An Improved Low-Bit-Rate Image Compression Framework Based on Semantic-Aware Model and Neighborhood Attention

In recent years, image compression techniques based on deep neural networks have achieved significant advancements, outperforming traditional methods in delivering higher compression efficiency at lower bit rates. However, existing approaches often result in degraded image quality at low bit rates, leading to distortions in critical regions such as faces and text during decoding. To address these limitations, we propose a robust image compression framework based on a semantic-aware model and a hyper-prior encoder with neighborhood attention. First, we utilize the encoder of the semantic-aware model to transform the input image into a latent space Z. To further improve the information representation within the latent space Z, we design a hyper-prior encoder, which leverages neighborhood attention to perform feature enhancement and transformation. This process can minimize quantization errors and facilitate efficient vector quantization. The refined latent space is then quantized using a residual vector quantization technique to ensure efficient and compact representation. Finally, entropy coding is applied to the quantized latent space, enabling high compression efficiency. Experimental results on public benchmark datasets show that our proposed framework achieves comparable results to current mainstream image compression methods. In addition, our proposed algorithm effectively mitigates distortions in facial and textual regions while preserving the structural integrity and visual fidelity. In future work, we plan to improve the model’s robustness under low-light conditions and enhance compression efficiency through lightweight optimization techniques, enabling broader real-world deployment.