Abstract: Quantitative inversion of fracture weakness, one of the key elastic parameters for anisotropic, naturally fractured reservoirs, is geologically important to the characterization of fracture density, connectivity, and fluids. Traditional analytical inversion methods based on anisotropic rock physics models often rely on simplified assumptions, e.g., homogeneous fracture distribution and weak anisotropy approximations. These simplifications struggle to accurately describe the seismic response mechanisms of complex fracture systems, a shortcoming that becomes particularly pronounced under strong anisotropy or heterogeneous fracture distribution, leading to substantial inversion errors and solution non-uniqueness. To this end, leveraging the advantages of convolutional neural networks in local feature extraction and spatial correlation modeling, this paper constructs an end-to-end deep learning model. By introducing geophysical initial models as prior geological constraints, a fractured parameter prediction model with multi-source data-driven capability is further established. Aiming at the non-uniqueness of geophysical inversion and the limitations of the deterministic output of neural networks, the approximate Bayesian computation method is incorporated. Through constructing the posterior probability distribution of model parameters, the uncertainty in the quantitative inversion of fracture intensity is quantified. Both synthetic model tests and real seismic data processing results indicate that this method can significantly improve inversion accuracy and reduce the confidence interval by more than 30%. Further application verification has confirmed that it demonstrates higher reliability and resolution in the quantitative characterization of parameters for complex fractured reservoirs, thus meeting the requirements of practical geological evaluation.