The microscopic morphology recognition of rock fractures plays a crucial role in revealing the fracture failure mechanism of rocks. Deep learning has shown significant success in image recognition, but it still faces issues such as strong subjectivity and complex features in rock fracture image recognition. To improve the accuracy of rock fracture image recognition, an improved InceptionV3 model was proposed. This model integrated Self-Attention and Bilinear Attention Mechanism to extract both local and global features, and combined feature concatenation fusion technology to enhance the model's ability to extract effective features, enabling rock fracture image recognition. The rock fracture dataset was used for analysis. The results show that the improved InceptionV3 model achieves an accuracy of 98.27% in sandstone fracture image recognition. The model is validated using a validation set, and the Precision, Recall, and F1-Score all reach 85%, confirming the model's efficiency and reliability. The research results provide a new solution for the automated and high-precision recognition of rock fracture images, which is beneficial for determining rock fracture failure and has important guiding significance for the prevention of geological engineering disasters.