Due to the limitations of the acquisition equipment, sensors, and the illumination or reflection characteristics of the ground, the acquired point clouds will inevitably be noisy. Noise degrades the quality of point clouds and hinders the subsequent point cloud processing tasks, so the denoising technique becomes a crucial step in point cloud processing. This paper proposes a point cloud denoising algorithm based on a hybrid attention mechanism, which takes into account the complexity of the internal features of point clouds and the randomness of point cloud transformations. Generates channel and spatial attention by parallel maximum pooling and average pooling of point cloud data, trains adaptive attention weights using a multilayer perceptron with shared weights, and serially fuses them, multiplies them with the input features to obtain more robust point cloud features, and connect to the score estimation module using the residuals. By studying and analyzing the mechanism proposed in this paper, it is experimentally demonstrated that the performance of the proposed model under various noise models is vastly improved over the baseline network and outperforms the advanced denoising methods without significantly increasing the network operation cost.