<p>Demographic characteristics influence the accuracy of facial recognition systems. Currently, practice has shown that researchers are increasingly interested in the problem of soft biometric recognition based on a person's face. However, the impact of demographic information, such as ethnicity, has not been sufficiently studied, and the results have been quite contradictory. The main goal is to develop a model to most accurately identify ethnicity. We combine transfer learning and extreme learning machine methodologies into an ensemble approach based on artificial neural network rules for the multi-classification of face images. The first part of the approach involves feature extraction based on a multi-branch model consisting of MobileNet and Inception combination (DeepMultiMI), and the second part considers extreme learning machine for classification into multiple ethnic groups. The output values of the models are combined into the final recognition result. The proposed approach is evaluated on two large datasets: UTKFace and FairFace. The performance of DeepMultiMI is compared with well-known machine learning methods for classifying facial images into multiple ethnic groups. Experimental results prove its high competitiveness and its applicability to demographic issues.</p>