Image classification method and system of neural network based on gradient direction parameter optimization

Abstract

The invention discloses an image classification method and system of a neural network based on gradient direction parameter optimization, and the method comprises the steps: obtaining an image data set, and carrying out the preprocessing of each image in the image data set; constructing a convolutional neural network, and setting a hyper-parameter and a loss function of the convolutional neural network; taking the preprocessed image and a known image classification label as input values of a convolutional neural network, and starting to train the convolutional neural network; calculating a loss function according to the output of the convolutional neural network and the image label, and then carrying out error back propagation; in an error back propagation process, adopting a gradient direction parameter optimization mode to realize parameter updating in a convolutional neural network training process, and after an end condition is satisfied, ending iterative training to obtain a trained convolutional neural network; and based on the trained convolutional neural network, classifying the to-be-classified image to obtain a classification label of the to-be-classified image. The method has the advantages of high image classification speed and high stability.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image classification method and system based on a gradient direction parameter optimized neural network. Background technique [0002] The statements in this section merely provide background related to the present disclosure and do not necessarily constitute prior art. [0003] In deep learning, multi-classification problems have always been a hot research field, and the most common multi-classification problem is image classification. When dealing with some image classification problems in deep learning, neural network models are required to be trained at any time. For example, in the face recognition access control system, when a new person enters the access control system, the face recognition neural network model must continuously train the face pictures of the newly added person. Therefore, accelerating the training speed of the neural network model is very im...

AI Technical Summary

Problems solved by technology

But the disadvantage of this basic gradient descent method is that the learning rate is a hyperparameter, and there is almost no theoretical basis for its setting. Basically, the empirical value of the frequently used learning rate is obtained by experiment. The algorithm does not converge; if the learning rate is set to a small value, the algorithm converges very slowly, and even stops before reaching the optimum point. In the actual deep learning training, it takes a lot of time for parameter tuning

However, Newton's method has high requirements for the objective function to be optimized, and requires the Hessian matrix to be positive definite. At the same time, the calculation is too complicated, which makes the amount of calculation and storage in the process of optimizing the parameters of the neural network larger, which leads to the real In solving the actual image multi-classification problem, the Newton method, including the later improved quasi-Newton method, is not faster and more convenient than the gradient descent method

Newton's method has a fast convergence speed in theory, but the main reason why it is not as good as the gradient descent method in the process of image multi-classification is that more gradient information is used, which increases the complexity of actual calculation and increases the amount of calculation.

Images