Muskmelon quality detection method based on hyperspectral image depth feature fusion

Abstract

The invention discloses a muskmelon quality detection method based on hyperspectral image depth feature fusion. The method comprises the following steps: constructing an end-to-end muskmelon sweetnessand hardness prediction framework, enabling the prediction framework to serve as an unsupervised network model, and enabling the framework to integrate the spectral features and spatial features of muskmelons into a unified neural network; carrying out 3 * 3 and 5 * 5 filling on the edge pixel points to ensure that the edge information of the image is rich; and constructing a small sample of correlation between characteristic pixels, calculating the distance between the characteristic pixels, taking the distance as a weight value, and then calculating the sweetness value and the hardness value of each pixel point in a measured area to estimate the sweetness and the hardness of the whole muskmelon. According to the method, the problem of model overfitting caused by too little annotation data is solved, and the edge information of the image is effectively extracted by filling the pixels of the edge neighborhood.

Description

technical field [0001] The invention relates to the field of melon quality detection, in particular to a method for melon quality detection based on deep feature fusion of hyperspectral images, which detects the sweetness and hardness of melon hyperspectral images through feature learning of spatial information and spectral information of images . Background technique [0002] The hyperspectral image has a high-resolution spectrum, and its spectral resolution can reach the nanometer level, which makes it have more spectral information than the visible light image, but there are two fundamental problems in the hyperspectral image, first: due to The spectral resolution is high, so the hyperspectral image has the disadvantages of huge data volume, serious data redundancy, and many bands; second: the spatial resolution of the hyperspectral image is limited, there are a large number of mixed pixels, and the detection accuracy will vary with the feature Hughes phenomenon in which...

AI Technical Summary

Problems solved by technology

[0004] The method of using deep neural network to process hyperspectral images has greatly solved the shortcomings caused by traditional methods, but most of the existing deep learning methods have certain limitations: First, most deep learning methods need to determine A large number of parameters, so the existing deep network model has high complexity, which requires a large number of training samples

Second, the existing hyperspectral image classification or prediction only uses the deepest features for prediction or classification, and does not integrate low-level information

Third, the edge information of the image also has a great influence on the final prediction. The existing hyperspectral image prediction or classification ignores the extraction of edge information, so in the final prediction, the edge of the image is extremely inaccurate. Optimistic, unable to meet the needs of practical applications

Images