Exploring the internal visualization techniques of deep neural networks, Class Activation Mapping (CAM) stands out as one of the most prominent directions in current interpretability methods. However, existing CAM methods introduced useless gradients to non-target regions, resulting in noises in saliency maps and limitations in generating fine-grained activation maps. To obtain finer-grained class activation maps and achieve more convincing visual explanations of neural networks, we propose WPG-CAM based on the Group-CAM method in this paper. In this approach, we first filter out noise information from the saliency maps and gradient matrices of the target layer by introducing pooling operations and addressing the low-resolution issue of image features through Gaussian upsampling. Secondly, we calculate the contribution ratio of information from different scales of activation maps using image information entropy. Lastly, we obtain refined activation maps by weighted average activation maps at different scales. Through implementation and analysis, our proposed WPG-CAM method effectively eliminates noise from saliency maps, outperforming other CAM methods in comprehensive metrics, thus providing more precise and higher granularity class activation images.