High-resolution remote sensing image classification method based on feature pooling and divisive normalization expression

Abstract

A remote sensing image classification method based on feature pooling and divisive normalization includes the steps that firstly, linear filtering is conducted on remote sensing images, and the characteristics of linear filtering results are merged through feature pooling; then, the high-order correlation in the remote sensing images is reduced through divisive normalization, and local descriptors are extracted in a dense mesh mode; finally, after Hellinger kernel mapping and feature dimension reduction, feature coding is conducted to form global expression of the remote sensing images, and after training and prediction, classification of the remote sensing images is completed. By means of the method, high image classification precision can be obtained, redundancy enhanced target position invariance of the low-order and high-order statistical correlation in the images is eliminated, feature compression is compact, the feature dimension is reduced, computing efficiency is improved, and better classification performance is obtained; the method is simple, easy to implement and high in precision.

Description

technical field [0001] The invention relates to a technology in the field of remote sensing image processing, in particular to a remote sensing image classification method based on feature pooling and denormalization. Background technique [0002] A key step in the classification of high-resolution remote sensing images is to extract appropriate image features, and widely used features include texture features and local features. Common texture features include Gabor texture and LBP texture, and local features include HOG features, SIFT features, and Dense SIFT features. However, there are problems in image classification such as target scale, inconsistent illumination, target occlusion, and low-order and high-order correlation. [0003] Most of the existing visual vocabulary-based remote sensing image object classification methods are based on Gaussian blur and sampling to generate a multi-layer Gaussian space pyramid; then through SIFT feature extraction and LBP feature e...

AI Technical Summary

Problems solved by technology

However, image classification methods based on visual vocabulary are essentially limited by the shortcomings of the bag-of-words model.

First, the way of using frequency histogram features is not the optimal image representation

Second, the quantization of low-level descriptors (such as SIFT features) by the bag-of-words model is a lossy process

However, the convolutional neural network in this type of technology contains feature extraction of many convolutional layers, and each layer involves the convolution operation of a large number of filter cores. When further considering the combination of multiple convolutional neural networks , as the number of combinations increases, its computational complexity will increase exponentially

Images