An Image Incremental Learning Method Based on Variational Autoencoder

Abstract

A method for image incremental learning based on variational autoencoders, comprising the following steps: 1) Constructing an encoding layer based on the AlexNet network layer structure, introducing a sampling layer and an output layer; 2) Constructing the above sampling layer and convolutional layer Based on the decoder, and add the BacthNorm layer; 3) train the encoder and decoder as an end-to-end whole on the data set; 4) separate the classification layer of the encoder and the decoder as the generation of old category data , and combined with new data for incremental learning. The invention makes the generation of anti-disturbance no longer limited by the influence of many environmental factors in practice, and has high practical value.

Description

technical field [0001] The present invention relates to an incremental learning method and a digital image processing technology, draws lessons from the idea of ​​variational auto encoders (VAEs), utilizes an improved encoder (Encoder) to classify, and the decoder (Decoder) generates as similar as possible to the input , and under the premise of maintaining the recognition accuracy of the classification layer, the incremental category data and the data generated by the decoder are jointly trained (Joint training), so as to achieve image incremental learning on the original model. Background technique [0002] With the rapid development of deep learning, although deep neural networks have shown superior performance in various fields such as image classification, semantic segmentation, object detection, speech recognition, medical image analysis, etc. The tasks train their own independent models. In order to improve the effect on multiple data sets at the same time and adapt ...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the problem that as time goes by, the amount of data continues to increase, the demand for space storage increases rapidly, and the final learning speed will lag behind the data update, and to improve the utilization value of the classifier in the actual scene of big data, the present invention proposes the following The variational autoencoder is used as a prototype, and the improved encoder is used as a classifier, and the decoder is used as a generator to generate old category data, and the newly added category and the generated old category data are jointly trained to achieve the purpose of incremental learning

The invention overcomes the shortcomings of traditional learning methods, effectively trains a classifier from increasing new data, and does not affect the recognition accuracy of old categories, solves the problem of time and space requirements, and adapts to the needs of practical application scenarios. It has important research and application value in the field of artificial intelligence

Images