A knowledge augmented image deblurring method with deep learning for in-situ quality detection of yarn production

In the in-situ quality detection of yarn production, image deblurring plays a critical role in the vision-based detection systems to restore a sharp image and provide more accurate input for inspection. However, image deblurring is still challenging since the current methods are mainly based on the pre-defined blur degree. In dynamic yarn production, the relationship between the defocus blur degrees and the poses of the yarn body is highly associated, which can be excavated to prior knowledge in image deblurring to achieve more effective restoration. Thus, a knowledge augmented deep learning model is proposed to adaptively deblur yarn images with variable defocus blur degrees. A pose classification module designed by prior knowledge is embedded into the deep neural network, which classifies the yarn poses and feeds them into multi-scale deblurring channels. In each channel, we incorporate the image gradient prior into the specially designed loss function to attract the attention of the deblurring network on the edge details of the yarn. The experimental results from actual spinning processes demonstrate that the proposed method performs a better effect not only in the variable-scale deblurring of the global image but also in the restoration of the edge details.