随着卷积神经网络的快速发展，卷积神经网络已经在图像识别和处理领域发挥了巨大作用，但是在不同频率的信息处理上仍然有可以提高的空间。本研究基于傅里叶变换设计低通滤波器，在传统的LeNet-5卷积神经网络模型上进行了特征提取的优化，发现了合适的低通滤波参数。在此基础上，经过每轮训练50次、迭代50次的三轮对比试验来对比原LeNet-5卷积神经网络模型和使用前置傅里叶变换低通滤波器的新LeNet-5卷积神经网络模型，发现测试集平均分类准确度提高了0.25%，平均训练时间缩短了约200秒。新模型在性能、可解释性、鲁棒性上较之传统模型均有提升。本研究通过应用傅里叶变换到卷积神经网络中，探索其在图像处理和模式识别任务中的潜在应用价值，并提出将来可以使用带通滤波器以及多重滤波器来进行进一步的优化研究。

With the rapid development of convolutional neural networks, convolutional neural networks have played a great role in the field of image recognition and processing, but there is still room for improvement in information processing of different frequencies. In this paper, the low-pass filter is designed based on Fourier transform, and the feature extraction is optimized on the traditional LeNet-5 neural network model, and the suitable low-pass filter parameters are found. On this basis, through three rounds of comparison tests with 50 training times per round and 50 iterations to compare the original LeNet-5 neural network model and the new LeNet-5 neural network model using the pre-Fourier transform low-pass filter, it is found that the average classification accuracy of the test set is increased by 0.25%, and the average training time is shortened by about 200 seconds. The new model has improved performance, interpretability and robustness compared with the traditional model. In this study, Fourier transform is applied to convolutional neural networks to explore its potential application value in image processing and pattern recognition tasks, and it is proposed that bandpass filters and multiple filters can be used for further optimization research in the future.