Semi-supervised deep network picture classification method based on swarm intelligence

Abstract

The invention discloses a semi-supervised deep network picture classification method based on swarm intelligence. The semi-supervised deep network picture classification method comprises the followingsteps of: 1), preprocessing a training data set and a testing data set of a digital picture, and carrying out normalization and centralization processing; 2) calculating network loss by using the training data set, specifically, 2-1) for marked data, calculating hybrid KL divergence loss between a predicted value and a real label value of the network, and 2-2) for unmarked data, calculating groupconsistency loss among a plurality of network model prediction values; 3) optimizing weight parameters of a deep network through a back propagation algorithm by using the hybrid KL divergence loss ofthe marked data and the group consistency loss of the unmarked data, and 4) classifying the test data set by using the trained deep network. The semi-supervised deep network picture classification method can realize efficient and accurate classification of the pictures.

Description

technical field [0001] The invention relates to a picture classification method, in particular to a semi-supervised deep network picture classification method. Background technique [0002] At present, in image processing and digital recognition, the fully supervised deep neural network performs excellently, but the labeled data required to train a better deep network is also huge and expensive, while the semi-supervised algorithm overcomes the above shortcomings, and the required labeled data Compared with the fully supervised method, it is greatly reduced, and the connection between unlabeled data is explored under the guidance of labeled data, thereby improving the accuracy of classification. [0003] The core difficulty of semi-supervised classification algorithms is how to effectively use unlabeled data. Now semi-supervised classification algorithms combined with deep networks can be roughly divided into two categories: pseudo-labeling approach and consistency principl...

AI Technical Summary

Problems solved by technology

[0004] Before the present invention was made, the current semi-supervised neural network used a supervised model and a supervised model for digital image recognition, and the deep neural network itself was easy to fall into the local optimal solution. When the supervised model fell into the local optimal solution , since the regulatory model only collects the information of a supervised model, it is impossible to effectively feed back the supervised model from a global perspective

Images