A rapid nondestructive detection method for rice grain mixed seeds based on hyperspectral imaging

Abstract

The present application relates to the technical field of rice grain detection, and particularly relates to a kind of rice grain mixed fast nondestructive testing method based on hyperspectral imaging, comprising the following steps: S1, collect rice hundred-grain plate hyperspectral image, carry out pre-processing and region of interest segmentation, extract the full-band spectral gray image of target area;S2, construct the multi-channel attention mechanism network model suitable for the nondestructive testing of rice grain mixed, input the full-band spectral gray image as the data set of classification training;The present application proposes a kind of rice grain mixed fast nondestructive testing method based on hyperspectral imaging, which fuses the spectral and spatial characteristics of rice grain, strengthens the spectral and spatial characteristic information of rice by multi-channel attention mechanism and self-attention mechanism module, further helps to improve the accuracy of rice grain classification, the multi-channel attention mechanism network model proposed in the present application fuses segmentation module and classification module, which facilitates to reduce the parameter amount when training high-resolution image.

Description

Technical Field

[0001] This invention relates to the field of rice grain detection technology, specifically to a rapid and non-destructive method for detecting mixed rice grains based on hyperspectral imaging. Background Technology

[0002] Since the 1980s, computer vision, spectroscopy and spectral imaging technologies, and electronic nose technologies have matured, significantly promoting the widespread application of hyperspectral technology in various fields. The combination of machine learning algorithms and hyperspectral imaging technology enables the rapid and accurate learning of effective spectral features from large amounts of data, thereby achieving efficient and non-destructive variety identification.

[0003] However, these methods rely too heavily on spectral information, and hyperspectral images have the characteristics of "same spectrum but different objects" and "same object but different spectra". With the continuous increase in various crop varieties, it has become difficult to accurately identify varieties based solely on spectral information.

[0004] Therefore, it is necessary to extract richer and more representative features to compensate for the shortcomings of single-spectral information in variety classification. In recent years, with the improvement of computing power, deep learning-based methods have been widely used in the field of hyperspectral seed identification.

[0005] However, most studies focus on seed identification of a single type, with fewer studies addressing the common real-world situation of mixed seed distribution. In such cases, seeds of different varieties are often mixed together, and directly classifying the collected mixed seed materials may lead to mixed features learned by the model, resulting in decreased accuracy.

[0006] Based on actual mixed-cropping situations, a multi-channel non-destructive detection model for rice grain varieties was constructed. Visible light hyperspectral imaging technology was used to obtain multi-channel spectral information and spatial characteristics of rice grains. Combined with a deep learning instance segmentation and classification model, effective detection of rice grain varieties can be achieved, providing a certain reference for subsequent protection of rice germplasm resources and breeding of improved varieties.

[0007] While existing technologies offer the advantages mentioned above, they also have disadvantages: Current research in crop seed variety identification primarily focuses on classifying individual varieties and morphological characteristics. However, in real-world scenarios, mixed seed distribution is common, and closely related seeds often appear similar, making morphological identification more challenging. Spectral information contains rich features but also significant redundancy, affecting model accuracy and efficiency. Furthermore, directly using mixed seed data for training and testing can lead to the model learning excessive spectral information, hindering training.

[0008] In conclusion, developing a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging remains a key issue that urgently needs to be addressed in the field of rice grain detection technology. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings in the existing technology for identifying crop seed varieties. Most studies focus on classifying individual varieties and morphological characteristics. However, in reality, mixed seed cultivation is common, and most closely related seeds look similar, making identification based solely on morphological features difficult. Spectral information contains rich features but also a lot of redundant information, which can affect the accuracy and efficiency of the model. Furthermore, directly using mixed seed data for training and testing can lead to the model learning too much spectral information, which is detrimental to model training.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] This invention provides a rapid and non-destructive method for detecting mixed rice grains based on hyperspectral imaging, comprising the following steps:

[0012] S1. Acquire hyperspectral images of 100 grains of rice, perform preprocessing and region of interest segmentation, and extract the full-band spectral grayscale image of the target region;

[0013] S2. Construct a multi-channel attention mechanism network model suitable for non-destructive detection of mixed rice grains, and input the full-band spectral grayscale image as the dataset for classification training;

[0014] S3. Using the trained multi-channel attention mechanism network model, the rice grain category information under mixed-crop conditions is detected based on the rice 100-grain plate hyperspectral image.

[0015] S4. Using the hyperspectral images of the rice grains at different sizes, input them into the trained target detection model and the multi-channel attention mechanism network model respectively for result testing and comparison.

[0016] Further, in step S1, the method for acquiring a hyperspectral image of 100 grains of rice, performing preprocessing and region-of-interest segmentation, and extracting the full-band spectral grayscale image of the target region is as follows:

[0017] The acquisition of 100-grain hyperspectral images of rice includes: randomly selecting rice grain samples of different varieties, dividing them into purebred rice grains and randomly mixed rice grains; acquiring 100-grain hyperspectral images of purebred rice grains and randomly mixed rice grains respectively using a hyperspectral imaging system in the 400-1000nm wavelength range; and defining continuous wavelength ranges. Corresponding wavenumber range: Set spatial coordinates as The acquired original hyperspectral image can then be represented as a ternary function:

[0018]

[0019] In the formula, Indicates at wavelength Lower space coordinates The type is Varieties are The hyperspectral image pixel values ​​corresponding to rice grains For the grain at wavelength The reflectivity function at that location satisfies , The spectral radiance of the light source, For the exposure time, It is additive Gaussian noise, and its variance varies with wavelength;

