It is key to prevent fracture failure by investigating the failure process and revealing the evolution mechanism of internal damage. Computed tomography (CT) can provide three-dimensional characterization of the internal damage evolution process, which supports the research of the internal damage evolution mechanism of materials. However, the quantitative recognition and extraction of damage evolution faces the challenge of weak damage feature signals being overshadowed by the complex structural signals of CT images. The idea of introducing mechanical parameters to guide neural networks was proposed. Three-dimensional strain fields obtained based on Digital Volume Correlation (DVC) were used as priori information of mechanical parameters to guide and constrain network training, enabling crack identification and extraction. Through quantitative evaluation and verification of actual CT experimental data, the method can improve micro crack identification precision and reduce the identification error rate.