Baseline deviation often occurs with the spectrum data acquisition, making the subsequent identification and analysis results deviate from the true values. Therefore, it is necessary to utilize the baseline correction technology to obtain more accurate spectrum data before the spectrum data analysis. The sparse Bayesian learning (SBL)-based baseline correction method can provide the optimal baseline correction results within the Bayesian framework, and it does not need to select parameters manually. However, the SBL framework is too simple to apply to complex sparsity structures. In practical implementations, if the peak of the pure spectrum is wide, the corresponding sparse representation vector would exhibit a block-sparsity property. The performance of the SBL method will be further improved if the additional block-sparse structure can be exploited appropriately. To this end, we introduce a coupling pattern model into the SBL framework to adaptively learn the block-sparse structure. Due to the inherent learning capability of the SBL framework, the proposed method can significantly improve the baseline correction performance. We conducted several simulations to evaluate the performance improvement, where the proposed method is compared to SSFBCSP and SBL-BC with different noise variances. The simulation results verify the superiority of the proposed method for wide peak recovery. Specifically, it has good stability when the noise level is high, but the performance of other methods degrades substantially. Monte Carlo simulation results further demonstrate that our method can significantly improve the pure spectrum fitting’s normalized mean square error (NMSE) performance. Finally, one real chromatogram dataset and three Raman datasets are used to validate the performance of the proposed method. The experimental results indicate that our method can produce smoother pure spectrum fitting and better denoising effects than others.