[0020] In step S1, the method for acquiring a hyperspectral image of 100 grains of rice, performing preprocessing and region of interest segmentation, and extracting the full-band spectral grayscale image of the target region is as follows:

[0021] The preprocessing and region of interest (ROI) segmentation include: using annotation tools to annotate the acquired hyperspectral images of the rice grain plate, distinguishing rice grains from the background, annotating semantic information, and using ROI segmentation technology, where the ROI is defined as the set of all grain pixels. Accurate segmentation is achieved using the level set method, and an evolution curve is defined. For the boundary of the region of interest, the energy functional is minimized:

[0022]

[0023] In the formula, Evolution curve The corresponding energy functional, The regularization parameter is used to control the smoothness of the evolution curve. It's a Dirac function. It is an evolution curve The magnitude of the gradient, The weighting parameters are used to adjust the proportion of terms related to the image gradient in the energy functional. It is an image With Gaussian kernel The magnitude of the gradient after convolution. It performs Gaussian smoothing on the image. The weighting parameter is used to adjust the proportion of terms related to the grain region indicator function in the energy functional. It is a semantic annotation function. The Heaviside function is used to divide the region around the evolution curve into inner and outer regions, select the target region for segmentation, and extract the full-band spectral grayscale image under 314 bands from the segmented target region.

[0024] In step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0025] The multi-channel attention mechanism network model includes: a segmentation module, a channel attention mechanism module, and a classification module;

[0026] The segmentation module uses the YOLO11-seg instance segmentation model to segment and train the input rice grain 100-grain plate image and its annotation information. After multiple iterations, it obtains the bounding box and mask of each rice grain in each image. The segmentation module outputs the first... The bounding box of each seed and binary mask ,in, With the center coordinates, Width and height, Let be the confidence level; its loss function is:

[0027]

[0028] In the formula, It is the total loss function of the segmentation module. The weighting coefficient for the location loss is used to adjust the proportion of the bounding box location prediction loss in the total loss. The generalized intersection-union loss function is used to evaluate the first... Predicted bounding box for each seed With the true bounding box The difference in position between them The weighting coefficient for the confidence loss is used to adjust the proportion of the bounding box confidence prediction loss in the total loss. The binary cross-entropy loss function is used to measure the first... Prediction confidence level of individual grains With true confidence The differences between them The weighting coefficient for the masking loss is used to adjust the proportion of masking prediction loss in the total loss. It is the DICE loss function, used to evaluate the first... Individual Seed Prediction Mask Compared to real masks The similarity between the grains is recorded, the coordinate frame information is recorded, and a unique ID number is assigned to each grain. The corresponding grayscale single-band image is read synchronously. At the same time, the image is cut out according to the recorded coordinate frame information to obtain a 314-channel image of a single rice grain.

[0029] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0030] In addition, the classification module with Transoformer as the backbone network introduces a channel attention mechanism at the input end; this mechanism consists of a global average pooling layer, two 1×1 convolutional layers, and LeakyReLU and Sigmoid activation functions.

[0031] Specifically, 314 multi-channel images of a single rice grain Perform global average pooling, expression:

[0032]

[0033] In the formula, This indicates the processing of a 314-channel image of a single rice grain. The resulting vector after global average pooling. Global average pooling is an abbreviation for a pooling operation. A 314-channel image of a single rice grain serves as input for the global average pooling operation. It is the number of pixels in the height direction of a single rice grain image. It is the number of pixels in the width direction of a single rice grain image. Image representing a single rice grain The Middle 314-dimensional channel data corresponding to each pixel position This represents the result after global average pooling. It is a 314-dimensional real vector.

[0034] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0035] It is important to note that the multi-channel image features weighted by the channel attention mechanism are fused and reduced in dimensionality through 1x1 convolution, and then further extracted spatial features and reduced the number of parameters through depthwise separable convolution. These features are then received by the Transformer encoder in the classification module for feature extraction and encoding; the output of the Transformer encoder is the first... The layer image feature encoding result is:

[0036]

[0037] In the formula, Indicates the Transformer encoder at the 1st... After layer processing Image feature encoding results corresponding to each rice grain A multilayer perceptron is a feedforward neural network containing multiple hidden layers used to perform further nonlinear transformations and processing on the input features. Layer normalization is a normalization technique that is used to... This part performs layer normalization. The Transformer encoder is in the... After layer processing Image feature encoding results corresponding to each rice grain Multi-head attention is an extended form of attention mechanism. It uses a multilayer perceptron for classification and discrimination, and then uses linear transformation and softmax activation to obtain the class probability mapping result of a single seed, outputting the class information of a single seed.

[0038] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0039] One judgment condition is set as follows: the multi-channel attention mechanism network model automatically identifies the segmentation and classification training status, and the weight sequence of the segmentation module is set as follows. ,in, (where the number of iterations is the threshold), define the optimal weight judgment condition:

[0040]

[0041] In the formula, This represents the optimal weight parameters for the segmentation modules. This indicates that the segmentation module is at the 1st... The weight parameters of the next iteration Find the parameter value that minimizes the subsequent loss function. The segmentation module is in the first... The loss function at the next iteration This means that the previous optimization is performed while the subsequent constraints are met. The segmentation module is in the first... The next iteration, with weight parameters as follows: The intersection-union ratio (IU) must be greater than or equal to the IU threshold. , The IoU threshold is set to [0.85-0.95], and the optimal segmentation module weights are read. At that time, the fixed segmentation module is triggered and used as a preprocessing tool for the classification module; input a purebred rice 100-grain plate image, and obtain multiple multi-channel rice single-grain images through the optimal segmentation module.

