高效Mamba驱动的端到端光场图像压缩

打开文本图片集

关键词：光场；图像压缩;端到端;Mamba

中图分类号：TP394.1 文献标识码：A

doi： 10. 37188/OPE. 20253318.2980 CSTR：32169.14.OPE.20253318.2980

Abstract： Light field images capture both spatial angular information light rays，providing richer visual information than traditional 2D images.However，their high-dimensional nature poses challenges for existing compression methods in terms global feature utilization，long-range correlation exploration， computational complexity，limiting the improvements compression performance eficiency.To address these issues，this paper proposed an eficient Mamba-driven end-to-end light field image compression method.Firstly，2D slices containing spatial epipolar plane information were extracted from the 4D light field image， Mamba was employed to fully capture their global contextual information. Secondly，to scan the light field image in multiple directions while avoiding a significant increase in computational complexity，a channel-eficient 2D selective scanning strategy was introduced to extract light field features accurately eficiently. Finally，on the decoding end，a residual reconstruction module was designed to enhance the reconstructed image quality while reducing the number parameters decreasing the encoding decoding time. The experimental results show that compared with the existing representative method SADN， the proposed method achieves an average bitrate reduction 7.4% a PSNR improvement O.37 dB on light field images with a 7×7 angular resolution，while also demonstrating superior subjective visual quality. In terms encoding decoding time，the proposed method has achieved a significant improvement 1O to 20 times.Furthermore，compared to the state--the-art method LFICDRASC， the proposed method achieves an average bitrate reduction 19.5% a PSNR improvement 0.58dB on light field imageswith a 13×13 angular resolution.

Key words： light field；image compression;end-to-end；Mamba

1引言

光场图像（LightFieldImages，LFI）能够同时记录光线的角度和空间信息，使其在数字重聚焦1、光谱光场成像、光场渲染3等任务中展现出独特的优势。（剩余24191字）