Accurate assessment of the metallographic grain size grade of steel can detect material deterioration and ensure the safety of equipment in service.In order to solve the problems that the traditional manual evaluation of steel metallographic grain size grade is time-consuming and easily influenced by manual experience,the evaluation results are not consistent and irreducible,etc,a deep learning-based steel metallographic grain size grade evaluation method is proposed.By adding a jump connection layer to the U-net model and reducing the number of downsampling to improve the segmentation accuracy and reduce the number of network parameters,the pixel accuracy is 93.86% and the mean pixel accuracy (MPA) is 86.89% on the 117 validation sets.The number of network parameters is only 2.02 M.The grain boundary prediction results are digitally processed and combined with the intercept point method to grade the grain size,and the average time taken to grade the grain size of steel on the test image is only 8.3 s/sheet.Compared with manual rating methods,this method is accurate,efficient and repeatable.