To mitigate the impact of multi-scale noise present in 3D point cloud data acquired through 3D laser scanning on subsequent processing stages, a denoising approach leveraging normal vector direction information for feature classification has been developed. Firstly, the algorithm removes the large-scale noise in the point cloud model by statistical filtering combined with radius filtering, and then uses the principal component analysis method to obtain the normal vector information of the point cloud, constructs the histogram of normal orientations through the angle between the local neighborhood normal vectors, divides the point cloud into a plane area with less feature details and a feature area with rich feature details. For different regions, bilateral filtering and adaptive guided filtering are used to denoise small-scale noise in point clouds. Experimental results show that the proposed method can effectively remove the multi-scale noise of the point cloud model, and also maintain the feature details of the model well.