A hyperspectral-based method for extracting banana fusarium wilt leaf phenotypic texture

Abstract

The application discloses a kind of based on hyperspectral banana wilt leaf phenotype texture extraction method, especially in the early stage of disease leaf yellowing phenotype texture extraction method.The implementation function of the present application includes two parts of feature spectral image preprocessing, texture calculation.In feature spectral image preprocessing part, mainly for the feature spectral image of banana wilt related image preprocessing is carried out, the position of banana leaf is extracted from the image;Texture calculation part mainly carries out filtering, gradient calculation, different waveband image fusion etc.to the feature spectral image after extracting leaf position, finally obtains the real leaf phenotype texture information.Compared with the method of extracting texture features of ordinary color image, the texture features extracted by hyperspectral camera under two characteristic wavebands in the present application are fused, which can reflect the real banana leaf texture better than ordinary camera, and provide phenotype information support for subsequent visual recognition and disease judgment.

Description

Technical Field

[0001] This invention relates to the field of banana wilt disease technology, specifically a method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging. Background Technology

[0002] Fusarium wilt of bananas is an incurable disease, often referred to as the "cancer of bananas." Currently, there is no specific cure for this disease; early detection and eradication are the most effective methods for managing it. There are two main methods for detecting Fusarium wilt: one is to sample the soil in the planting area and use methods such as PCR to identify the presence of pathogens. However, this method is time-consuming, requires a large number of samples, and is labor-intensive, making it difficult to apply in actual cultivation. The second method is to judge based on the characteristics of the banana plant's leaves. Infected leaves will show signs of chlorosis, yellowing, drooping, and curling, eventually withering starting from the edges. However, this method requires professional judgment based on leaf observation, which is difficult for inexperienced farmers to make accurate assessments. Furthermore, for farmers with large planting areas, manual judgment is insufficient to cover the entire planting area.

[0003] Currently, computer vision for identifying crop diseases has gradually entered the agricultural planting field. The emergence of new equipment such as handheld detectors and multispectral drones has provided a more efficient and accurate identification method than manual labor. The use of hyperspectral cameras combined with machine vision technology to automatically identify banana wilt disease helps farmers detect the disease in a timely manner and reduces manual labor; this research is of great significance. Summary of the Invention

[0004] The purpose of this invention is to provide a method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging. By performing a fusion operation on the texture features extracted by a hyperspectral camera under two characteristic bands, the method can better reflect the true texture of banana leaves than a conventional camera, providing phenotypic information support for subsequent visual identification and judgment of the disease, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging includes two parts: feature spectral image preprocessing and texture calculation.

[0007] The characteristic spectral image preprocessing is performed on the characteristic spectral image of banana wilt disease to extract the position of banana leaves from the image;

[0008] The texture calculation includes filtering, gradient calculation, and image fusion processing of different bands on the feature spectral image after extracting the leaf position, and finally obtaining the real leaf phenotypic texture information.

[0009] Furthermore, the specific method for preprocessing the characteristic spectral image is as follows:

[0010] S1: Use a hyperspectral camera to capture hyperspectral images of banana tree leaves to obtain hyperspectral image information of the leaves;

[0011] S2: Remove the background from the image. Extract the 430nm and 725nm layers from the hyperspectral image and remove the background.

[0012] S3: Perform difference fusion between the 725nm layer and the 430nm layer to obtain an image with low background reflectivity and high blade reflectivity. Then, perform image processing on the image by first dilation and then erosion.

[0013] S4: Use Gaussian blur to process the image and reduce excessive numerical differences in some neighboring areas;

[0014] S5: Perform binarization segmentation on the image, adaptively calculate the segmentation threshold using the maximum inter-class variance method, and perform binarization segmentation on the image to obtain a black and white image, where the leaves are white with a value of 255 and the background is black with a value of 0.

[0015] S6: Then perform dilation and erosion on the image to obtain the position mask of the banana leaf;

[0016] S7: Perform an AND operation between the obtained position mask and the 680nm and 725nm images to extract the banana leaf positions within the 680nm and 725nm images.