[0042] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0043] It is important to note that two batch_size values ​​are set. The first is used to determine the number of rice 100-grain images input to the segmentation module, and the second is used to determine the number of multi-channel single-grain rice data used for training the classification module. All single-grain rice images obtained by the segmentation module and their corresponding annotation information are randomly shuffled and then divided into datasets, merged into batches, and input into the classification module of the multi-channel attention mechanism network model for training and validation.

[0044] Furthermore, the classification module underwent multiple iterations, selecting the optimal classification module weight parameters based on classification accuracy, class precision and recall, and cross-entropy loss. The expression is:

[0045]

[0046] In the formula, This represents the weight parameters of the selected optimal classification module. Represents the weight parameters of the classification module Find the parameter value that maximizes the latter expression. The weighting coefficients are used to balance classification accuracy. and The impact of scores on selecting optimal weights It is the classification accuracy. For the sample All categories The predicted probability, For the sample The true category, The F1 score is the accuracy rate. and recall rate The harmonic mean, Represents the cross-entropy loss of the classification module. Must be less than or equal to the threshold , It is a validation set. The precision rate is calculated by dividing the number of true positives by the sum of true positives and false positives. The recall rate is calculated by dividing true positives by the sum of true positives and false negatives.

[0047] Further, in step S3, the method for detecting rice grain category information under mixed-crop conditions using the trained multi-channel attention mechanism network model and based on the rice 100-grain plate hyperspectral image is as follows:

[0048] The hyperspectral image of the hybrid rice 100-grain plate is input into the trained multi-channel attention mechanism network model to obtain the ID number and category information of each individual rice grain. The ID number and category information are then plotted above the corresponding individual rice grain in the original hyperspectral image of the rice 100-grain plate to form a text drawing area. This text drawing area is defined as the rectangular area above the coordinate frame of the ID number. ,expression:

[0049]

[0050] In the formula, Indicates the first The text drawing area corresponding to each rice grain These are the x and y coordinates of the pixels within the text drawing area. It is the first The x-coordinate of the center of the bounding box corresponds to each rice grain. It is the first The width of the bounding box corresponding to each rice grain. It is the first The center ordinate of the bounding box corresponding to each rice grain. It is the first The height of the bounding box corresponding to each rice grain It is the height of the text.

[0051] Furthermore, in step S4, the method for testing and comparing the results by inputting hyperspectral images of the rice 100-grain plate of different sizes into the trained target detection model and the multi-channel attention mechanism network model is as follows:

[0052] The target detection model is the YOLO11 target detection model. Purebred rice 100-grain plate data are input into the YOLO11 target detection model for iterative training. The optimal YOLO11 target detection model is saved. Two rice grain mixed images with different resolutions are selected and input into the trained YOLO11 target detection model and the multi-channel attention mechanism network model respectively for result detection. The grain detection rate, detection accuracy, and ratio of the YOLO11 target detection model and the multi-channel attention mechanism network model under the two different resolution images are calculated and used as the performance evaluation index of the YOLO11 target detection model and the multi-channel attention mechanism network model for result detection of rice grain mixed images at different resolutions.

[0053] Beneficial effects

[0054] Compared with known public technologies, the technical solution provided by this invention has the following beneficial effects:

[0055] The present invention proposes a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging. This method integrates the spectral and spatial characteristics of rice grains and enhances the spectral and spatial feature information of rice through a multi-channel attention mechanism network model and a multi-channel attention mechanism, thereby further improving the accuracy of rice grain classification.

[0056] The multi-channel attention mechanism network model proposed in this invention integrates segmentation and classification modules, which not only reduces the number of parameters during training of high-resolution images, but also facilitates the adjustment of the model structure, making it adaptable to different application scenarios and needs. Compared with existing object detection models, it has greater flexibility.

[0057] The multi-channel attention mechanism network model proposed in this invention integrates segmentation and classification modules, which helps the model to extract and identify rice grain features independently during training and testing. It is less susceptible to interference from features of other varieties and has stronger generalization ability on unknown mixed datasets. Attached Figure Description

[0058] Figure 1 This is a flowchart of a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging, according to the present invention.

[0059] Figure 2 The diagram shows the YOLO11 backbone structure, Transformer structure, and flowchart of the multi-channel attention mechanism for a rapid and non-destructive detection method for rice grain hybridization based on hyperspectral imaging, as presented in this invention. Figure 1 .

[0060] Figure 3The diagram shows the YOLO11 backbone structure, Transformer structure, and flowchart of the multi-channel attention mechanism for a rapid and non-destructive detection method for rice grain hybridization based on hyperspectral imaging, as presented in this invention. Figure 2 .

[0061] Figure 4 The diagram shows the YOLO11 backbone structure, Transformer structure, and flowchart of the multi-channel attention mechanism for a rapid and non-destructive detection method for rice grain hybridization based on hyperspectral imaging, as presented in this invention. Figure 3 .

[0062] Figure 5 This image shows the training results of the YOLO11-seg instance segmentation model for a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging, according to the present invention.

[0063] Figure 6 This is a training and evaluation index curve of a multi-variety rice grain classification model for a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging, according to the present invention.

[0064] Figure 7 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 1 .

[0065] Figure 8 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 2 .

[0066] Figure 9 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 3 .

[0067] Figure 10 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 4 .

[0068] Figure 11 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 5 .

[0069] Figure 12 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 6 .

[0070] Figure 13 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 7 .

