Image incremental learning method based on dynamic correction vector

Abstract

The invention discloses an image incremental learning method based on a dynamic correction vector. The method comprises the following steps: S1, a backbone network with a ResNet-32 network layer structure as a model is constructed, an Adam training optimizer is adopted, and meanwhile, a basic classification loss function is a Kullback-Leibler Diverge relative entropy loss function; s2, knowledge distillation is introduced into a loss function, a classification loss function is combined, a new model is helped to learn knowledge in an old category, and the catastrophic forgetting problem is relieved; s3, a ResNet-32 model is trained in a training mode by adopting a representative memory method and a dynamic correction vector method; and S4, the optimal model trained in the previous incremental stage is reloaded for repeating the steps S2 to S3, and the performance on all test sets is evaluated until all incremental data is trained. According to the method, the incremental learning task recognition capability is improved, and the method has high practical value.

Description

technical field [0001] The present invention relates to Knowledge Distillation technology and Representative Memory method. By using the technology of Dynamic Correction Vector, while maintaining the accuracy of classification and recognition of old categories, the accuracy of new category data is improved at the same time. Classification accuracy, so as to realize the incremental learning recognition task on the original data set. Background technique [0002] In recent years, Deep Convolutional Neural Networks (DCNNs) have been widely used in various fields of detection, segmentation, object recognition and images. Despite the success of Convolutional Neural Networks, it was in the ImageNet competition that computer vision and machine learning teams began to pay attention. In 2012, AlexNet easily won the ImageNet large-scale visual recognition challenge by implementing Deep-CNN and pushing DCNNs into the public eye. (ISLVRC). Since then, DCNNs have dominated ISLVRC and ...

AI Technical Summary

Problems solved by technology

[0006] In order to solve the problem that the depth model trains dynamically changing data sets in practical application scenarios, reduce the dependence on distributed computing systems, and save a lot of computing overhead and system memory, the present invention proposes a 32-layer residual network ResNet- 32, through the introduction of knowledge distillation technology and representative memory method, using the technique of dynamic correction vector, the problem of catastrophic forgetting is alleviated, and the performance of incremental learning is improved

Images