[0017] Furthermore, S3 performs the following image processing method of first dilating and then eroding: by dilating, the edges of the blades are enlarged and the noise holes inside the blades are filled, and then the enlarged texture after dilution is eroded.

[0018] Furthermore, the texture calculation process is as follows:

[0019] S1: The bilateral filtering algorithm is used to filter the images of the 680nm and 725nm bands respectively, smoothing the images, reducing noise, and preserving edge information;

[0020] S2: Calculate the magnitude and direction of the gradient of the two images respectively, and roughly calculate the edge position and edge intensity of the images;

[0021] S3: Use non-maximum suppression to find the edge with the strongest intensity among all edge information, and remove the edges with weaker intensity on both sides of the edge to make the edge more refined;

[0022] S4: Use hysteresis thresholds to filter out high-intensity edges, medium-intensity edges, and low-intensity edges from all edge information;

[0023] S5: After classifying the three types of edges, edges belonging to high intensity edges are identified as real edges, edges belonging to low intensity edges are eliminated, and medium intensity edges need to be judged a second time. If a medium intensity edge has an uninterrupted edge connecting to a high intensity edge, it is considered an edge; otherwise, it is eliminated.

[0024] S6: After the hysteresis thresholding is applied to the 680nm and 725nm images, two edge texture maps are obtained respectively. The two edge texture maps are then weighted and fused, with both weights being 0.5. After the weighted fusion, the intensity of overlapping edges does not change much, while the intensity of non-overlapping edges is weakened.

[0025] S7: The weighted fused edge texture map is subjected to hysteresis thresholding again to obtain high-intensity edges, medium-intensity edges and low-intensity edges again. Among them, overlapping edges will continue to be classified as high-intensity edges because the intensity changes little. Non-overlapping edges will become medium-intensity edges or low-intensity edges after processing because the intensity is weakened. Low-intensity edges will be removed, and medium-intensity edges will be judged again to retain the edges that are connected to high-intensity edges.

[0026] S8: The final image is a phenotypic texture map of banana leaves, which reflects the veins, edges, and folds within the leaves.

[0027] Furthermore, in S4, two thresholds, high and low, are needed to distinguish between high-intensity edges, medium-intensity edges, and low-intensity edges. The threshold calculation is based on the idea of ​​maximum entropy. By calculating the entropy values ​​of the foreground and background after image segmentation, the segmentation threshold is determined. The greater the entropy of the foreground and background, the richer the information contained in the foreground and background of the image. At this time, the threshold used to segment the image is used as the high threshold for segmenting high-intensity edges and medium-intensity edges. The formula is as follows:

[0028]

[0029]

[0030] H(q)=H0(q)+H1(q) (3)

[0031] Where q is the threshold for segmentation, with a value ranging from 0 to 255, P0(q) represents the cumulative probability of background pixels, P1(q) represents the cumulative probability of foreground pixels, H0 represents the background entropy, H1 represents the foreground entropy, and H(q) represents the total entropy of the image when the threshold is q. The segmentation threshold q with the maximum entropy is selected as the high threshold TH through 256 iterations.

[0032] Furthermore, after obtaining the high threshold TH, the idea of ​​maximum inter-class variance is used to find the low threshold TL from the range of 0-TH for segmenting medium-intensity edges and low-intensity edges, as shown in the following formula:

[0033] μ i =ω0×μ0+ω1×μ1 (5)

[0034] σ i =ω0×(μ0-μ i ) 2 +ω1×(μ1-μ i ) 2 (6)

[0035] i is the threshold for segmentation, with a value ranging from 0 to TH. ω0 is the proportion of foreground pixels in the image, ω1 is the proportion of background pixels in the image, μ0 is the average gray level of the foreground, μ1 is the average gray level of the background, μi is the average gray level of the image when the segmentation threshold is i, and σi is the inter-class variance. The maximum inter-class variance is found through TH iterations, and the segmentation threshold at this time is used as the lower limit threshold TL.

