Open set category mining and extending method based on depth neural network and device thereof

Abstract

The invention discloses a sample categorization method based on a depth neural network. The open set category mining and extending method based on a depth neural network comprises steps of using a sample set comprising defined category samples to train a categorized model to be extended, obtaining categorization threshold value information, sending a sample set comprising undefined category samples into the categorization model to be extended, determining at least part of the undefined category samples according to the categorization threshold value information of the categorization model to be extended, artificially marking the undefined category samples, adding a number of columns of a weight transfer matrix in a categorization layer of the depth nerve network in order to increase a total number of model recognition categories, wherein the added weight columns comprise first information associated with global categorization and second information associated with connection between categories and using the undefined category samples which are artificially marked to increase the models which already finish training and updating. The open set category mining and extending method based on the depth neural network and the device thereof extend the depth neural network through modifying a depth neural network categorization layer weight transfer matrix, dynamically increases the number of the recognized categories so as to process the open set recognition problem and can be applied to a scene which is closer to a real scene.

Description

technical field [0001] The present invention relates to the field of deep learning, in particular to a deep neural network-based method and device for discovering and expanding open-set categories. Background technique [0002] Deep neural networks have achieved remarkable results on many visual recognition problems and have led to many high-impact academic research and successful commercial applications. Some recent studies on image classification and human action recognition problems have shown excellent performance. However, most recognition systems are designed for a static closed world, based on the assumption that the recognition category is a priori knowledge. However, in the real world, the recognition scenarios are ever-changing, and thousands of different recognition scenarios cover countless recognition categories. Even in a specific scenario, we can define a specific recognition category, and events such as anomalies will inevitably occur. In addition, the tra...

AI Technical Summary

Problems solved by technology

That is to say, in this technical solution, the self-encoder is used to extract features first, and then the classifier is used to classify. However, only the newly marked samples and the marked samples are used to retrain the entire autoencoder for the entire incremental process. However, there are many disadvantages. First, such an incremental process requires a large number of labeled samples of new categories. Second, the process of re-updating the entire model is cumbersome and expensive.

Images