CN-122023150-A - Crop disease image enhancement system and method

CN122023150ACN 122023150 ACN122023150 ACN 122023150ACN-122023150-A

Abstract

The application provides a crop disease image enhancement system and method, which are characterized in that depth features of a visible light image and a near infrared image are respectively extracted through a feature extraction network, an initial disease-spot thermodynamic diagram of a target crop blade is further generated, an association weight vector of the near infrared depth feature and the visible light depth feature is determined, the visible light depth feature is subjected to channel weighting through the association weight vector, a disease-spot-background discrimination degree diagram of the target crop blade is fused with the initial disease-spot thermodynamic diagram based on a weighting result, a dynamic weighted disease recognition loss function is constructed according to the disease-spot-background discrimination degree diagram in a training process of the feature extraction network, and then the visible light image is decoded and reconstructed according to the feature extraction network after parameter optimization to obtain a final image with obviously enhanced disease-spot region visual features. By adopting the scheme of the application, the image dynamic enhancement can be carried out on the target crop blade by combining the morphological prior of the visible light image and the spectrum prior from the multispectral analysis.

Inventors

CHEN JIAHE
CHEN YONGBING
ZHANG YINGYI
ZHANG HAILI

Assignees

温州科技职业学院

Dates

Publication Date: 20260512
Application Date: 20251027

Claims (10)

1. The image dynamic enhancement method based on the disease spot-background discrimination guidance is applied to a crop disease image enhancement system and is characterized by comprising the following steps: Synchronously acquiring a visible light image and a near infrared image of a target crop blade; Respectively extracting depth features of the visible light image and the near infrared image through a feature extraction network to obtain visible light depth features and near infrared depth features; Generating an initial disease spot thermodynamic diagram of a target crop blade according to the visible light depth characteristic, and determining a relevance weight vector of the near infrared depth characteristic and the visible light depth characteristic in a channel dimension; channel weighting is carried out on the visible light depth features through the relevance weight vector, and then the obtained weighted result and the initial disease spot thermodynamic diagram are fused into a disease spot-background discrimination map of the target crop leaf; In the training process of the feature extraction network, a dynamically weighted disease recognition loss function is constructed according to the disease spot-background discrimination map so as to optimize network parameters, and then the visible light image is decoded and reconstructed according to the feature extraction network after parameter optimization, so that a final image with obviously enhanced visual features of a disease spot area is obtained.
2. The method of claim 1, wherein extracting depth features of the visible light image and the near-infrared image, respectively, by a feature extraction network, the obtaining the visible light depth features and the near-infrared depth features specifically comprises: Configuring a deep convolution network with shared weights as a feature extraction network; And respectively inputting the visible light image and the near infrared image into the feature extraction network, and further obtaining visible light depth features and near infrared depth features through forward propagation calculation.
3. The method of claim 1, wherein generating an initial plaque thermodynamic diagram of a target crop leaf from the visible light depth features comprises: constructing a disease spot area prediction module comprising a convolution layer and an up-sampling layer; Inputting the visible light depth features into the lesion area prediction module; And outputting the initial plaque thermodynamic diagram through forward calculation of the plaque area prediction module.
4. The method of claim 1, wherein determining the correlation weight vector of the near infrared depth feature and the visible depth feature in the channel dimension specifically comprises: and carrying out channel interaction and global pooling on the visible light depth features and the near infrared depth features, and outputting the relevance weight vector through a weight generation network.
5. The method of claim 1, wherein channel weighting the visible light depth features by the relevance weight vector specifically comprises: carrying out channel-by-channel multiplication operation on the relevance weight vector and the visible light depth characteristic; and carrying out channel compression and space activation on the characteristics after multiplication operation to generate a spectrum enhancement weight map.
6. The method of claim 1, wherein fusing the initial plaque thermodynamic diagram with the plaque-background discrimination map of the target crop leaf based on the weighted result specifically comprises: Obtaining a spectrum enhancement weight graph obtained by channel weighting; Spatially aligning the spectral enhancement weight map with the initial plaque thermodynamic diagram; And carrying out weighted fusion operation on the spectrum enhancement weight map and the initial disease spot thermodynamic diagram after spatial alignment, and outputting a fused disease spot-background discrimination map.
7. The method of claim 1, wherein constructing a dynamically weighted disease recognition loss function from the lesion-background discrimination map during training of the feature extraction network to optimize network parameters specifically comprises: Predicting a pixel-level disease area based on the visible light depth characteristic, and further determining pixel-level basic disease identification loss; Combining the disease spot-background discrimination map as a space weight matrix with the pixel-level basic disease identification loss to construct a dynamically weighted disease identification loss function; And driving a back propagation process through the dynamically weighted disease identification loss function to update parameters of the feature extraction network and related modules.
8. A crop disease image enhancement system comprising an image dynamic enhancement unit, wherein the image dynamic enhancement unit comprises: The acquisition module is used for synchronously acquiring a visible light image and a near infrared image of the target crop blade; The processing module is used for respectively extracting depth features of the visible light image and the near infrared image through a feature extraction network to obtain visible light depth features and near infrared depth features; The processing module is further used for generating an initial disease spot thermodynamic diagram of the target crop leaf according to the visible light depth feature, and determining a relevance weight vector of the near infrared depth feature and the visible light depth feature in a channel dimension; the processing module is further used for carrying out channel weighting on the visible light depth characteristic through the relevance weight vector, and further fusing the weighted result and the initial disease spot thermodynamic diagram into a disease spot-background discrimination degree diagram of the target crop leaf; And the execution module is used for constructing a dynamically weighted disease recognition loss function according to the disease spot-background discrimination degree graph in the training process of the feature extraction network so as to optimize network parameters, and further decoding and reconstructing the visible light image according to the feature extraction network after parameter optimization to obtain a final image with obviously enhanced visual features of the disease spot area.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the plaque-background discrimination directed image dynamic enhancement method according to any one of claims 1 to 7 when the computer program is executed.
10. A computer-readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the plaque-background discrimination directed image dynamic enhancement method according to any one of claims 1 to 7.

