Providing a hybrid method for face detection and gender recognition by a transfer learning and fine-tuning approach in deep convolutional neural networks and the Yolo algorithm

The present study aims to assess a combined method for face detection and gender recognition using deep learning by evaluating the shortcomings of face detection methods accurately. Deep learning algorithms can learn high-level features and have attracted a lot of attention for use in the field of machine vision. The present study names a hybrid method called Hyper-Yolo-face and utilizes a clear image using deep Convolution Neural Networks (CNNs), Yolo algorithm, and local binary patterns (LBPs) to identify the face and recognize the gender. Reducing the number of parameters is regarded as an extremely important challenge in deep networks in terms of memory consumption and the amount of computing in the network. The proposed method is based on the AlexNet model and generalization in the loss function of version 3 of the Yolo algorithm, which leads to improved precision. The present study focuses on applying small filters in transfer learning and fine-tuning network layers and using a new regression loss function in the Yolo algorithm to make it more appropriate for multiscale face detection. The face images are detected and cut by the presented Yolo in the proposed method. Then, an LBP operator is applied so that richer information and images enter the AlexNet network to estimate other parameters including gender recognition. Based on the experiments on the AFLW, FDDB, and PASCAL datasets, the proposed method improves recognition precision significantly.