[0036] Edges with an intensity higher than TH are classified as high-intensity edges, those with an intensity lower than TL are classified as low-intensity edges, and those in between are classified as medium-intensity edges.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] This invention provides a method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging. The method involves preprocessing the characteristic spectral images related to banana wilt in the feature spectral image preprocessing section to extract the location of banana leaves from the images. The texture calculation section performs filtering, gradient calculation, and image fusion of different bands on the feature spectral images after leaf location extraction to ultimately obtain accurate leaf phenotypic texture information. Compared with methods for extracting texture features from ordinary color images, this invention uses texture features extracted by a hyperspectral camera in two characteristic bands for fusion calculation, which better reflects the true texture of banana leaves than ordinary cameras, providing phenotypic information support for subsequent visual identification and disease judgment. Attached Figure Description

[0039] Figure 1 This is a flowchart of the characteristic spectral image preprocessing process of the present invention;

[0040] Figure 2 This is a flowchart of the texture calculation and processing flow of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This invention provides a hyperspectral-based method for extracting the phenotypic texture of banana wilt leaves, comprising two parts: feature spectral image preprocessing and texture calculation. The feature spectral image preprocessing part mainly performs image preprocessing on feature spectral images related to banana wilt, extracting the location of banana leaves from the images. The texture calculation part mainly performs filtering, gradient calculation, and image fusion of different bands on the feature spectral images after extracting the leaf locations, ultimately obtaining the true phenotypic texture information of the leaves. Compared with methods for extracting texture features from ordinary color images, this method uses texture features extracted by a hyperspectral camera under two feature bands for fusion calculation, which reflects the true texture of banana leaves more accurately than ordinary cameras, providing phenotypic information support for subsequent visual recognition and disease judgment.

[0043] Please see Figure 1 Specifically, the methods for feature spectral image preprocessing are as follows:

[0044] Step 1: First, use a hyperspectral camera to capture hyperspectral images of banana tree leaves to obtain hyperspectral image information of the leaves;

[0045] Step 2: The background needs to be removed from the image. The position of the leaf in the image is extracted. According to the hyperspectral reflectance curve of the leaf, the reflectance of healthy banana leaves and diseased leaves is similar in the 725nm band, while the reflectance of the background is low. However, the reflectance of banana leaves is low in the 430nm band, while the reflectance of the background is high. The 430nm and 725nm layers are extracted from the hyperspectral image for background removal.

[0046] Step 3: Perform interpolation fusion on the 725nm layer and the 430nm layer to obtain an image with low background reflectivity and high blade reflectivity. This image contains noise from the interpolation fusion. To eliminate this noise, perform image processing: first dilation, then erosion. Dilation enlarges the blade edges and fills noise holes within the blades, then erosion shrinks the enlarged texture.

[0047] Step 4: Process the image using Gaussian blur. Because the image after difference fusion is prone to large pixel value differences in local neighborhoods, which can affect subsequent processing algorithms, Gaussian blur can smooth the image and reduce excessive numerical differences in some neighborhoods.

[0048] Step 5: Perform binarization segmentation on the image. After the above processing, the data at the position of the banana leaf in the image will be white, while the background will be close to black. At this time, the Otsu's method is used to adaptively calculate the segmentation threshold and perform binarization segmentation on the image to obtain a black and white image, where the leaf is white with a value of 255 and the background is black with a value of 0.

[0049] Step Six: Perform dilation followed by erosion on the image. Because there are a few holes due to noise after binarization, the dilation operation can fill the holes, and the erosion operation restores the edge expansion caused by the dilation operation, finally obtaining the position mask map of the banana leaf;

[0050] Step 7: Perform an AND operation between the position mask obtained in the previous step and the 680nm and 725nm images to extract the banana leaf positions within the 680nm and 725nm images.

[0051] Please see Figure 2 Specifically, the texture calculation processing steps are as follows:

[0052] Step 1: Use a bilateral filtering algorithm to filter the images in the 680nm and 725nm bands respectively, smoothing the images, reducing noise, and preserving edge information;

[0053] Step 2: Calculate the magnitude and direction of the image gradient for each of the two images, and roughly calculate the edge position and edge intensity of the images;

[0054] Step 3: Use non-maximum suppression to find the edge with the highest intensity among all edge information, and remove the edges with lower intensity on both sides of the edge to make the edge more refined;

