基于多维损失函数优化FSRNet的气象观测图像超分辨率重构算法  

Optimization of FSRNet Super-Resolution Reconstruction Algorithm of Meteorological Observation Image Based on Multi-Dimensional Loss Function

作　　者：王瑞祥 WANG Ruixiang(Meteorological Observation Data Center of Henan Province,Zhengzhou 450003,China;CMA·Henan Key Laboratory of Agrometeorological Support and Applied Technique,Zhengzhou 450003,China)

机构地区：[1]河南省气象探测数据中心,河南郑州450003 [2]中国气象局·河南省农业气象保障与应用技术重点实验室,河南郑州450003

出　　处：《河南科技》2023年第5期37-41,共5页Henan Science and Technology

基　　金：中国气象局/河南省农业气象保障与应用技术重点实验室应用技术研究基金项目(KQ202211)。

摘　　要：【目的】传统FSRNet生成的超分辨率图像存在伪影、模糊等情况,作为志愿气象观测的试点省份,河南省气象局招募了大批志愿观测站,以期解决部分气象观测图像存在较低的分辨率的问题。【方法】首先引入热图损失、目标注意力损失和对抗性损失构成多维损失函数,对FSRNet进行模块优化,使用转置卷积放大低分辨率的图像。其次对模型进行分步训练,先对低分辨率观测图像进行粗略SR网络训练,再使剩余网络融入训练。【结果】多维损失训练的引入,降低了算法复杂度,提升了粗略SR网络的性能,解决了SR网络训练时调参困难等相关问题,提高了超分辨率气象观测图像的质量。【结论】试验结果证明,本方法在气象领域观测图像的优化相比于其他算法效果更佳,可以生成质量更高、细节更加清晰的目标观测图像。[Purposes]The super-resolution images generated by the traditional FSRNet have artifacts,blurring,etc.As a pilot province for voluntary meteorological observation,Henan Meteorological Bureau has recruited a large number of voluntary observation stations in order to solve the problem of low resolu⁃tion of some meteorological observation images.model of meteorological observation images is con⁃structed.[Methods]Firstly,heat map loss,target attention loss and adversarial loss were introduced to form a multidimensional loss function.The module of FSRNet was optimized,and transposed convolution was used to enlarge the low-resolution image.Secondly,the model is trained step by step.First,the low resolution observation images are trained by rough SR network,and then the remaining network is inte⁃grated into the training.[Findings]The introduction of multi-dimensional loss training reduces the algo⁃rithm complexity,improves the performance of rough SR network,solves related problems such as the difficulty of parameter adjustment during SR network training,and improves the quality of super resolution meteorological observation image.[Conclusions]Experimental results show that this method is more ef⁃fective than other algorithms in optimizing observation images in meteorological field,and can generate observation images with higher quality and clearer details.

关 键 词：志愿气象观测 超分辨率 多维损失 先验信息 生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]