To address the low efficiency and accuracy of manual testing in railway computer interlocking systems, this study proposes a deep learning-based method for text localization and recognition in interlocking interface images. First, a text localization model based on the connectionist text proposal network (CTPN) is developed. By comparing multiple backbone networks (ResNet50, AlexNet, ZF and VGG16), VGG16 is selected as the feature extractor to enhance high-level semantic representation and improve the detection of small text regions. Second, the generalization ability and robustness of the CTPN model are improved through performance comparison with common object detection models and the incorporation of dropout. A projection-based segmentation method, combining horizontal and vertical projections, is further employed to address text adhesion issues in the interface. Finally, an improved AlexNet model is used for text recognition. Experimental results on a railway interlocking interface dataset in the TensorFlow environment show that the proposed method achieves a localization accuracy of 87.98%, a recall of 73.33%, and an F-score of 80.39%, while the recognition accuracy reaches 89%. These results demonstrate that the proposed approach can effectively locate and recognize interface text, providing reliable data support for automated routing and test result analysis in interlocking system testing.