Convolutional neural network (CNN) models have been widely used for fundus-based disease screening, but the model deployment is challenging due to the large demand for computing resources. The low-rank decomposition is usually used to compress CNN models. However, the reduction of model parameters often leads to performance degradation. Therefore, we propose a factorized convolution with spectral normalization, named FConvSN, to reduce the model complexity while maintaining an ideal performance. FConvSN applies the spectral norm to constrain the weight in the direction of the spectral norm to achieve weight decay, thereby improving the generalizability. Since the features of fundus images appear as a highly skewed distribution, factorized convolution can be used to promote the sharing of convolution parameters, and spectral normalization can further prevent excessive weight in the spectral norm direction. We have conducted experiments on the fundus dataset to prove that our FConvSN can achieve performance comparable to standard convolution.