[0055] Step 4: Use hysteresis thresholding to filter out high-intensity edges, medium-intensity edges, and low-intensity edges from all edge information. This requires calculating two thresholds, high and low, to distinguish these three categories. The high threshold calculation is based on the idea of ​​maximum entropy. The segmentation threshold is determined by calculating the entropy values ​​of the foreground and background after image segmentation. The higher the entropy of the foreground and background, the richer the information contained in the foreground and background of the image. The threshold used for image segmentation in this case serves as the high threshold for separating high-intensity edges from medium-intensity edges. The formula is as follows:

[0056]

[0057]

[0058] H(q)=H0(q)+H1(q) (3)

[0059] Where q is the segmentation threshold, ranging from 0 to 255, P0(q) represents the cumulative probability of background pixels, P1(q) represents the cumulative probability of foreground pixels, H0 represents the background entropy, H1 represents the foreground entropy, and H(q) represents the total entropy of the image when the threshold is q. The segmentation threshold q with the maximum entropy is selected as the high threshold TH through 256 iterations.

[0060] After obtaining the high threshold TH, the idea of ​​maximum inter-class variance is used to find the low threshold TL from the range of 0-TH to segment medium-intensity edges and low-intensity edges. The formula is as follows:

[0061] μ i =ω0×μ0+ω1×μ1 (5)

[0062] σ i =ω0×(μ0-μ i ) 2 +ω1×(μ1-μ i ) 2 (6)

[0063] i is the segmentation threshold, ranging from 0 to TH; ω0 is the proportion of foreground pixels in the image; ω1 is the proportion of background pixels in the image; μ0 is the average gray level of the foreground; μ1 is the average gray level of the background; μi is the average gray level of the image when the segmentation threshold is i; and σi is the inter-class variance. The segmentation threshold at the maximum inter-class variance is found through TH iterations and is used as the lower threshold TL.

[0064] Edges with an intensity higher than TH are classified as high-intensity edges, those with an intensity lower than TL are classified as low-intensity edges, and those in between are classified as medium-intensity edges.

[0065] Step 5: After classifying the three types of edges, edges belonging to high intensity are identified as real edges, edges belonging to low intensity are eliminated, and medium intensity edges need to be judged a second time. If a medium intensity edge has an uninterrupted edge connecting to a high intensity edge, it is considered an edge; otherwise, it is eliminated.

[0066] Step 6: After the hysteresis thresholding is applied to the 680nm and 725nm images, two edge texture maps are obtained respectively. The two edge texture maps are then weighted and fused, with both weights being 0.5. After the weighted fusion, the intensity of overlapping edges does not change much, while the intensity of non-overlapping edges is weakened.

[0067] Step 7: Perform hysteresis thresholding on the weighted fused edge texture map again to obtain high-intensity edges, medium-intensity edges, and low-intensity edges again. Among them, overlapping edges will continue to be classified as high-intensity edges because the intensity change is not significant. Non-overlapping edges will become medium-intensity or low-intensity edges after processing because their intensity is weakened. Low-intensity edges will be removed, and medium-intensity edges will be judged again to retain the edges that are connected to high-intensity edges.

[0068] Step 8: The final image obtained is the phenotypic texture map of the banana leaf.

[0069] This phenotypic texture map can reflect the veins, edges, and folds within banana leaves. The phenotypic texture information helps neural network algorithms identify banana leaves infected with Fusarium wilt.

