More trainable inception-ResNet for face recognition

In recent years, applications of face recognition have increased significantly. Despite the successful application of deep convolutional neural network (DCNN), training such networks is still a challenging task that needs a lot of experience and carefully tuning. Based on the Inception-ResNet network, we propose a novel method to mitigate the difficulty of training such deep convolutional neural network and improve its performance simultaneously. The residual scaling factor used in the Inception-ResNet module is a manually set fixed value. We believe that changing the value to a trainable parameter and initializing it to a small value can improve the stability of the model training. We further adopted a small trick of alternating the ReLU activation function with the Leaky ReLU and PReLU. The proposed model slightly increased the number of training parameters but improved training stability and performance significantly. Extensive experiments are conducted on VGGFace2, MS1MV2, IJBB and LFW datasets. The results show that the proposed trainable residual scaling factor (TRSF) and PReLU can promote the accuracy notably while stabilizing training process.