A Multispectral Image Sharpening Method Based on Transfer Learning Neural Network

Abstract

The invention discloses a multispectral image sharpening method based on a transfer learning neural network, comprising the following steps: reading the original multispectral image and a panchromatic image registered therewith, preprocessing the image data, and obtaining training samples; building Convolutional neural network structure; the training samples are input into the convolutional neural network, and the adaptive moment estimation algorithm is used to reduce the loss error to an acceptable range, so as to obtain the optimal solution of the network structure parameters; the same preprocessed The complete multispectral test sample is input into the optimal convolutional neural network structure, and the output is processed to obtain a high-resolution multispectral image; if it is necessary to sharpen the multispectral image of the missing band, it is necessary to fine-tune the trained network before testing . The invention can enhance the migration ability of the trained neural network to the multispectral image sharpening processing of the lost band while maintaining the original sharpening effect.

Description

technical field [0001] The invention relates to the field of remote sensing image processing, in particular to a multispectral image sharpening method based on a transfer learning neural network. Background technique [0002] With the characteristics of large amount of information and wide coverage, remote sensing images play an important role in many fields. In the military field, it can conduct all-round detection and monitoring of targets, and facilitate the collection of intelligence from all parties; in the civilian field, it is widely used in navigation, disaster detection and forecasting, resource surveying, etc. However, due to the limitations of the sensor imaging mechanism, commonly used remote sensing satellites cannot provide multispectral images with high spatial resolution and spectral resolution at the same time. In order to make up for this deficiency, most satellites today generally have two different types of sensors at the same time, respectively acquirin...

AI Technical Summary

Problems solved by technology

Since the convolutional neural network is based on the learning of training samples, its sharpening effect is highly dependent on the similarity between training samples and test samples. However, in the practical application of multispectral images, it is often accompanied by spectral band damage. Phenomenon, the convolutional neural network that has been trained is no longer suitable for such images. At this time, the network structure can only be adjusted and retrained using multispectral images of missing bands.

Images