Abstract: Due to the small size of the tool and the difficulty of precise positioning in the micro-milling process,the resolution of the tool image obtained by the CCD camera is difficult to meet the needs of high-precision wear monitoring.A super-resolution reconstruction network based on deep learning was proposed for low resolution tool images obtained from machining sites.The network used a cross residual structure and an efficient attention module to learn the mapping relationship between high and low resolution image pairs.To enhance the texture detail reconstruction effect of tool images,a loss function based on image gradient information was proposed for network training.Compared with bicubic interpolation and typical super-resolution networks,the proposed network had a clearer texture of reconstructed tool images,significantly improved peak signal-to-noise ratio and structural similarity index,and reduced the error of wear estimation while maintaining excellent reconstruction speed.

Key words: micro-milling, deep learning, super-resolution, loss function, tool monitoring

摘要： 在微铣削加工过程中,由于刀具尺寸较小和难以精确定位的特点,导致通过CCD相机获得的刀具图像的分辨率难以应对高精度磨损监测的需求。针对加工现场获得的低分辨率刀具图像,提出一种基于深度学习的单幅图像超分辨率重建网络。该网络采用交叉残差结构与高效注意力模块,学习高低分辨率图像对之间的映射关系。为增强刀具图像纹理细节重建效果,提出一种融合图像梯度信息的损失函数用于网络训练。与双三次插值以及典型超分辨率网络相比,所提出的网络在保持优异重建速度的前提下,重建的刀具图像纹理更加清晰,峰值信噪比与结构相似性指标得到显著提升,进行磨损值估计的误差显著降低。

关键词: 微铣削, 深度学习, 超分辨率, 损失函数, 刀具监测