[0071] Figure 14 This invention presents the prediction results of different-sized rice hybrid datasets on a model for a rapid and non-destructive detection method for rice grain hybrids based on hyperspectral imaging. Figure 8 . Detailed Implementation

[0072] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0073] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0074] The present invention will now be described in further detail with reference to the accompanying drawings:

[0075] Example 1:

[0076] like Figure 1 - Figure 14 As shown, this invention provides a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging, comprising the following steps:

[0077] S1. Acquire hyperspectral images of 100 grains of rice, perform preprocessing and region of interest segmentation, and extract the full-band spectral grayscale image of the target region;

[0078] As a preferred embodiment of the present invention, in step S1, the method for acquiring a hyperspectral image of a rice grain sample, performing preprocessing and region of interest segmentation, and extracting the full-band spectral grayscale image of the target region is as follows:

[0079] The acquisition of 100-grain hyperspectral images of rice includes: randomly selecting rice grain samples of different varieties, dividing them into purebred rice grains and randomly mixed rice grains; acquiring 100-grain hyperspectral images of purebred rice grains and randomly mixed rice grains respectively using a hyperspectral imaging system in the 400-1000nm wavelength range; and defining continuous wavelength ranges. Corresponding wavenumber range: Set spatial coordinates as The acquired original hyperspectral image can then be represented as a ternary function:

[0080]

[0081] In the formula, Indicates at wavelength Lower space coordinates The type is Varieties are The hyperspectral image pixel values ​​corresponding to rice grains For the grain at wavelength The reflectivity function at that location satisfies , The spectral radiance of the light source, For the exposure time, It is additive Gaussian noise, and its variance varies with wavelength;

[0082] More specifically, in step S1, the method for acquiring a hyperspectral image of 100 grains of rice, performing preprocessing and region-of-interest segmentation, and extracting the full-band spectral grayscale image of the target region is as follows:

[0083] The preprocessing and region of interest (ROI) segmentation include: using annotation tools to annotate the acquired hyperspectral images of the rice grain plate, distinguishing rice grains from the background, annotating semantic information, and using ROI segmentation technology, where the ROI is defined as the set of all grain pixels. Accurate segmentation is achieved using the level set method, and an evolution curve is defined. For the boundary of the region of interest, the energy functional is minimized:

[0084]

[0085] In the formula, Evolution curve The corresponding energy functional, The regularization parameter is used to control the smoothness of the evolution curve. It's a Dirac function. It is an evolution curve The magnitude of the gradient, The weighting parameters are used to adjust the proportion of terms related to the image gradient in the energy functional. It is an image With Gaussian kernel The magnitude of the gradient after convolution. It performs Gaussian smoothing on the image. The weighting parameter is used to adjust the proportion of terms related to the grain region indicator function in the energy functional. It is a semantic annotation function. The Heaviside function is used to divide the region around the evolution curve into inner and outer regions, select the target region for segmentation, and extract the full-band spectral grayscale image under 314 bands from the segmented target region.

[0086] In this embodiment, the 400-1000nm band covers key spectral feature regions of rice grains, such as chlorophyll absorption peaks and starch reflectance peaks, and distinguishes between purebred and mixed samples. This provides diverse data for subsequent training of multi-channel attention mechanism network models, facilitating the standardization and comprehensiveness of data collection. The level set method, through energy functional optimization, facilitates accurate separation of grains and background, avoiding background noise interference with subsequent analysis and improving the high precision of segmentation. The extracted full-band spectral grayscale image retains the essential spectral features of the grains. Compared with traditional RGB images, it can more effectively distinguish varieties with similar morphology but different compositions, providing high-quality input for subsequent multi-channel attention mechanism models and improving the accuracy and reliability of mixed detection.

[0087] S2. Construct a multi-channel attention mechanism network model suitable for non-destructive detection of mixed rice grains, and input the full-band spectral grayscale image as the dataset for classification training;

[0088] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0089] The multi-channel attention mechanism network model includes: a segmentation module, a channel attention mechanism module, and a classification module;

[0090] The segmentation module uses the YOLO11-seg instance segmentation model to segment and train the input rice grain 100-grain plate image and its annotation information. After multiple iterations, it obtains the bounding box and mask of each rice grain in each image. The segmentation module outputs the first... The bounding box of each seed and binary mask ,in, With the center coordinates, Width and height, Let be the confidence level; its loss function is:

[0091]

[0092] In the formula, It is the total loss function of the segmentation module. The weighting coefficient for the location loss is used to adjust the proportion of the bounding box location prediction loss in the total loss. The generalized intersection-union loss function is used to evaluate the first... Predicted bounding box for each seed With the true bounding box The difference in position between them The weighting coefficient for the confidence loss is used to adjust the proportion of the bounding box confidence prediction loss in the total loss. The binary cross-entropy loss function is used to measure the first... Prediction confidence level of individual grains With true confidence The differences between them The weighting coefficient for the masking loss is used to adjust the proportion of masking prediction loss in the total loss. It is the DICE loss function, used to evaluate the first... Individual Seed Prediction Mask Compared to real masks The similarity between the grains is recorded, the coordinate frame information is recorded, and a unique ID number is assigned to each grain. The corresponding grayscale single-band image is read synchronously. At the same time, the image is cut out according to the recorded coordinate frame information to obtain a 314 multi-channel image of a single rice grain.

[0093] More specifically, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0094] In addition, the classification module with Transoformer as the backbone network introduces a channel attention mechanism at the input end; this mechanism consists of a global average pooling layer, two 1×1 convolutional layers, and LeakyReLU and Sigmoid activation functions.

