Long-tail distributions are an inherent flaw in scene graph task-related datasets, where the uneven number of relationship samples leads to a model biased toward predicting high-frequency relationships. The debiasing technique, by improving the model structure, not only produces a complex model structure but also increases the number of model parameters. To address this problem, we propose a model-independent scene graph debiasing algorithm based on probabilistic correlation, comprising two modules: correlation prediction and weight fusion. First, the biased predictions of the algorithm are statistically analyzed to obtain the correlation confusion matrix between the predicted values and the actual values; second, the debiasing method based on the correlation confusion matrix is designed; and finally, the original predictions of the fusion are corrected to correct the biased results of the model. Our experiments on the Visual Genome dataset demonstrate that the method presented in this paper exhibits higher debiasing ability than similar algorithms and stronger fine-grained relationship prediction capabilities.