SFNIC: Hybrid Spatial-Frequency Information for Lightweight Neural Image Compression

Neural image compression (NIC) has shown remarkable rate-distortion (R-D) efficiency. However, the considerable computational and spatial complexity of most NIC methods presents deployment challenges on resource-constrained devices. We introduce a lightweight neural image compression framework designed to efficiently process both local and global information. In this framework, the convolutional branch extracts local information, whereas the frequency domain branch extracts global information. To capture global information without the high computational costs of dense pixel operations, such as attention mechanisms, Fourier transform is employed. This approach allows for the manipulation of global information in the frequency domain. Additionally, we employ feature shift operations as a strategy to acquire large receptive fields without any computational cost, thus circumventing the need for large kernel convolution. Our framework achieves a superior balance between rate-distortion performance and complexity. On varying resolution sets, our method not only achieves rate-distortion (R-D) performance on par with versatile video coding (VVC) intra and other state-of-the-art (SOTA) NIC methods but also exhibits the lowest computational requirements, with approximately 200 KMACs/pixel. The code will be available at https://github.com/baoyu2020/SFNIC.