Method, system and device for collaborative target recognition based on decorrelated binary network

Abstract

The present application provides a collaborative target recognition method, system and device based on a decorrelation binary network. In this application, by minimizing the target quantization error in the training process of the binary network model, the minimized target quantization error is decomposed into maximum likelihood estimation and maximum a posteriori estimation based on Bayesian learning to determine the activation tensor Value error Lc, weight tensor binarization error Ld, in order to realize that in addition to considering the conventional prediction error Ls when training the binary network model, Lc and Ld are further considered to optimize the noise influence in the process of neural network binarization , to avoid affecting the training of the binary network model due to the disappearance or non-existence of the gradient of the parameters during the model training process, to ensure that the final trained binary network model is more stable and to complete the model convergence better, and to further improve the application of the binary network model For example, the accuracy of recognized target objects (such as human faces) and / or abnormal behaviors (such as arson, robbery).

Description

technical field [0001] The present application relates to Internet of Things technology, in particular to a method, system and device for collaborative target recognition based on decorrelation binary networks. Background technique [0002] Binary Neural Network (BNN: Binary Neural Network), also known as binary network, is used to reduce the network size and speed up network training. In the application, the weight tensor and activation tensor of each layer network contained in the binary network will be converted into binary values, such as +1 or -1. The core of the binary network is to effectively reduce the multiplication and addition operations and reduce the storage space of the weights while maintaining a small decrease in accuracy, so as to provide effective feasibility for the deployment of the mobile terminal. [0003] However, since the weight tensor and activation tensor of each layer network in the binary network are simply scaled to binary values, such as +1 o...

AI Technical Summary

Problems solved by technology

[0003] However, since the weight tensor and activation tensor of each layer network in the binary network are simply scaled to binary values, such as +1 or -1, this simple scaling method will cause parameters in the binary network such as weight tensors to Gradients for quantities and activation tensors etc. do not exist (disappear)

The disappearance or non-existence of the gradient of the parameters will affect the training of the binary network model, which will lead to the application of the final trained binary network model, such as the recognized target object (such as a human face) and / or abnormal action behavior (such as arson , robbery, fighting, stealing) etc. are not accurate

Images