Abstract: With the advancement of facial recognition technology, the identification of a single object has become highly mature. However, in some complex scenarios, applying the traditional single-face recognition technique alone struggles to meet the demands for precision and efficiency. This paper addresses the need for multi-object facial recognition in complex scenarios by designing and implementing a convolutional neural network-based multi-object facial recognition method. The core facial detection functionality is achieved through the standardized design and training of the AMTCNN (advanced multi task cascaded convolutional networks) model, which builds upon the MTCNN (multi task cascaded convolutional networks) model. During the training phase, joint learning of facial detection and alignment, along with the early stopping strategy, is employed to enhance the model's generalization performance. In the facial detection phase, a multi-scale input strategy is applied to improve the recall rate for small-sized faces. Finally, in the facial recognition phase, a deep learning library is used to extract 128-dimensional facial feature vectors for accurate identity recognition. Comprehensive comparative experiments demonstrate that AMTCNN outperforms the traditional HOG feature-based method.