Interbedded shale oil reservoirs exhibit strong heterogeneity and are vertically characterized by frequent alternations of thin sandstone and mudstone layers. Due to the interference of random logging noise and surrounding rock shielding effects, the original logging data of thin interbeds have a low signal-to-noise ratio and indistinct features, making it difficult to achieve fine lithology prediction. To address this challenge, this study integrates time-frequency domain signal analysis with artificial intelligence and proposes an adaptive deconvolution method for logging curve reconstruction. First, based on the wavelet transform multi-scale decomposition and reconstruction principle, high-frequency noise in the original logging curve is filtered out to suppress random noise. Then, an improved U-Net neural network is applied to perform deconvolution on the wavelet-processed logging data to reconstruct high-resolution logging curves. Finally, the model is calibrated using actual core data to obtain a high-resolution logging curve reconstruction model. The results show that the proposed logging curve reconstruction method based on time-frequency domain signal analysis and U-Net neural network achieves higher accuracy than conventional globally invariant deconvolution methods. The matching rate between the reconstructed curves and the forward-modeled logging curves from core data is improved by an average of 14.3%. Moreover, the reconstructed logging curves perform best in lithology prediction tasks. Specifically, for thin interbedded lithological assemblages, the lithology identification accuracy is improved by 27.6% compared with the original logging curves, demonstrating excellent application performance. This study presents a high-resolution logging curve reconstruction method that effectively enhances the logging response of thin beds, providing technical support for fine logging interpretation of interbedded shale reservoirs.