Abstract: Various defect structures exist in coal and rock masses, including cracks and pores, which significantly affect the occurrence and migration of coalbed methane. Additionally, the instability and destruction of coal rock structures are closely related to the changes and expansion of coal rock pore and crack structures, leading to mining dynamic disasters. Therefore, identifying coal rock cracks is of great practical significance for studying the migration law of coal seam gas and the stability of the surrounding rock. The models proposed in recent studies for improving the identification accuracy of internal cracks in coal and rock masses still have the following shortcomings. (1) The impact of noise interference in CT (Computed tomography) images on model performance is often overlooked. (2) The potential loss of small-crack features during downsampling and the insufficient fusion of multilayer features during upsampling have not received sufficient attention. (3) The differences within the crack categories were not fully considered. To address the challenge of accurately identifying microcracks in CT scan images of coal and rock, this study proposes a crack segmentation model based on explicit visual prompting for coal and rock CT images (EViP-CTCrack) by reorganizing the convolution kernel, down-sampling and up-sampling processes, and classifier functions. This model effectively filters out noise interference and enhances its ability to capture the key details of cracks by introducing an explicit visual cue generator. The segmentation accuracy of the model was improved by improving the downsampling and upsampling strategies, preserving small-crack features, and enhancing multiscale feature fusion capabilities. Using a multi-representative vector classifier, the internal differences in the crack categories can be fully described to improve the generalization ability of the model. The main results show that (1) the crack segmentation model EViP-CTCrack mainly consists of a residual mixed-connection convolution module, a cross-attention upsampling module, a multi-representative vector classifier, and an explicit visual prompt generator module, which can effectively improve the shortcomings of existing models in terms of noise interference, feature loss, and crack category difference description, significantly enhancing the robustness and generalization ability of the model. (2) This study proposes a high-quality coal and rock CT image dataset, CTRock, and conducts a comparison and ablation study on CTRock to verify the effectiveness of the targeted design. The algorithm is validated using the proposed CTRock dataset. The average intersection-to-union ratio and accuracy of EViP-CTCrack on the CTRock dataset were 88.1% and 94.4%, respectively, indicating good crack segmentation results. (3) After CT scanning the rock samples collected from the Xinjie Coal Mine, EViP-CTCrack was used to identify cracks in all layers of the reconstructed data slices. A porosity compressive strength equation was established that can quickly calculate the uniaxial compressive strength and provide new ideas for studying the physical and mechanical properties of coal rock.