Human identity will be recognized by measuring and analyzing human characteristics and it will be assumed as face recognition. In this technique, face recognition will be performed by examining the facial image through the computer using the database. The face information of identified people is recorded in this database. Identifying criminals, credit cards, security system are among the applications of gender detection. This research aims to increase the accuracy of gender recognition from facial images and reduce false positive rate. The innovative aspect of the research, we used the ResNet-50 pre-trained convolutional neural network in order to extract the optimal features from the images. This neural network extracted 1000 feature vector for each image using 50 convolution layers. In the classification stage, by removing the fully connected layers in the convolutional neural network, the learning was transferred to a multi-layer perceptron neural network. In general, it can be said that using the transfer learning method allows us to benefit from the advantages of convolutional neural network and traditional neural networks at the same time. In the classification stage, by removing the fully connected layers in the convolutional neural network, the learning was transferred to an optimized multi-layer perceptron neural network utilizing the Firefly algorithm. In this study, in order to adjust the weight and bias of the Perceptron neural network, the Firefly algorithm was used. In general, it can be said that using the transfer learning method allows us to benefit from the advantages of convolutional neural network and traditional neural networks at the same time. Finally, using the proposed method, we were able to achieve 97% accuracy, which was an improvement compared to other existing methods.
