PINCurve-S: Physics-Informed Neural Curves with Spatial Attention for Efficient Low-Light Image Enhancement
Anubhav Jain, Nikhil Panwar, Vivek Kumar, Alireza Alaei, Partha Pratim Roy
Indian Institute of Technology Roorkee Southern Cross University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Existing low-light image enhancement methods face a fundamental trade-off between restoration quality and computational efficiency, limiting their deployment on resource-constrained edge devices. In this paper, a novel physics-informed neural network called PINCurve-S is proposed that formulates enhancement as a task of interpretable curve parameter estimation rather than pixel-wise mapping. The method trains a lightweight encoder to predict three physically meaningful parameters, namely gamma, exposure, and contrast, which drive a differentiable curve transformation for dynamic range adjustment. This physics-grounded formulation constrains the solution space to valid image transformations, enabling extreme model compression while preserving enhancement capability. The curve estimation is specially designed considering brightness preservation, color consistency, and spatial smoothness through carefully formulated loss functions. PINCurve-S is appealing in its efficiency, requiring only 20K parameters and 0.125 GFLOPs, which represents over 125 times reduction compared to recent quality-focused methods. Despite this simplicity, competitive performance is achieved with 19.96 dB PSNR on LOL-v2. Extensive experiments on LOL, LOL-v2, and SID benchmarks demonstrate the advantages of the proposed method in establishing a new Pareto frontier for the quality-efficiency trade-off.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIImage Enhancement Techniques
Advanced Image Processing Techniques · Image and Signal Denoising Methods