[0095] Specifically, 314 multi-channel images of a single rice grain Perform global average pooling, expression:

[0096]

[0097] In the formula, This indicates the processing of a 314-channel image of a single rice grain. The resulting vector after global average pooling. Global average pooling is an abbreviation for a pooling operation. A 314-channel image of a single rice grain serves as input for the global average pooling operation. It is the number of pixels in the height direction of a single rice grain image. It is the number of pixels in the width direction of a single rice grain image. Image representing a single rice grain The Middle 314-dimensional channel data corresponding to each pixel position This represents the result after global average pooling. It is a 314-dimensional real vector;

[0098] As a preferred embodiment of the present invention, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of mixed rice grains is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0099] It is important to note that the multi-channel image features weighted by the channel attention mechanism are fused and reduced in dimensionality through 1x1 convolution, and then further extracted spatial features and reduced the number of parameters through depthwise separable convolution. These features are then received by the Transformer encoder in the classification module for feature extraction and encoding; the output of the Transformer encoder is the first... The layer image feature encoding result is:

[0100]

[0101] In the formula, Indicates the Transformer encoder at the 1st... After layer processing Image feature encoding results corresponding to each rice grain A multilayer perceptron is a feedforward neural network containing multiple hidden layers used to perform further nonlinear transformations and processing on the input features. Layer normalization is a normalization technique that is used to... This part performs layer normalization. The Transformer encoder is in the... After layer processing Image feature encoding results corresponding to each rice grain Multi-head attention is an extended form of attention mechanism. It uses a multilayer perceptron for classification and discrimination, and then uses linear transformation and softmax activation to obtain the class probability mapping result of a single seed, and outputs the class information of a single seed.

[0102] More specifically, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0103] One judgment condition is set as follows: the multi-channel attention mechanism network model automatically identifies the segmentation and classification training status, and the weight sequence of the segmentation module is set as follows. ,in, (where the number of iterations is the threshold), define the optimal weight judgment condition:

[0104]

[0105] In the formula, This represents the optimal weight parameters for the segmentation modules. This indicates that the segmentation module is at the 1st... The weight parameters of the next iteration Find the parameter value that minimizes the subsequent loss function. The segmentation module is in the first... The loss function at the next iteration This means that the previous optimization is performed while the subsequent constraints are met. The segmentation module is in the first... The next iteration, with weight parameters as follows: The intersection-union ratio (IU) must be greater than or equal to the IU threshold. , The IoU threshold is set to [0.85-0.95], and the optimal segmentation module weights are read. At that time, the fixed segmentation module is triggered and used as a preprocessing tool for the classification module; input a purebred rice 100-grain plate image, and obtain multiple multi-channel rice single-grain images through the optimal segmentation module;

[0106] Further, in step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed, and the full-band spectral grayscale image is input as the dataset for classification training; the method is as follows:

[0107] It is important to note that two batch_size values ​​are set. The first is used to determine the number of rice 100-grain images input to the segmentation module, and the second is used to determine the number of multi-channel single-grain rice data used for training the classification module. All single-grain rice images obtained by the segmentation module and their corresponding annotation information are randomly shuffled and then divided into datasets, merged into batches, and input into the classification module of the multi-channel attention mechanism network model for training and validation.

[0108] Furthermore, the classification module underwent multiple iterations, selecting the optimal classification module weight parameters based on classification accuracy, class precision and recall, and cross-entropy loss. The expression is:

[0109]

[0110] In the formula, This represents the weight parameters of the selected optimal classification module. Represents the weight parameters of the classification module Find the parameter value that maximizes the latter expression. The weighting coefficients are used to balance classification accuracy. and The impact of scores on selecting optimal weights It is the classification accuracy. For the sample All categories The predicted probability, For the sample The true category, The F1 score is the accuracy rate. and recall rate The harmonic mean, Represents the cross-entropy loss of the classification module. Must be less than or equal to the threshold , It is a validation set. The precision rate is calculated by dividing the number of true positives by the sum of true positives and false positives. The recall rate is calculated by dividing true positives by the sum of true positives and false negatives.

[0111] In this embodiment, a multi-module collaborative design balances the advantages of accurate localization and classification. The refined loss function of the segmentation module helps avoid grain omissions and overlaps. The attention mechanism focuses on key spectral features, which helps solve the problem of hyperspectral data redundancy. The deep encoding capability of the classification module improves variety differentiation, especially for morphologically similar varieties. An automatic selection strategy for optimal weights avoids overfitting. The fixed segmentation module is used as a preprocessing tool, which reduces training costs and ensures the consistency of input to the classification module. The dual batch size setting facilitates a balance between efficiency and accuracy, making it suitable for large-scale breeding sample detection, such as processing thousands of 100-grain plates daily. This helps improve the accuracy of mixed-type detection while shortening the detection time for a single sample, providing efficient and reliable technical support for rice variety purity identification and breeding screening.

[0112] S3. Using the trained multi-channel attention mechanism network model, the rice grain category information under mixed-crop conditions is detected based on the rice 100-grain plate hyperspectral image.

