Multi-dimensional seed quality evaluation method and device

Abstract

The invention discloses a multi-dimensional seed quality evaluation method and device. The problems that an existing seed quality evaluation method is single, low in efficiency and limited in accuracy are solved. The method comprises the following steps: collecting a seed optical image and an X-ray image; segmenting to obtain a single seed optical image and an X-ray image; seed appearance characteristic parameters are calculated according to the single-seed optical image, and seed internal structure parameters are calculated according to the single-seed X-ray image; inputting the single-seed optical image and the X-ray image into a seed vigor identification model, and outputting vigor indexes of each seed; and obtaining a seed quality score according to the appearance characteristic parameters, the internal structure parameters and the vitality indexes of the seeds. According to the method, the seed quality evaluation is output in combination with the data of multiple dimensions, the evaluation dimensions are more, the evaluation angle is more comprehensive, and the quality evaluation is more accurate and objective.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seed quality evaluation, in particular to a multi-dimensional seed quality evaluation method and device. BACKGROUND

[0002] According to the national standard, the germination rate of vegetable seeds is about 85%-90%. If the germination rate does not meet the standard, it will not meet the technical needs of intensive and factory seedling of vegetables. At the same time, it is not conducive to mechanized sowing and later automatic and intelligent management. In order to achieve intensive and intelligent management, it is necessary to increase the cost of artificial management during seedling stage and increase the substrate and facilities.

[0003] To improve the germination rate of seeds, it is necessary to accurately evaluate the seeds to automatically screen and improve the quality of seeds. The existing seed quality evaluation methods mainly include: 1. Physical measurement method, 2. Physiological and biochemical measurement method, and 3. Seed vigor detection. These methods are low in efficiency and accuracy. On the other hand, the existing seed evaluation devices on the market have single detection dimension, that is, some products simply detect the size, length-width ratio, color and weight of seeds, and some can only measure the purity of seeds, but cannot achieve multi-dimensional detection of seed quality. For example, the application number 202211369888.7, the name of the peanut seed selection evaluation and grading method based on network model, identifies the seed type according to the seed appearance characteristics, and then evaluates the seed quality. This patent has the problem of single detection dimension and limited quality detection accuracy. The existing detection methods and devices cannot meet the current market demand for high-quality seeds, and there is an urgent need for an evaluation method and device that can integrate multi-dimensional image information and automatically extract seed quality characteristics to realize rapid, objective and comprehensive non-destructive detection of seed quality. SUMMARY

[0004] The present application mainly solves the problems of single, low efficiency and limited accuracy of the existing seed quality evaluation method, and provides a multi-dimensional seed quality evaluation method and device.

[0005] The above technical problems of the present application are mainly solved by the following technical scheme: a multi-dimensional seed quality evaluation method, comprising the following steps: Collecting seed optical images and X-ray images; Segmenting to obtain single-seed optical images and X-ray images; Calculating seed appearance feature parameters according to single-seed optical images and calculating seed internal structure parameters according to single-seed X-ray images; Inputting single-seed optical images and X-ray images into a seed vigor recognition model to output various seed vigor indexes; Obtaining seed quality scores according to seed appearance feature parameters, internal structure parameters and vigor indexes.

[0006] The application calculates appearance characteristic parameters by taking pictures of the seed outside through a camera, calculates internal structure parameters by taking pictures of the embryo and endosperm of the seed inside through an X-ray machine, and obtains seed vigor indexes through a deep learning method, and combines multiple dimension data to output seed quality evaluation together, so that the evaluation dimension is more, the evaluation angle is more comprehensive, and the quality evaluation is more accurate and objective. The technical problems of single seed quality evaluation, low efficiency and limited accuracy in the prior art are solved.

[0007] As a preferred scheme, the seed optical image is identified and segmented through a semantic segmentation network to obtain a plurality of single-seed optical images.

[0008] The seed optical image collected by the camera is a multi-seed optical image, and the semantic segmentation network is needed to identify and locate a plurality of seeds in the multi-seed optical image, and the identified seeds are segmented to generate a plurality of single-seed optical images. The semantic segmentation network of the application preferably adopts a pixel-level segmentation model based on a U-net convolutional neural network. The model extracts multi-scale image features through an encoder, combines a decoder for high-resolution recovery, and uses a Softmax layer to output the class label of each pixel, so as to realize accurate identification and segmentation of the seed region. The model has good edge detection capability and spatial detail preservation capability, and can accurately extract the boundary profile of each seed in a complex background.

[0009] As a preferred scheme, the seed X-ray image and a plurality of single-seed optical images are generated into a plurality of single-seed X-ray images through an association algorithm.

[0010] The seed X-ray image collected by the X-ray machine is a multi-seed X-ray image, and a plurality of seeds in the multi-seed X-ray image need to be identified and located, and the identified seeds are segmented to obtain a plurality of single-seed X-ray images. The application combines the seed X-ray image with a plurality of single-seed optical images, transfers the seed positioning information through an association algorithm, and generates a plurality of single-seed X-ray images from the multi-seed X-ray image. In this scheme, the association algorithm preferably adopts a method combining ORB feature extraction and matching with RANSAC affine transformation estimation to complete spatial mapping and positioning projection between different modal images, ensure one-to-one correspondence between the single X-ray image and the target optical image, and ensure the accuracy of subsequent parameter calculation and vigor index evaluation.

