The computed tomography (CT) scanning test during the destruction of coal rock specimens can achieve visual characterization of the internal fracture structure of coal rock specimens. However, during the loading process, the displacement and deformation of the specimens will cause positioning errors in the CT scanning layers, which will affect the accuracy of the test results. To solve this problem, the image similarity retrieval technique was utilized to retrieve the most similar CT images of the loaded specimens under different stress levels, to more accurately describe the fracture evolution law of the loaded coal samples. The results show that during the triaxial compressive damage process, the stress-strain curve of the specimen is divided into five distinct phases, the initial compression-density phase, the linear-elastic deformation phase, the plastic yielding phase, the peak damage phase, and the residual deformation phase. The image retrieval technique based on the perceptual Hash algorithm can accurately identify and locate CT images with the highest similarity in different stress stages, and CT image similarity decreases with increasing axial load. Throughout the triaxial loading process, the mean absolute difference in similarity of CT images shows two distinct stages of change: a slowly increasing stage and a rapidly increasing stage. The fractal dimension can quantitatively describe fracture evolution, and the fractal dimension mainly experienced three stages of slow decrease, slow increase, and rapid increase during the damage process of loaded samples.