Hyperspectral image noise reduction method based on three-dimensional quasi-recurrent neural network

Abstract

The invention discloses a hyperspectral image noise reduction method based on a three-dimensional quasi-recurrent neural network, and belongs to the field of image processing. The hyperspectral imagenoise reduction method based on the three-dimensional quasi-recurrent neural network comprises the steps of constructing the three-dimensional quasi-recurrent neural network taking three-dimensional convolution, quasi-recurrent pooling and an alternating direction structure as a core according to space-spectrum correlation and spectrum global correlation; training a three-dimensional quasi-recurrent neural network through the noise reduction training data set; and removing complex noise in the noisy hyperspectral image by using the trained three-dimensional quasi-recurrent neural network. According to the method, the noise reduction recovery of the hyperspectral image can be completed with high quality, the recovery result can be ensured to have space and spectrum fidelity, the recovery efficiency of the hyperspectral image is greatly improved, the method can be used for the hyperspectral image acquired by an imaging system with any spectral resolution, and the application range of thehyperspectral image is expanded. The method can be applied to the fields of remote sensing imaging, geological exploration, agricultural production and biomedicine.

Description

technical field [0001] The patent of the invention relates to a hyperspectral image denoising method based on a three-dimensional quasi-recurrent neural network, in particular to a method capable of reconstructing a noiseless hyperspectral image from a noisy hyperspectral image collected in a complex environment, which belongs to the field of image processing. Background technique [0002] Hyperspectral images record discrete spectral information at each spatial location in real scenes. Compared with common color RGB images, rich spectral details can reflect richer scene information, such as scene lighting and materials, which makes hyperspectral images widely used in geological exploration, agricultural production and biomedical fields. However, due to the limited light intensity of each spectral segment, hyperspectral images are usually affected by Gaussian noise, streak noise, dead line noise, and impulse noise during the acquisition process, and the resulting image quali...

AI Technical Summary

Problems solved by technology

While the latter, due to the lack of representation capabilities for global spectral correlations, tend to perform inferior to model-based methods

The trade-off between the representational power of the model and the flexibility of the spectral dimension poses a fundamental limitation on the real-world application of such methods

Images