Rainy-day images often suffer from severe background blur and detail loss, significantly hindering computer vision applications to some extent. Effective rain removal hinges on capturing rich and accurate image features. This paper proposes the MMA, a novel image rain removal network based on multi-level mixed attention. MMA integrates a multi-scale attention module into a UNet architecture, and introduces a multi-head attention intermediate block to better capture global and local features by strengthening the perception of relationships between different image regions. Moreover, a feature fusion module is devised to integrate multi-scale features and augment the model's ability to process detailed features. Extensive experiments on synthetic and real datasets show that, compared to baseline models, the MMA network achieves an average improvement of 1.86 dB in peak signal-to-noise ratio and 1.71% in structural similarity index, demonstrating its effectiveness in image rain removal.