Early fatigue crack detection and remaining life evaluation are essential for preventive maintenance of steel structures. However, the efficiency and accuracy of the tiny and initial crack detection are limited due to the time-consuming manual processes. This study develops a novel computer vision-based framework for simultaneous real-time fatigue crack detection and prediction. Linear elastic fracture mechanics (LEFM) is the theoretical basis of the proposed framework. A DeepLabV3+ model was used to automatically extract geometric features of cracks from visual images for high-precision crack identification and quantification. A physics-informed neural network (BPINN) was constructed with prior intervals of physical parameters. Prior knowledge of LEFM was embedded into the deep learning framework to establish a nonlinear relationship between crack length and fatigue cycle count. The fatigue cycle gradient with respect to crack length was computed using automatic differentiation. A composite loss function was then designed by integrating data error, physical residuals, and parameter prior constraints. In addition, subnetworks were employed to adaptively update physical parameters. The representation of individual structural characteristics was improved by quantifying parameter uncertainty in fatigue crack growth life (FCGL) predictions, which ensures physical consistency. Finally, the effectiveness of this method was validated through two case studies. The analytical results show that the FCGL predictions and crack detection results align well with the experimental findings. The proposed data-physics fusion approach demonstrates superior performance in FCGL evaluations, while the identification of individualized uncertainty parameters refines the accuracy of lifetime predictions.