Description

Crop disease image enhancement system and method Technical Field The application relates to the technical field of image enhancement, in particular to a crop disease image enhancement system and method. Background Image enhancement is a basic image processing method whose core purpose is not to restore the "true" condition of an image, but rather to improve the visual display of an image or convert it into a form more suitable for a particular task (e.g., human eye observation or machine analysis) by selectively highlighting or suppressing certain information in the image, common methods including contrast stretching, histogram equalization, sharpening, filtering, etc. The crop disease image enhancement is a specific application of image enhancement in a specific field, and the core purpose of the crop disease image enhancement is to purposefully highlight the visual characteristics of disease spot areas (such as disease leaves and disease stems) in a crop image, and inhibit the interference of healthy tissues and complex backgrounds, so that a crop image which is easier to be recognized by human eyes or accurately analyzed and diagnosed by a computer automatic recognition system is generated, in the visible light image, early disease spots, slight disease spots, shadows, water stains, soil stains, natural aging and the like are very similar in color and texture, while the traditional enhancement method (such as global contrast enhancement) can enhance all low-contrast areas in a same-eye manner, but amplifies background interference, so that the false detection rate is high, and therefore, how to dynamically enhance the image of a target crop blade by combining the morphological prior of the visible light image and the spectrum prior from multispectral analysis becomes a difficult problem faced by the industry. Disclosure of Invention Based on the above, the application provides a crop disease image enhancement system and a crop disease image enhancement method for dynamically enhancing images of target crop leaves by combining a morphological prior of a visible light image and a spectrum prior from multispectral analysis. In a first aspect, the present application provides an image dynamic enhancement method based on a lesion-background discrimination guide, applied to a crop disease image enhancement system, the method comprising the steps of: Synchronously acquiring a visible light image and a near infrared image of a target crop blade; Respectively extracting depth features of the visible light image and the near infrared image through a feature extraction network to obtain visible light depth features and near infrared depth features; Generating an initial disease spot thermodynamic diagram of a target crop blade according to the visible light depth characteristic, and determining a relevance weight vector of the near infrared depth characteristic and the visible light depth characteristic in a channel dimension; channel weighting is carried out on the visible light depth features through the relevance weight vector, and then the obtained weighted result and the initial disease spot thermodynamic diagram are fused into a disease spot-background discrimination map of the target crop leaf; In the training process of the feature extraction network, a dynamically weighted disease recognition loss function is constructed according to the disease spot-background discrimination map so as to optimize network parameters, and then the visible light image is decoded and reconstructed according to the feature extraction network after parameter optimization, so that a final image with obviously enhanced visual features of a disease spot area is obtained. In some embodiments, extracting depth features of the visible light image and the near-infrared image through a feature extraction network respectively, to obtain the visible light depth feature and the near-infrared depth feature specifically includes: Configuring a deep convolution network with shared weights as a feature extraction network; And respectively inputting the visible light image and the near infrared image into the feature extraction network, and further obtaining visible light depth features and near infrared depth features through forward propagation calculation. In some embodiments, generating an initial plaque thermodynamic diagram of the target crop leaf from the visible light depth features specifically includes: constructing a disease spot area prediction module comprising a convolution layer and an up-sampling layer; Inputting the visible light depth features into the lesion area prediction module; And outputting the initial plaque thermodynamic diagram through forward calculation of the plaque area prediction module. In some embodiments, determining the relevance weight vector of the near infrared depth feature and the visible light depth feature in the channel dimension specifically includes: and carrying out channel interaction and global