Scalable modular image recognition method based on generative adversarial network

Abstract

The invention discloses a scalable modular image recognition method based on a generative adversarial network. The method comprises the following steps: 1, sorting image data according to types, dividing the image data into N types if there are N types, and enabling each type to only contain the same type of image data; 2, preparing N generative adversarial network models according to the type number N; 3, training the image data of the ith type, wherein i belongs to {1, 2,..., N}; training a generative adversarial network Gi until the similarity between the image data generated by the generative adversarial network and original data set image data reaches a preset value; 4, fixing parameters of a discriminator Di in the generative adversarial network, and training a generator Gi accordingto a gradient inverse direction; 5, fixing parameters of the generator Gi in the generative adversarial network, and training the discriminator Di until training is completed; 6, repeating step 3 tostep 5 on all types of image data until the training of the N discriminators is completed; 7, arranging and combining the N discriminators in parallel to form a discriminator group; 9, inputting the pictures into a discriminator group, wherein each discriminator outputs a prediction result to the pictures; and 10, calculating final prediction results.

Description

technical field [0001] The invention relates to a scalable modular image recognition method based on a generative confrontation network, which is used for image recognition tasks of various scales and quantities, and specifically belongs to the technical field of image recognition and classification. Background technique [0002] With the development of computer information technology and the rise of deep learning, more and more image classification and recognition tasks in the field of computer vision are processed using deep learning. Among them, the deep neural network is the most widely used method for processing image classification and recognition tasks. But we still face many problems when using deep neural networks to deal with practical tasks in the industry. Currently, the commonly used recognition methods for image data are deep neural networks and support vector machines. [0003] 1. Deep learning. Deep learning is a new research direction in the field of mach...

AI Technical Summary

Problems solved by technology

But this will cause the following problems: 1) The model is very complex, it is difficult to modify it according to the current task, and requires researchers to have rich experience and skills, which is difficult to meet in the actual environment.

2) For tasks of different scales, a single model cannot be adjusted for the task scale, resulting in waste of resources

3) The deep neural network cannot identify negative sample data

In solving image classification problems, another common practice is to use multiple support vector machines, each of which is responsible for the identification of a classification, but the same support vector machines have the following disadvantages: 1) support vector machines are not suitable for large-scale training samples Difficult to implement

Since the support vector machine uses quadratic programming to solve the support vector, and solving the quadratic programming will involve the calculation of the m-order matrix (m is the number of samples), when the number of m is large, the storage and calculation of the matrix will consume a lot machine memory and computing time

2) There are difficulties in solving multi-classification problems with support vector machines

Images