A feature selection method based on a hierarchical deep network

Abstract

The invention provides a feature selection method based on a hierarchical deep network. The invention discloses a feature selection algorithm for performing selective orthogonality on depth features of different levels of a tree classifier. the features extracted by each layer of classifier better meet the requirements of respective classification tasks; the characteristic separability is improved; the influence of similarity characteristics among the categories on the network image recognition capability is effectively inhibited; and during back propagation, the knowledge graph is utilized toguide the network to update the feature selection parameters, so that the network can pay more attention to the similarity among the categories in the coarse category during rough classification, andcan pay more attention to the difference among the similar categories during fine classification. According to the method, the effectiveness and separability of the features are improved, the recognition capability of the whole network structure is improved, and the classification accuracy is improved. According to the method, a better classification effect is achieved on each data set.

Description

technical field [0001] The invention relates to the field of image recognition, in particular to a feature selection method of a deep network. Background technique [0002] With the advent of the Internet age, the scale of digital images has become larger and larger, and the number of images has not only increased exponentially, but also the types of images have also exploded. At present, research based on deep learning emerges in an endless stream, and also shows better and better performance. However, large-scale image classification is still a challenging problem. Therefore, it is very attractive to develop a new generalizable algorithm for large-scale image classification to improve the recognition accuracy. [0003] At present, the traditional deep convolutional neural network has made great achievements in the field of image recognition and is the most effective tool for image recognition. The original pixel-level features are very sensitive to the geometric transfo...

AI Technical Summary

Problems solved by technology

As the number and category of images increase, if a CNN is integrated for each classifier of each layer of the hierarchical classifier for feature extraction, the computational complexity will increase linearly

Images