Cross-domain image classification model construction method and device based on transfer learning

Abstract

The invention discloses a cross-domain image classification model construction method and device based on transfer learning. The cross-domain image classification model construction method performs image feature extraction on data of a source domain data set and a target domain data set by using a convolution layer of a pre-trained Inception-V3 model, and deletes the part, which is greatly different from the target domain, in the source domain data through the clustering algorithm, so as to reduce the noise influence possibly generated by the source domain data, solve the noise influence caused by the difference between the source domain and the target domain, improve the similarity between the source domain data and the target domain data, and reduce the noise influence of the source domain data. An attention mechanism is added to extract image feature information of more target domain data when model fine adjustment is carried out on the target domain. Therefore, the utilization rateof the target domain data is improved, and more image feature information of the target domain data is extracted, and the cross-domain image classification accuracy is improved, and the adaptabilityof the migration model to the target domain is enhanced, so that the accuracy of the finally constructed image classification model is ensured.

Description

technical field [0001] The present application relates to the technical field of image classification and recognition, in particular to a method and device for constructing a cross-domain image classification model based on transfer learning. Background technique [0002] With the rapid development of the Internet, the number of multimedia images such as images and videos is increasing day by day. Image classification refers to the classification of categories according to their characteristic information by means of certain algorithms or models. It has important applications in the fields of image detection, video surveillance and target recognition. At present, the most commonly used method in the field of image classification is a classification method based on deep learning, that is, using models such as CNN (Convolutional Neural Networks, Convolutional Neural Networks) to extract features from images, and using labeled image datasets to train image classification models....

AI Technical Summary

Problems solved by technology

[0005] 1. When using the transfer learning method, due to the difference in image features between the source domain and the target domain, the noise generated by the source domain data will affect the migration effect. If the difference between the source domain and the target domain is greater, there may also be negative transfer phenomenon, making the source domain The data negatively impacts the target domain data;

[0006] 2. Since the image in the target domain is the correct image that needs to be classified, it is necessary to extract more image features of its target domain data, but the utilization rate of unlabeled data in the target domain is low in the current method

Images