融合坐标注意力机制的植物低照度多光谱图像增强模型  

Low-light multispectral plants image enhancement model incorporating with coordinate attention mechanism

作　　者：张伯驹 朱启兵 黄敏 赵鑫 ZHANG Boju;ZHU Qibing;HUANG Min;ZHAO Xin(Key Laboratory of Advanced Process Control for Light Industry of Ministry of Education,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学轻工过程先进控制教育部重点实验室,无锡214122

出　　处：《激光技术》2025年第1期79-86,共8页Laser Technology

基　　金：国家自然科学基金资助项目(62273166)。

摘　　要：为了解决短时曝光条件下高(多)光谱图像照度和信噪比偏低问题,提出了一种基于改进残差(Res-UNet)的高(多)光谱图像增强模型。该模型以Res-UNet为主干网络,采用空洞空间金字塔池化和坐标注意力机制强化模型的特征聚合能力,并引入标准分数损失函数以改善模型对光谱曲线的重构能力,利用图像增强质量指标和下游任务(干旱叶片分割精度)对所提模型的性能进行评价。结果表明,改进后模型的峰值信噪比、结构相似度和光谱角映射分别达到0.9852、39.71和3.120,优于对比算法;其对于干旱叶片的分割精确率也高于各类主流算法。该模型对植物低照多光谱图像增强的有效性,可为各类光谱图像下游任务提供信息支持。In order to solve the problem of low illumination and signal-to-noise ratio of hyperspectral(multispectral)images under short exposure conditions,a hyperspectral(multispectral)image enhancement model was proposed based on the improved residual U-Net(Res-UNet).The Res-UNet was taken as the backbone network,atrous spatial pyramid pooling and coordinate attention were used to enhance the feature aggregation ability of the model,and the Z-score loss function was introduced to improve the reconstruction ability of the model on spectral curves.The performance of the proposed model was evaluated using image enhancement quality metrics and a downstream task(drought leaf segmentation accuracy).The peak signal-to-noise ratio,structural similarity and spectral angular mapping of the improved model reach 0.9852,39.71 and 3.120,which are better than the comparative algorithms.And its segmentation accuracy for arid leaves is higher than that of various mainstream algorithms.The result shows the effectiveness of this model for plant low-light multispectral image enhancement,which can provide information support for various spectral image downstream tasks.

关 键 词：图像处理 多光谱图像 注意力机制 低照度图像 图像增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]