Image classification method based on direct push semi-supervised depth learning

Abstract

The invention relates to an image classification method based on direct push semi-supervised depth learning, which comprises the following steps: preparing a semi-supervised image data set, and dividing the training data into a training data set and a verification data set, wherein the training data set has a part of data labeled and the other part unlabeled; verifying the labeled data; on the labeled training dataset, training the general depth neural network image classification model. When the trained model achieves the expected precision on the verification dataset, the network model parameters are saved. A direct push semi-supervised depth convolution neural network model based on the Min-Max principle is built and both labeled and unlabeled data in the training data set are used totrain the model circularly; when the number of cycles reaches the maximum number of cycles, the parameters of the network model are saved. The trained model is used to calculate the recognition accuracy of the label or test data set of the test image. The TSSDL algorithm provided by the invention has good portability.

Description

technical field [0001] The invention belongs to the technical field of computer vision image classification, and in particular relates to an image classification method based on Min-Max rule-based transductive semi-supervised deep learning. Background technique [0002] To date, deep convolutional neural networks have demonstrated high-level performance in many computer vision applications, such as image classification, object detection, face recognition, and image translation. Large-scale training datasets containing millions of annotated images are one of the important factors driving the success of deep convolutional neural networks. However, manually annotating to create a large-scale, high-quality training set is very time-consuming, expensive, or even difficult to accomplish (such as the training set for image semantic segmentation). At the same time, a vast amount of unlabeled images can be easily obtained from the Internet by web crawlers and search engines. Theref...

AI Technical Summary

Problems solved by technology

Since the feature quality output by the deep convolutional neural network model is poor in the initial stage of training, and the feature quality gradually improves as the training process progresses, it is difficult to combine these two types of methods with the training of deep convolutional neural networks.

Secondly, the traditional SSL and TSSL methods treat each unlabeled sample equally, and cannot reasonably deal with singular samples and uncertain data samples, which affects the stability and effect of model training.

This problem is more obvious in the training of deep convolutional neural network, because the feature quality generated in the initial stage of network model training is poor and unstable, which may mislead the model to train in the wrong direction

Images