Tree network method for testing and updating through WordNet embedding

Abstract

The invention provides a tree network method for testing and updating through WordNet embedding and relates to the intelligent classification of pictures. The method comprises a step of building a tree network, a step of carrying out pre-training, a step of carrying out dynamic pruning based on an SVM, dynamically pruning a node with a too low activation value of characteristic mapping and a subtree of the node in a test process, wherein if the activation value of the node is too low, the probability of a class represented by the node is low, so the probability of a descendant node is low andnegligible, a step of carrying out the acceleration of a model in the test process of using the SVM since the sum of the activation values of characteristic maps has strong linear separability, anda step of carrying out online update based on a branch and transmitting a sample with a high prediction probability back to the tree network for training with a detected image as a training sample.

Description

technical field [0001] The invention relates to intelligent classification of pictures, in particular to a tree network method for testing and updating through WordNet embedding. Background technique [0002] Deep convolutional neural networks have led to a series of breakthroughs in several computer vision tasks, such as image classification [1-5] ,Target Detection [6-7] and semantic segmentation [8-10] Wait. Deep convolutional neural networks bring many powerful advantages: low-dimensional / medium-dimensional / high-dimensional feature integration [11] ; end-to-end training and increased accuracy as the number of layers increases. Much recent work has focused on stacking network depths for better accuracy, but leads to increasingly poor interpretability. While deeper and deeper networks achieve excellent evaluation metrics, it brings more forward propagation time and poorer interpretability. [0003] In addition, decision trees have excellent interpretability and fast t...

AI Technical Summary

Problems solved by technology

However, due to its small model space, it does not have strong generalization ability like neural networks when dealing with natural language processing and computer vision problems.

Images