[0113] As a preferred embodiment of the present invention, in step S3, the method for detecting rice grain category information under mixed-crop conditions using the trained multi-channel attention mechanism network model and based on the rice 100-grain plate hyperspectral image is as follows:

[0114] The hyperspectral image of the hybrid rice 100-grain plate is input into the trained multi-channel attention mechanism network model to obtain the ID number and category information of each individual rice grain. The ID number and category information are then plotted above the corresponding individual rice grain in the original hyperspectral image of the rice 100-grain plate to form a text drawing area. This text drawing area is defined as the rectangular area above the coordinate frame of the ID number. ,expression:

[0115]

[0116] In the formula, Indicates the first The text drawing area corresponding to each rice grain These are the x and y coordinates of the pixels within the text drawing area. It is the first The x-coordinate of the center of the bounding box corresponds to each rice grain. It is the first The width of the bounding box corresponding to each rice grain. It is the first The center ordinate of the bounding box corresponding to each rice grain. It is the first The height of the bounding box corresponding to each rice grain It is the height of the text;

[0117] In this embodiment, automated detection significantly improves efficiency. Compared to manual grain-by-grain identification, this invention completes the detection of the entire plate within 10 seconds, while avoiding human error. Precise visual annotation facilitates subsequent analysis; breeders can trace grain locations using ID numbers and quickly calculate hybridization ratios based on category information, such as 32% japonica rice in this sample, providing quantitative evidence for varietal purity assessment. Standardized text area settings enhance annotation consistency, avoiding text overlap or obscuring grain features, preserving original image information while clearly presenting detection results, facilitating subsequent data archiving and secondary analysis. This approach is suitable for large-scale breeding screening scenarios, reducing detection costs and improving result reliability.

[0118] S4. Using the hyperspectral images of the rice grains from the 100-grain plate of different sizes, input them into the trained target detection model and the multi-channel attention mechanism network model respectively to test and compare the results;

[0119] As a preferred embodiment of the present invention, in step S4, the method for testing and comparing the results of the rice 100-grain plate hyperspectral images of different sizes by inputting them into the trained target detection model and the multi-channel attention mechanism network model is as follows:

[0120] The target detection model is the YOLO11 target detection model. Purebred rice 100-grain plate data are input into the YOLO11 target detection model for iterative training. The optimal YOLO11 target detection model is saved. Two rice grain mixed images with different resolutions are selected and input into the trained YOLO11 target detection model and the multi-channel attention mechanism network model respectively for result detection. The grain detection rate, detection accuracy, and ratio of the YOLO11 target detection model and the multi-channel attention mechanism network model under the two different resolution images are calculated and used as the performance evaluation index of the YOLO11 target detection model and the multi-channel attention mechanism network model for result detection of rice grain mixed images at different resolutions.

[0121] In this embodiment, by comparing with the optimal YOLO11 target detection model, the advantages of the multi-channel attention mechanism network model in the rice grain mixed-species detection scenario can be objectively verified, avoiding the problem of "the performance of self-developed models cannot be quantified". Through cross-resolution testing, the actual application needs are met. Field detection often obtains low-resolution images due to equipment and environmental limitations, while laboratory detection requires high-precision images. By comparing the indicators under the two resolutions, the applicability of the multi-channel attention mechanism network model in different scenarios can be clarified, highlighting its advantage of maintaining high detection accuracy in low-resolution images. This provides data support for subsequent practical promotion, such as portable field detection equipment and high-precision laboratory analysis systems, while reducing planting losses or quality misjudgments caused by variety mixing.

[0122] Example 2:

[0123] like Figure 1 - Figure 14 As shown, this invention provides a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging, comprising the following steps:

[0124] The dataset used in this invention is divided into two parts: one part is the data of a hundred-grain plate made up of pure rice grains, and the other part is the data of a hundred-grain plate made up of a mixture of different varieties of rice grains.

[0125] Each image contains 60 rice grains. A total of 200 images of 100-grain plates of rice from 20 different varieties were collected, with 1200 grains per rice grain. The rice grains were distinguished from the background using the ISAT semi-automatic annotation tool to obtain annotation information.

[0126] The multi-channel attention mechanism network model proposed in this invention, suitable for non-destructive testing of mixed rice grains, integrates a segmentation module and a classification module. The segmentation module is used to acquire single rice grain data and corresponding 314-band grayscale images from a 100-grain plate. The classification module uses the VisionTransformer network as its backbone, removes the patch bedding module, and adds a channel attention mechanism module to learn the spectral characteristics of rice grains.

[0127] In this invention, a multi-channel attention mechanism network model is trained by collecting images of 100 grains of purebred rice and annotating semantic information data.

[0128] The training was conducted on a Linux system provided by Ubuntu 20.04LTS. The CPU used was a 16vCPU Intel Xeon Gold 6430 (Sapphire Rapids architecture, 2.1GHz clock speed, 60MB L3 cache), and the GPU was an NVIDIA RTX 4090 graphics card with 24GB of video memory and 16384 CUDA cores. Parallel computation of the deep learning model was achieved based on the CUDA 11.8 acceleration library.

[0129] The multi-channel attention mechanism network model is implemented using the PyTorch deep learning framework and Python language. The multi-channel attention mechanism network model reads the 100-grain plate image of rice grains, and obtains multi-channel single-grain rice image data and the position ID of each rice grain in the image through the segmentation module. All the segmented grains are randomly mixed and the dataset is divided. The model is trained according to the set batch size and the number of iterations is set to 400. After multiple rounds of training, the trained weights are obtained and the segmentation module is fixed and used as a preprocessing tool for the classification module, and then used for inference testing.

[0130] Based on the collected images of 100 grains of mixed rice, the data is input into a trained multi-channel attention network model to detect the rice grain category information under mixed conditions. The recorded location ID and category information are then returned to the original image, achieving non-destructive detection of mixed rice varieties.

[0131] To compare the performance of the model proposed in this invention, mixed rice images of different sizes were used to test the trained target detection model, and the prediction results were compared with those of a rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging proposed in this invention.

[0132] The YOLO11 target detection model was selected. After labeling the rice grain 100-grain plate data, the YOLO11 target detection model was trained for 1000 rounds of iteration until it converged. The optimal YOLO11 target detection model was then saved.

