<p>Surface damage in single crystal superalloys during the machining process severely affects their high-performance capabilities. The complexity of the boundary information of surface damage makes it difficult to quantitatively evaluate the machining quality. To achieve efficient and accurate damage assessment and prediction, this study proposes a convolutional neural network based image characterization method for recrystallization damage, combining plastic deformation to predict recrystallization damage morphology under specific machining conditions. Through grinding and scratching experiments, two datasets containing over 2000 surface damage scanning electron microscope images are collected and created. The encoder-decoder neural network is used to achieve adaptive quantitative segmentation of surface recrystallization damage morphology. The optimal model reached a quantitative accuracy of 89.87%, and the error in quantitative parameters such as damage area and maximum depth, compared to manual measurements, ranged from 3 to 10%, enabling direct and accurate evaluation. By quantifying the damage morphology, the variation trends of geometric features in the two regions under the same machining conditions are established, enabling the prediction of geometric morphology from plastic deformation to recrystallization damage. Under high scratching force, the recrystallization area of the material is 20% larger than the plastic deformation area, while the maximum depth remains consistent. Under flexible grinding conditions, the average depth of recrystallization damage is more than 50% greater than that of plastic deformation. This deep learning-based image characterization method facilitates precise assessment and rapid prediction of recrystallization damage in single crystal superalloys machining.</p>