In order to address the growing complexity of models, introducing neural network pruning during training and deployment is a good option. We propose a pruning method that addresses two common issues in neural network pruning through a two-module approach. First, we use a pruning structure determination method based on the receptive field and effective receptive field. By modeling the ratio of convolutional layers as a measure of their effectiveness in feature extraction, we calculate the pruning ratio for each layer based on the overall reduction in floating-point operations. Next, we assess filter importance using a method based on the mean of feature maps. In pretrained models, filters with high activation contribute more significantly to the output and are more worthwhile to retain. Our two-module approach does not introduce any additional training parameters and is a non-learning framework. We conducted experiments on the CIFAR-10 and ImageNet-1k datasets, and the results indicate that our method is highly effective.