Traditional face recognition methods have poor recognition performance； deep learning-based methods face difficulty recognizing under unrestricted conditions， face features are weakly differentiated， and recognition accuracy is easily affected by pose and expression. To address this， a twin neural network model structure that introduces a Convolutional Block Attention Module （CBAM） hybrid attention mechanism is proposed. First， the algorithm structure based on the basic framework of the Siamese neural network was improved and the improved VGG11 into the framework is introduced. The BN model for feature extraction is used， which introduces batch normalization （BN） technology based on the VGG11 structure. Second， a feature extraction network incorporating a CBAM mixed attention mechanism was introduced based on the original model structure. Finally， in response to the lack of facial recognition training for Asians， the CASIA-FaceV5 dataset was employed， which is more aligned with Asian facial features， for recognition training. The experimental results show that the algorithm’s accuracy reaches 96.67% in face recognition， and the accuracy on CAS-PEAL-R1 face dataset is 6.05% and 6.7% higher than that of SRGES and VGG11+siamese algorithms， respectively. The algorithm in this study can better verify facial recognition under multiple factors， has good robustness， and greater application value.