Efficiency of binary neural networks for object detection on an image

Deep convolutional neural networks are widely used for object detection. However, modern deep convolutional neural network models are computationally expensive hindering their deployment in resource-constrained mobile and embedded devices. Binary neural networks help to alleviate the resource requirements of devices. Activations and weights in binary neural networks are limited by binary values $(-1, 1)$. The proposed method implements balancing and standardization of floating-point weights in forward propagation and two-stage sign function approximation in back propagation. The paper presents the results of detection accuracy on the PASCAL Face dataset as well as the results of speed and model size on the mobile device for the proposed method, the model without binarization, the TinyML network, and Bi-Real Net and ABC-Net methods.