Image recognition model training method, device and network and image recognition method

Abstract

The invention relates to a training method, device and network of an image recognition model and an image recognition method, and the training method comprises the steps: obtaining a prediction label value corresponding to a convolutional neural network, inputting feature maps outputted by a plurality of intermediate layers in the convolutional neural network into respective corresponding preset ViT networks for feature extraction, the method comprises the following steps: obtaining a predicted label value and a first preset loss function value corresponding to each preset ViT network, then carrying out weight and bias updating on each preset ViT network, and calculating according to the predicted label value corresponding to the convolutional neural network and the predicted label value corresponding to each preset ViT network to obtain an integrated predicted label value; and according to the integrated predicted label value, a second preset loss function corresponding to the convolutional neural network and the real label value, calculating to obtain a second preset loss function value corresponding to the convolutional neural network, and generating an image recognition model. A traditional convolutional neural network structure and a ViT network can be fused.

Description

technical field [0001] The application relates to the field of image recognition, in particular to a training method, device, network, image recognition method and equipment terminal of image recognition model. Background technology [0002] At present, the application of Vit network model (vision transformer) in computer vision to replace CNN (convolutional neural networks) is a hot spot in computer vision research. Vit network model essentially uses the visual self attention network mechanism to pay attention to the important information of each part of the picture, so as to output the corresponding prediction results. [0003] Because the above method is relatively novel, it uses many special operators that are not commonly used or have a low frequency in convolutional neural networks, and these special operators are often not well supported by mobile devices, which leads to the fact that the existing effective methods in the field of computer vision can not be directly combin...

AI Technical Summary

Problems solved by technology

[0004] In view of this, the present application provides a training method, device, network and equipment terminal of an image recognition model, which can further combine the advantages of the ViT network on the basis of utilizing the traditional convolutional neural network structure, and integrate the traditional convolutional neural network Structure and ViT network are integrated to overcome the shortcomings that the existing effective methods in the field of computer vision often cannot be directly combined with this new vision method using ViT network

Images