[0133] Rice hybrid images of different resolutions were selected and input into two trained models (the optimal YOLO11 object detection model and the trained multi-channel attention mechanism network model). The performance of the two models was compared based on their detection rate, detection accuracy, and the ratio between the two models.

[0134] The results show that the trained multi-channel attention mechanism network model and the optimal YOLO11 object detection model have comparable prediction performance on simple tasks. For rice grains with similar appearances, the trained multi-channel attention mechanism network model has an advantage. However, for rice grains with more distinct appearance features, the optimal YOLO11 object detection model shows superior performance. For the prediction of sample images of different resolutions, the trained multi-channel attention mechanism network model has stronger generalization ability and is better able to handle image compression. Furthermore, the independent classification module helps improve the accuracy of the trained multi-channel attention mechanism network model, while YOLO11 is significantly affected.

[0135] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rapid and non-destructive method for detecting mixed rice grains based on hyperspectral imaging, characterized in that, Includes the following steps: S1. Acquire hyperspectral images of 100 grains of rice, perform preprocessing and region of interest segmentation, and extract the full-band spectral grayscale image of the target region; S2. Construct a multi-channel attention mechanism network model suitable for non-destructive detection of mixed rice grains, using the full-band spectral grayscale image as the dataset for classification training. The multi-channel attention mechanism network model includes a segmentation module, a channel attention mechanism module, and a classification module. The classification module, with Transoformer as the backbone network, introduces a channel attention mechanism at the input. This mechanism consists of a global average pooling layer, two 1×1 convolutional layers, and LeakyReLU and Sigmoid activation functions. A judgment condition is set as follows: the multi-channel attention mechanism network model automatically identifies the segmentation and classification training status. The weight sequence of the segmentation module is set as follows. ,in, Define the optimal weight judgment condition for the number of iterations: ， In the formula, This represents the optimal segmentation module weight parameters. This indicates that the segmentation module is at the 1st... The weight parameters of the next iteration Find the parameter value that minimizes the subsequent loss function. The segmentation module is in the first... The loss function at the next iteration This means that the previous optimization is performed while the subsequent constraints are met. The segmentation module is in the first... The next iteration, with weight parameters as follows: The intersection-union ratio (IU) must be greater than or equal to the IU threshold. , The IoU threshold is set to 0.9, and the optimal segmentation module weights are read. At that time, the fixed segmentation module is triggered and used as a preprocessing tool for the classification module; input a purebred rice 100-grain plate image, and obtain multiple multi-channel rice single-grain images through the optimal segmentation module; S3. Using the trained multi-channel attention mechanism network model, the rice grain category information under mixed-crop conditions is detected based on the rice 100-grain plate hyperspectral image. S4. Using the rice grain 100-grain hyperspectral images of different sizes, input them into the trained target detection model and multi-channel attention mechanism network model respectively for result testing and comparison.

2. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 1, characterized in that, In step S1, the method for acquiring a hyperspectral image of 100 grains of rice, performing preprocessing and region of interest segmentation, and extracting the full-band spectral grayscale image of the target region is as follows: The acquisition of 100-grain hyperspectral images of rice includes: randomly selecting rice grain samples of different varieties, dividing them into purebred rice grains and randomly mixed rice grains; acquiring 100-grain hyperspectral images of purebred rice grains and randomly mixed rice grains respectively using a hyperspectral imaging system in the 400-1000nm wavelength range; and defining continuous wavelength ranges. Corresponding wavenumber range: Set spatial coordinates as The acquired original hyperspectral image can then be represented as a ternary function: ， In the formula, Indicates at wavelength Lower space coordinates The type is Varieties are The hyperspectral image pixel values ​​corresponding to rice grains For the grain at wavelength The reflectivity function at that location satisfies , The spectral radiance of the light source, For the exposure time, It is additive Gaussian noise, and its variance varies with wavelength.

3. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 2, characterized in that, Step S1 also includes: The preprocessing and region of interest (ROI) segmentation include: using annotation tools to annotate the acquired hyperspectral images of the rice grain plate, distinguishing rice grains from the background, annotating semantic information, and using ROI segmentation technology, where the ROI is defined as the set of all grain pixels. Accurate segmentation is achieved using the level set method, and an evolution curve is defined. For the boundary of the region of interest, the energy functional is minimized: ， In the formula, Evolution curve The corresponding energy functional, The regularization parameter is used to control the smoothness of the evolution curve. It's a Dirac function. It is an evolution curve The magnitude of the gradient, The weighting parameters are used to adjust the proportion of terms related to the image gradient in the energy functional. It is an image With Gaussian kernel The magnitude of the gradient after convolution. It performs Gaussian smoothing on the image. The weighting parameter is used to adjust the proportion of terms related to the grain region indicator function in the energy functional. It is a semantic annotation function. The Heaviside function is used to divide the region around the evolution curve into inner and outer regions, select the target region for segmentation, and extract the full-band spectral grayscale image of 314 bands from the segmented target region.

4. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 3, characterized in that, In step S2, a multi-channel attention mechanism network model suitable for non-destructive detection of rice grain hybridization is constructed. The method for inputting the full-band spectral grayscale image as the dataset for classification training is as follows: The segmentation module is trained on the input rice grain 100-grain plate image and annotation information based on the YOLO11-seg instance segmentation model. After multiple iterations, the position coordinate box and mask of each rice grain in each image are obtained. The segmentation module outputs the first... The bounding box of each seed and binary mask ,in, With the center coordinates, Width and height, Let be the confidence level; its loss function is: ， In the formula, It is the total loss function of the segmentation module. The weighting coefficient for the location loss is used to adjust the proportion of the bounding box location prediction loss in the total loss. The generalized intersection-union loss function is used to evaluate the first... Predicted bounding box for each seed With the true bounding box The difference in position between them The weighting coefficient for the confidence loss is used to adjust the proportion of the bounding box confidence prediction loss in the total loss. The binary cross-entropy loss function is used to measure the first... Prediction confidence level of individual grains With true confidence The differences between them The weighting coefficient for the masking loss is used to adjust the proportion of masking prediction loss in the total loss. It is the DICE loss function, used to evaluate the first... Individual Seed Prediction Mask Compared to real masks The similarity between the grains is recorded, the coordinate frame information is recorded, and a unique ID number is assigned to each grain. The corresponding grayscale single-band image is read synchronously. At the same time, the image is cut out according to the recorded coordinate frame information to obtain a 314-channel image of a single rice grain.

5. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 4, characterized in that, Step S2 also includes: 314-channel images of a single rice grain Perform global average pooling, expression: ， In the formula, Represents a 314-channel image of a single rice grain. The resulting vector after global average pooling. This is a global average pooling operation. A 314-channel image of a single rice grain serves as input for the global average pooling operation. It is the number of pixels in the height direction of a single rice grain image. It is the number of pixels in the width direction of a single rice grain image. Image representing a single rice grain The Middle 314-dimensional channel data corresponding to each pixel position This represents the result after global average pooling. It is a 314-dimensional real vector.

6. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 5, characterized in that, Step S2 also includes: Multi-channel image features weighted by a channel attention mechanism are fused and reduced in dimensionality across channels using 1x1 convolutions, followed by depthwise separable convolutions to further extract spatial features and reduce the number of parameters. These features are then received by the Transformer encoder in the classification module for feature extraction and encoding. The output of the Transformer encoder is the first... The layer image feature encoding result is: ， In the formula, Indicates the Transformer encoder at the 1st... After layer processing Image feature encoding results corresponding to each rice grain It is a feedforward neural network containing multiple hidden layers, used to perform further nonlinear transformations and processing on the input features. It is a normalization technique, which is used for... This part undergoes a normalization operation. The Transformer encoder is in the... After layer processing Image feature encoding results corresponding to each rice grain For multi-head attention, classification is performed using a multilayer perceptron, followed by linear transformation and softmax activation to obtain the class probability mapping result of a single seed, and the class information of a single seed is output.

7. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 6, characterized in that, Step S2 also includes: Two batch_size values ​​are set. The first is used to determine the rice 100-grain plate images input to the segmentation module, and the second is used to determine the number of multi-channel single-grain rice data used for training the classification module. All single-grain rice images obtained by the segmentation module and their corresponding annotation information are randomly shuffled and divided into datasets, which are then merged into batches and input into the classification module of the multi-channel attention mechanism network model for training and validation. The classification module undergoes multiple iterations, selecting the optimal classification module weight parameters based on classification accuracy, class precision and recall, and cross-entropy loss. The expression is: ， In the formula, This represents the weight parameters of the selected optimal classification module. Represents the weight parameters of the classification module Find the parameter value that maximizes the latter expression. The weighting coefficients are used to balance classification accuracy. and The impact of scores on selecting optimal weights It is the classification accuracy. For the sample All categories The predicted probability, For the sample The true category, It's the F1 score, which is the accuracy rate. and recall rate The harmonic mean, Represents the cross-entropy loss of the classification module. Must be less than or equal to the threshold , It is a validation set. It's the accuracy rate, calculated using the formula for true positives. Divide by true positive False positives The sum of It is the recall rate, calculated using the formula for true positives. Divide by true positive False negative sum.

8. The rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 7, characterized in that, In step S3, the method for detecting rice grain category information under mixed-crop conditions using the trained multi-channel attention mechanism network model and the hyperspectral image of the rice 100-grain plate is as follows: The hyperspectral image of the hybrid rice 100-grain plate is input into the trained multi-channel attention mechanism network model to obtain the ID number and category information of each individual rice grain. The ID number and category information are then plotted above the corresponding individual rice grain in the original hyperspectral image of the rice 100-grain plate to form a text drawing area. This text drawing area is defined as the rectangular area above the coordinate frame of the ID number. ,expression: In the formula, Indicates the first The text drawing area corresponding to each rice grain These are the x and y coordinates of the pixels within the text drawing area. It is the first The x-coordinate of the center of the bounding box corresponds to each rice grain. It is the first The width of the bounding box corresponding to each rice grain. It is the first The center ordinate of the bounding box corresponding to each rice grain. It is the first The height of the bounding box corresponding to each rice grain It is the height of the text.

9. A rapid and non-destructive detection method for mixed rice grains based on hyperspectral imaging according to claim 8, characterized in that, In step S4, the method for testing and comparing the results of the rice 100-grain plate hyperspectral images of different sizes, respectively, is as follows: The target detection model is the YOLO11 target detection model. Rice grain data (100-grain plate data) is input into the YOLO11 target detection model for iterative training. The optimal YOLO11 target detection model is saved. Two rice grain mixed images at different resolutions are selected and input into the trained YOLO11 target detection model and the multi-channel attention mechanism network model respectively for result detection. The grain detection rate, detection accuracy, and ratio of the YOLO11 target detection model and the multi-channel attention mechanism network model are calculated for the two different resolution images. These ratios are used as performance evaluation indicators for the YOLO11 target detection model and the multi-channel attention mechanism network model in detecting rice grain mixed images at different resolutions.

Images