Enhance Feature Representation of Dual Networks for Attribute Prediction

Traditional single branch CNN could not extract all the details of the input, which may lose some vital information, resulting in a decrease in recognition accuracy. In this paper, we propose a novel dual branch adversarial neural network named D-BANN. Inspired by adversarial learning, we drive parallel networks to extract complementary features and adopt a novel loss function to extend the application domain of the model. Moreover, we divide the network training procedure into multi-steps to alternatively optimize the loss functions. In order to evaluate the proposed method, we carry out comprehensive experiments on three attribute datasets. The results on facial attributes demonstrate that the proposed method can outperform other single task networks in face attribute recognition. Also, D-BANN achieves competitive results in two pedestrian datasets compared to the state-of-the-art multi-task methods. We visualize the D-BANN using Grad-CAM to verify the effectiveness of feature annotation.