Artificial Neural Networks (ANNs) has been applied in the face detection task because of its ability to capture the complex probability distribution conditioned to the class of face patterns. However, many works use Back-Propagation (BP) to adapt the weights of the ANNs. The problem of using BP is that it has many disadvantages related to the appropriate choice of its parameters, as the learning rate and momentum. Furthermore, since BP assumes a fixed architecture for the ANN, an inappropriate choice of the architecture can make it have a sub-optimal performance. In this paper we investigate the application of the IPSONet in the facial detection task. IPSONet is a training technique for neural networks like multilayer perceptron (MLP) that uses an improved PSO to evolve simultaneously structure and weights of ANNs. Thus, the IPSONet produces ANNs with higher generalization ability if compared to BP. The system developed in this work, which includes the feature extraction process of the input image and the training of a MLP net using IPSONet is called IPSONetFD. The experiments using the MIT CBCL Face Database showed that the proposed technique is robust in the sense that it can detect faces with a wide variety of pose, lighting and face expression. The results showed that the IPSONetFD had better performance than others ANN's architectures (PyraNet and I-PyraNet, in this study), and an equivalent performance if compared to SVM. Thus, the proposed technique demonstrated that ANNs trained by IPSONet has better performance than ANNs trained by BP in the face detection task.