军事知识服务平台
基于全面零空间正则化的工业铸件表面图像超分辨率方法
A Super-Resolution Method for Industrial Casting Surface Image Based on Comprehensive Null-Space Regularization
摘要 图像超分辨率技术在工业场景中被广泛应用,可用于恢复铸件表面细节、增强纹理,并支持异常检测等自动化任务.由于感知-失真权衡的存在,现有的基于回归的图像超分辨率方法虽然能实现较低的失真度,但细节恢复效果较差,会导致明显的模糊;而基于生成图像的超分辨率方法虽能提升感知质量,却容易引入不真实的伪纹理.为解决上述问题,重建具有真实丰富细节的铸件超分辨率图像,本文提出一种基于全面零空间正则化(CNSR)的图像超分辨率方法.该方法通过引入零空间约束,分别在先验提取层面、纹理生成层面和训练层面对扩散模型进行全面的正则化约束,在保留扩散模型强大的细节生成能力的同时,约束其生成忠实于原图的细节纹理.实验表明:本文提出的方法在工业铸件表面图像超分辨率数据集上取得了优越的性能,并能有效提升后续工业异常检测任务的精度.
Abstract:
Image super-resolution is widely applied in industrial scenarios,which be used to restore the surface details of casting images,enhance textures,and support automated tasks such as anomaly detection.Due to the existence of the perception-distortion trade-off,existing regression-based image super-resolution methods can achieve low distortion but poor detail restoration,resulting in obvious blurring.On the other hand,generation-based image super-resolution methods can improve the perceptual quality,but they are prone to introducing unrealistic artifacts.To address the above issues and reconstruct super-resolution images with realistic and rich details,this paper proposed an image super-resolution method based on Comprehensive Null-Space Regularization(CNSR).This method introduces null-space constraints to comprehensively regularize the diffusion model at the prior extraction level,texture generation level,and training level.While retaining the powerful detail generation ability of the diffusion model,it restricts the model to generate texture details that are faithful to the original image.Experiments show that the method proposed in this paper has achieved superior performance on the industrial casting surface image super-resolution dataset,and effectively enhanced the accuracy of subsequent industrial anomaly detection tasks.
年,卷(期): 2026, 65(5)
分类号: TP391.4
机标分类号: TP391.41 TN911.73 F713.55
在线出版日期: 2026年11月4日
基金项目: 62072112:国家自然科学基金,