An Active Crowdsourcing Image Learning Method Based on Semi-Supervised Variational Autoencoders

Abstract

The invention discloses an active crowdsourcing image learning method based on a semi-supervised variational autoencoder, specifically: acquiring a picture set, randomly selecting a certain proportion of images, distributing them to the annotators on the network for annotating, and obtaining Crowdsourcing labeling; building a crowdsourcing learning network model based on a semi-supervised variational autoencoder; inputting a dataset into the model to construct a loss function; training the model end-to-end based on stochastic gradient descent; selecting the entropy of predicting the true label The largest image, query the real mark; combine with the training set used in the previous round of iteration to generate a new data set, return to the third step until the current number of iterations reaches the threshold; delete the model reconstruction part and the crowdsourcing mapping layer part, and use the rest of the network as a classifier. The invention utilizes crowdsourcing data and unmarked data at the same time to reduce the cost of data labeling. And by introducing a small number of real markers, the overfitting of the model to noise can be alleviated and the generalization performance of the model can be improved.

Description

technical field [0001] The invention relates to an active crowdsourcing image learning method based on a semi-supervised variational autoencoder, and belongs to the technical field of image labeling. Background technique [0002] Traditional supervised learning requires a large number of expert images and requires relatively high labeling costs. To reduce the cost, various methods have been proposed, such as crowdsourcing learning, semi-supervised learning, and active learning. [0003] The process of crowdsourcing learning is to first publish the labeling task on the network platform. Anyone can perform this labeling task, and the collected data has a higher proportion of noise than expert labeling. In order to solve the uncertainty of non-expert annotators in crowdsourcing, each image is usually annotated by multiple annotators, that is, repeated annotations. How to utilize this crowdsourcing data with noise and repeated labels has become the key point of crowdsourcing l...

AI Technical Summary

Problems solved by technology

However, in actual situations, the real labeling of a small number of images will not increase too much labeling cost, but it is likely to greatly improve the generalization performance of the model

Images