[0011] As a preferred scheme, the seed appearance characteristic parameters are calculated according to the single-seed optical image, including: The seed basic geometry and structure features are extracted from the single-seed optical image, the area, defect area, size, aspect ratio of the seed in the single-seed optical image are calculated, and the seed color is identified.

[0012] The calculation of the seed appearance feature parameters can be performed by using the set appearance feature recognition network, which can extract the basic geometric and structural features of the seed, calculate the area and defect area of a single seed, identify the color of the seed, and calculate the size and aspect ratio of the seed by using the minimum circumscribed matrix and minimum circumscribed circle of OpenCV.

[0013] As a preferred solution, the calculation of the internal structure parameters of the single seed from the X-ray image of the single seed comprises: The embryo and endosperm regions are separated and extracted from the X-ray image of the single seed by using the image gray scale-based analysis method, the area proportion of the embryo and endosperm is calculated, and the internal transparency feature of the seed is obtained by using the image overall gray value mean and distribution method.

[0014] The present solution mainly calculates the internal structure parameters of the seed, the area proportion of the embryo and endosperm is calculated by analyzing the embryo and endosperm regions in the X-ray image of the single seed, and the median internal transparency is determined by analyzing the gray value.

[0015] As a preferred solution, the seed vigor recognition model comprises two input layers, the two input layers are connected to the first branch and the second branch respectively, the first and second branches are connected to the fusion layer, the fusion layer is connected to the output layer through the full connection layer, and the first and second branches each comprise three layers of convolution layers, a spatial attention mechanism, a reshaping layer, a 3D convolution layer and a temporal attention mechanism.

[0016] The two input layers are respectively used for inputting the single seed optical image and the single seed X-ray image, the input layer is inputted with a video frame sequence, it is assumed that each video frame sequence has T frames, and the size of each frame is HxWxC (heightxwidthxchannel number).

[0017] Convolution layer: 2D local feature extraction layer (Local Feature Layer) is used to extract local spatial features; the 2D local feature extraction layer is composed of Con2D+BatchNorm+ReLU activation function.

[0018] Spatial attention mechanism: a spatial attention mechanism is added after the convolution layer to calculate the attention weight of each feature map. Specifically, a 1x1 convolution layer is used to produce a spatial attention map, and the spatial attention map is multiplied with the convolution feature map to emphasize important regions.

[0019] Reshaping layer: the dimension of the input is changed, and the shape of the input data is adjusted to a specified shape.

[0020] 3D convolution layer: the time dimension feature is extracted, and a bidirectional long short-term memory network (Bi-LSTM) can also be used instead.

[0021] Temporal attention mechanism: After the extraction of temporal features, the temporal attention mechanism is applied to calculate the attention weight of the time step, and the temporal attention map is multiplied with the temporal features to emphasize important time steps.

[0022] Fusion layer: the features of the two branches are fused.

[0023] Fully connected layer: the extracted features are flattened and classified or regressed through the fully connected layer. The Dropout layer is used to prevent overfitting.

[0024] Output layer: according to the task requirements, the vitality index is output. In this scheme, the vitality index is the vitality grade or vitality score.

[0025] As a preferred scheme, The seed vitality recognition model is input with a single seed optical image and an X-ray image, and the extracted vitality index includes multi-level deep features for representing seed vitality.

[0026] In another scheme, multi-level deep features capable of representing seed vitality are extracted from a single seed optical image and a single seed X-ray image by a seed vitality recognition model. The multi-level deep features refer to image feature combinations that can reflect the seed vitality state from different levels and dimensions. For example, the clarity of the seed endosperm texture, the visibility of the embryo structure, and the distribution characteristics of the seed internal density. In this scheme, the seed vitality recognition model is a pre-set calculation model, including calculation steps for obtaining each deep feature. By inputting a single seed optical image and a single seed X-ray image, the calculated results are output.

[0027] As a preferred scheme, the seed quality score is obtained according to the seed appearance feature parameters, internal structure parameters, and vitality index, which includes: A quality evaluation model is established, which obtains corresponding matching values according to the input seed appearance feature parameters, internal structure parameters, and vitality index, and calculates the quality grade according to the matching values.

[0028] This scheme establishes a quality evaluation model according to a set quality grade calculation method. Specifically, a multi-stage matching range and corresponding matching value are set for the input data, and the final quality grade is calculated according to the matching value, such as the weighted sum method, and the calculated data is converted into the corresponding quality grade. The input seed appearance feature parameters, internal structure parameters, and vitality index are input into the quality evaluation model and converted into corresponding matching values, and the corresponding quality grade is calculated and converted according to the matching values. The present application evaluates the seed quality from multiple dimensions such as seed external morphology and internal structure, as well as vitality index, which has more evaluation dimensions, more comprehensive evaluation angles, and more accurate reflection of seed quality.

