Classification optimization method and device

Abstract

An embodiment of the present application provides a classification optimization method and device. The method includes initializing a pre-constructed neural network, constructing a parameter matrix, and orthogonally initializing parameters of a classification layer in the neural network by using the parameter matrix; importing a training image into the neural network to obtain a feature vector outputted after the training image passes through the network layer; according to the parameter matrix and eigenvector, obtaining the loss function; according to the loss function, performing gradient calculation on the network layer weights corresponding to each network layer and the classification layer weights corresponding to the classification layer to update the network layer weights and the classification layer weights; and training the input image according to the neural network which updates the weights of the network layer and the classification layer. The classification optimization scheme adds the constraint of orthogonality between the class parameters, which makes the different classes relatively independent, and makes the neural network model more discriminant.

Description

technical field [0001] The present invention relates to the technical field of image recognition, in particular to a classification optimization method and device. Background technique [0002] The multi-classification problem is one of the common problems in practical application scenarios. In deep neural networks, Softmax is the most commonly used loss function. This is because Softmax has the characteristics of simple structure and efficient gradient operation. However, many studies have shown that when only the Softmax loss function is used to supervise the network model for training, the obtained features are not sufficiently discriminative. To be precise, the network's classification of images is prone to errors when the images are relatively similar. This is mainly because Softmax only optimizes the inter-class distance, and does not shrink the distance of samples within the class. [0003] In the existing schemes, there are improved methods based on Softmax and C...

AI Technical Summary

Problems solved by technology

However, the improved method based on Softmax loss mainly focuses on the additive distance, such as AM-Softmax, but neither constrains the category center, but dynamically learns with the model training

Although the multi-group method restricts the distance between samples, it is difficult to generate high-quality sample pairs, and the gradient update operation is more complicated.

Images