Remote sensing image scene accurate classification method based on Dual-ResNet small sample learning

Abstract

The invention discloses a remote sensing image scene accurate classification method based Dual-ResNet small sample learning, which comprises the following steps of: firstly, performing feature extraction on two simultaneously input remote sensing scene images X1 and X2 by using ResNet 18 to respectively obtain features Gw (X1) and Gw (X2), and sharing a weight by the two ResNet-18 in a training process; secondly, calculating the distance between the two ResNet18 network extraction features, then comparing the distance between the set of features with a preset threshold value to obtain a predicted category label, and finally outputting the predicted category label. According to the method, the Dual-ResNet is used for training the two input remote sensing images at the same time, the featuredistance between the two images is obtained through respective learning, compared with common deep learning, the intra-class distance can be reduced and the inter-class distance can be increased while image features are learned, the dependence of deep learning on the data size is improved, the method has a very good effect on remote sensing image scene classification with a small sample data volume.

Description

technical field [0001] The invention belongs to the field of image processing, in particular to a method for accurately classifying remote sensing image scenes based on Dual-ResNet small-sample learning. Background technique [0002] Remote sensing image scene classification is a research hotspot in the field of image processing, and it is also a very challenging task in computer vision. With the rapid development of remote sensing image technology, the amount of optical remote sensing image data has grown rapidly, and remote sensing images contain very rich information. , such as spatial information, texture information, geometric structure information of ground objects, etc. Objects in images have the characteristics of large differences in the same category and high similarity between some categories. Therefore, how to effectively and automatically classify and recognize remote sensing image scenes has attracted the attention of many researchers. With the development of ...

AI Technical Summary

Problems solved by technology

[0007] (1) The design of the deep neural network is complex, which increases the computing time in the training process, and an over-deep neural network can easily lead to over-fitting, gradient disappearance, etc.

[0008] (2) The distribution of remote sensing image data is not well considered. For the scene classification of remote sensing images, the number of samples of remote sensing images in each category often has an unbalanced distribution. If this is ignored in the training process, It can easily lead to lower prediction accuracy for sample categories with fewer distributions

[0009] (3) The training process of the neural network has a strong dependence on the number of samples. Generally, the more abundant the number of samples, the better the final prediction result of the neural network trained. However, for images such as remote sensing, high artificial The cost of labeling makes the amount of remote sensing image data for training not particularly large, so small sample learning will become a major development direction in the field of remote sensing and even in the field of computer vision

Images