基于零次学习SelfDeblur机器视觉动态检测图像去模糊方法

Deblurring the image for machine vision-based detection in dynamic scenes via SelfDeblur with zero-shot learning

机器视觉动态检测容易生成模糊图像,研究图像去模糊方法是提高机器视觉检测性能重要手段。端到端图像去模糊方法受庞大数据集以及成对输入要求的制约,SelfDeblur则通过零次学习降低数据集要求,可用于机器视觉动态检测。根据机器视觉动态检测特定应用场景,将SeDeblur应用于非盲去模糊,并改进其损失函数、网络结构以及训练优化方法,实现数据集低要求的端到端图像去模糊。试验表明所述方法图像去模糊质量优于现有方法,无参考指标平均提高34.45%。

Machine vision-based detection in dynamic scenes generates motion blurred images easily,therefore,the image deblurring is considered as an appropriate solution to tackle this problem.The end-to-end image deblurring is limited by the demands such as massive data set and paired inputting,while the SelfDeblur is applied in machine vision-based detection via zero-learning to lower the data set.According to specific scene for machine vision-based detection,Self Deblur is implemented for non-blind image deblurring,and improving its loss function,network architecture and training optimization,which has addressed the issue for end-to-end image deblurring with low demand of data set.The experimental results have shown that the proposed method outperformed the existing approaches,and the referenceless image quality evaluator increased by 34.45%on average.