte Carlo segmentation method to obtain pre-trained dictionaries, this enables the dictionary to learn the characteristics of the signal to a greater extent. First, the variance of all the data blocks is calculated, a uniformly distributed random number is generated for each data block, and if the variance of the block is greater than the random number, the block is selected as an atom in the dictionary. Then the dictionary is trained and updated by KSVD algorithm, and the interference of noise in dictionary atoms is filtered by median filter to make the dictionary more prominent in the feature of seismic data. After that, the trained dictionary is used for seismic data denoising, and the denoising performance of the proposed method is tested by synthetic and real data. Finally, the influence of different block selection methods and different data preprocessing methods on the denoising results is discussed in detail, and the block selection strategy of the proposed method is determined, and the future work is prospected.