[0070] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging, comprising two parts: feature spectral image preprocessing and texture calculation, characterized in that: The characteristic spectral image preprocessing is performed on the characteristic spectral images of banana wilt disease to extract the location of banana leaves from the images; the specific method of characteristic spectral image preprocessing is as follows: S1: Use a hyperspectral camera to capture hyperspectral images of banana tree leaves to obtain hyperspectral image information of the leaves; S2: Remove the background from the image. Extract the 430nm and 725nm layers from the hyperspectral image and remove the background. S3: Perform difference fusion between the 725nm layer and the 430nm layer, and then perform image processing on the image by first dilation and then erosion. S4: Process the image using Gaussian blur; S5: Perform binarization segmentation on the image, and adaptively calculate the segmentation threshold using the maximum inter-class variance method to obtain a black and white image, where the leaves are white with a value of 255 and the background is black with a value of 0. S6: Then perform dilation and erosion on the image to obtain the position mask of the banana leaf; S7: Perform an AND operation between the obtained position mask and the 680nm and 725nm images to extract the banana leaf positions in the 680nm and 725nm images; The texture calculation includes filtering, gradient calculation, and image fusion of different bands on the feature spectral image after extracting the leaf position, ultimately obtaining the true leaf phenotypic texture information; the processing steps of the texture calculation are as follows: S1: The bilateral filtering algorithm is used to filter the images of the 680nm and 725nm bands respectively, smoothing the images, reducing noise, and preserving edge information; S2: Calculate the magnitude and direction of the image gradient for the 680nm and 725nm bands processed in S1, and calculate the edge position and edge intensity of the image respectively. S3: Use non-maximum suppression to find the edge with the strongest intensity among all edge information, and remove the edges with weaker intensity on both sides of the edge to make the edge more refined; S4: Use hysteresis thresholds to filter out high-intensity edges, medium-intensity edges, and low-intensity edges from all edge information; S5: After classifying the three types of edges, edges belonging to high intensity edges are identified as real edges, edges belonging to low intensity edges are eliminated, and medium intensity edges need to be judged a second time. If a medium intensity edge has an uninterrupted edge connecting to a high intensity edge, it is considered an edge; otherwise, it is eliminated. S6: After the images at 680nm and 725nm are processed by hysteresis thresholding, two edge texture maps are obtained respectively. The two edge texture maps are then weighted and fused, with each having a weight of 0.

5. S7: The weighted fused edge texture map is subjected to hysteresis thresholding again to obtain high-intensity edges, medium-intensity edges and low-intensity edges again. Among them, overlapping edges will continue to be classified as high-intensity edges because the intensity changes little. Non-overlapping edges will become medium-intensity edges or low-intensity edges after processing because the intensity is weakened. Low-intensity edges will be removed, and medium-intensity edges will be judged again to retain the edges that are connected to high-intensity edges. S8: The final image is a phenotypic texture map of banana leaves, which reflects the veins, edges, and folds within the leaves.

2. The method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging as described in claim 1, characterized in that, The image processing method of first dilating and then eroding the image is as follows: the edge of the blade is enlarged by dilation and the noise holes in the blade are filled. Then, the enlarged texture after dilation is shrunk by erosion.

3. The method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging as described in claim 1, characterized in that, Two thresholds, high and low, need to be calculated to distinguish between high-intensity, medium-intensity, and low-intensity edges. The threshold calculation is based on the idea of ​​maximum entropy. By calculating the entropy values ​​of the foreground and background after image segmentation, the segmentation threshold is determined. The higher the entropy of the foreground and background, the richer the information contained in the foreground and background of the image. The threshold used for image segmentation in this case serves as the high threshold for separating high-intensity and medium-intensity edges. The formula is as follows: (1) (2) (3) Where q is the threshold for segmentation, with a value ranging from 0 to 255, P0(q) represents the cumulative probability of background pixels, P1(q) represents the cumulative probability of foreground pixels, H0 represents the background entropy, H1 represents the foreground entropy, and H(q) represents the total entropy of the image when the threshold is q. The segmentation threshold q with the maximum entropy is selected as the high threshold TH through 256 iterations.

4. The method for extracting the phenotypic texture of banana wilt leaves based on hyperspectral imaging as described in claim 3, characterized in that, After obtaining the high threshold TH, the idea of ​​maximum inter-class variance is used to find the low threshold TL from the range of 0-TH to segment medium-intensity edges and low-intensity edges. The formula is as follows: (5) (6) i is the threshold value for segmentation, ranging from 0 to TH; ω0 is the proportion of foreground pixels in the image; ω1 is the proportion of background pixels in the image; µ0 is the average gray level of the foreground; µ1 is the average gray level of the background; and µ... i σ is the average gray level of the image when the segmentation threshold is i. i The inter-class variance is used to find the maximum inter-class variance through TH iterations. The segmentation threshold at this point is used as the lower limit threshold TL. Edges with an intensity higher than TH are classified as high-intensity edges, those with an intensity lower than TL are classified as low-intensity edges, and those in between are classified as medium-intensity edges.

Images