In the study of biometrics recognition, the fusion experiment of soft biometrics has always adopted the fusion model between the main features. Soft biometrics are different from the main features in terms of application scope, recognition ability, etc. The fusion mechanism between the main features cannot effectively mine and utilize the discrimination information in the soft biometrics, resulting in less ideal and accurate experimental results. Based on this, this paper deeply analyzes the characteristics of soft biometrics, and puts forward the requirements of ‘effective complementarity’ and ‘personalized fusion’. Based on the quantified effective complementarity, and combined with the goal of ‘minimum error rate’, this paper establishes a local fusion model for each user using the least squares method. The proposed fusion model is applied in face and fingerprint recognition experimental scenarios to verify the rationality of the analysis and the effectiveness of the proposed method in promoting the recognition performance.