In recent years, learning image compression has attracted widespread attention from researchers. Existing learning image compression methods have failed to fully capture the complex texture representation of images, hindering further improvement in image compression efficiency. Therefore, a learning image compression method based on multi-scale feature fusion is proposed in the paper. Firstly, a multi-scale feature fusion module is constructed using generalized split normalization at the encoder to capture multi-scale fine-grained feature information of the image. Then, an enhanced inverted bottleneck residual blocks is introduced at the decoder to expand the high-dimensional channel space of texture detail representation. Finally experimental results show that compared with SOTA methods, the proposed method can averagely reduce over 4% bitrate under similar quality on public datasets.