Feature Fusion Method Based on Stacked Autoencoder

Abstract

The invention provides a feature fusion method based on a stacked autoencoder. The technical solution includes the following content: First, extract the local three-patch binary pattern texture feature of the image, select and extract several baseline features of the image by using a feature selection method, and concatenate all the obtained features to obtain a concatenation vector. Then, the concatenated vectors are normalized and then whitened. The whitened result is used as the input of SAE, and SAE is trained by layer-by-layer greedy training method. Finally, using the trained SAE, the SAE is fine-tuned through the softmax classifier to minimize the loss function, and the output of the SAE is a highly discriminative fusion feature vector. The features selected by the present invention have little redundancy, and provide richer information for feature fusion.

Description

technical field [0001] The invention belongs to the technical field of image fusion, and relates to a feature fusion method based on SAE (Stacked Autoencoder, stacked autoencoder), which improves the discrimination and efficiency of fusion features. Background technique [0002] Feature fusion refers to the technology of comprehensive analysis and fusion processing of the extracted feature information. In image understanding, the use of feature fusion can not only increase the feature information of the image, but also effectively integrate the advantages of the original features to obtain a more comprehensive feature expression of the target. Classical feature fusion algorithm (feature fusion algorithm can refer to Wang Dawei, Chen Dingrong, He Yizheng. A review of multi-feature image fusion technology for target recognition[J]. Avionics Technology, 2011,42(2):6-12.), directly Combining features directly in a certain way does not essentially consider the influence of the r...

AI Technical Summary

Problems solved by technology

[0004] Existing SAE-based feature fusion algorithm (SAE-based feature fusion algorithm can refer to the literature Chen Y, LinZ, Zhao X, et al..Deep Learning-Based Classification of Hyperspectral Data[J].IEEE Journal of Selected Topics in Applied Earth Observations&RemoteSensing,2014,7(6):2094-2107.), the selected feature dimension is high, the fusion network structure is complex, the training time is long, and it cannot meet the real-time requirements

In addition, the redundancy between the features is large, and the complementarity is small, and the feature discrimination is not significantly improved after fusion.

Images