To address the issue of insufficient frequency feature perception in image classification models, we propose an image classification network based on frequency-aware convolutions. First, we introduce the Fourier transform to design frequency-aware convolutions, embedding them into residual branches to fuse frequency features and enhance the model's sensitivity to frequency characteristics. Second, A multi-branch feature selection module has been designed to enhance information exchange between channels while simultaneously extracting multi-level features in a collaborative manner. By leveraging frequency-aware convolutions and the multi-branch feature selection module, the model achieves joint perception of both the spatial and frequency domains of images. Finally, experimental validation is conducted on datasets including CIFAR10, CIFAR100, SVHN, ImageNet, and ImageWoof. Results demonstrate that compared to current mainstream image classification models, the proposed model achieves optimal classification accuracy across all tests, validating its effectiveness in enhancing image classification performance.