[0029] A multi-dimensional seed quality evaluation device, comprising: An image acquisition module acquires seed optical images and X-ray images respectively; A processor module identifies and locates the seeds in the acquired images, divides to form single-seed optical images and X-ray images, calculates appearance feature parameters and internal structure parameters respectively according to the single-seed optical images and X-ray images, inputs the single-seed optical images and X-ray images into a vigor identification model to obtain a seed vigor index, inputs the seed appearance feature parameters, internal structure parameters and vigor index into a quality evaluation model, and outputs to obtain a seed quality grade.

[0030] The image acquisition module includes a camera and an X-ray machine. The camera is a traditional high-pixel camera, which is used to shoot the appearance features of the seeds and transmit them to the processor module in the form of optical images. The X-ray machine is used to shoot the internal endosperm and embryo of the seeds and transmit them to the processor module in the form of X-ray images. The processor module preferably adopts JestonNano microprocessors, which process the input seed optical images and X-ray images, including four parts.

[0031] The first part divides to obtain single-seed optical images and single-seed X-ray images. A semantic segmentation network is used to identify and locate several seeds in the acquired seed optical images, and the identified seeds are divided to generate several single-seed optical images. The semantic segmentation network adopts a pixel-level segmentation model based on a U-Net convolutional neural network, which extracts multi-scale image features through an encoder, combines a decoder for high-resolution recovery, and uses a Softmax layer to output the class label of each pixel, thereby realizing accurate identification and segmentation of the seed region. This network has good edge detection ability and spatial detail preservation ability, and can accurately extract the boundary contour of each seed in a complex background. According to the single-seed optical images and the seed X-ray images, the seed positioning information is transmitted through a correlation algorithm to generate several single-seed X-ray images. The correlation algorithm uses a combination of ORB feature extraction and matching and RANSAC affine transformation estimation to complete the spatial mapping and positioning projection between different modal images, thereby realizing high-precision cross-modal seed matching and sub-image cropping. This process not only improves the positioning accuracy, but also ensures the one-to-one correspondence between the single-grain X-ray image and the corresponding optical image target, ensuring the accuracy of subsequent parameter calculation and vigor evaluation.

[0032] The second part calculates key seed parameters, including seed appearance features and internal structure parameters. Seed appearance feature calculations primarily extract basic geometric and structural features from segmented single-seed optical images, calculate seed area and defect area in a single optical image, identify seed color, and use OpenCV's minimum bounding matrix and minimum bounding circle to calculate seed size and aspect ratio. Internal structure parameter calculations assess parameters such as the proportion of the embryo and endosperm in a single X-ray image, as well as seed transparency. A grayscale-based analysis method is used to separate and extract the embryo and endosperm regions, thereby calculating their area proportions. Simultaneously, the overall grayscale mean and distribution of the image are used to evaluate the seed's internal transparency characteristics.

[0033] The third part calculates seed vigor indices. As one implementation scheme, a seed vigor recognition model is constructed. Single-seed optical images and X-ray images are input into the model, which outputs a seed vigor index, which is a vigor score indicating vigor level. As another implementation scheme, a seed vigor recognition model is constructed. Single-seed optical images and X-ray images are input. The model extracts multi-level depth features characterizing seed vigor from both images, including endosperm texture clarity, germ structure visibility, and internal seed density distribution characteristics.

[0034] The fourth part calculates the seed quality grade; a quality assessment model is constructed, and the seed appearance characteristic parameters, seed internal structure parameters, and seed vigor index obtained from the previous three parts are input into the set quality assessment model to output the seed quality grade, so as to achieve a comprehensive evaluation of seed quality.

[0035] As a preferred option, it also includes: The encoder module collects the conveyor belt distance and sends it to the processor module to trigger the image acquisition module to work based on the conveyor belt distance. The display module acquires the calculation results from the processing module and displays the seed's appearance characteristics, internal structure parameters, vigor index, and quality grade.

[0036] The device includes a conveyor belt that transports seeds to a location below a camera and X-ray machine in the image acquisition chamber. It's important to note that the conveyor belt speed should not be too high, and the seeds should be spread out to minimize contact or overlap, which could affect subsequent seeding. The camera and X-ray machine capture optical and X-ray images of the seeds. An encoder module is installed on the conveyor belt. This module collects the conveyor distance and transmits the information to a processor module. The processor module then triggers the camera and X-ray machine to take images based on the conveyor distance. Simultaneously, the processor module outputs motor control signals to control the conveyor belt speed. Specifically, the processor module obtains encoder information or outputs signals to the motor via GPIO ports. A display module is connected to the processor module. Specifically, the display module outputs the calculation results to the display module via an HDMI interface for display, including seed appearance parameters, internal structural parameters, vigor indicators, and quality grades, providing a visual representation of multi-dimensional quality evaluation information for each seed.

[0037] Therefore, the advantages of this invention are: it calculates appearance characteristic parameters by photographing the seed's exterior with a camera, calculates internal structural parameters by photographing the seed's embryo and endosperm with an X-ray machine, and obtains seed vigor indicators through deep learning methods. Combining data from multiple dimensions, it outputs a seed quality evaluation, resulting in a more comprehensive and accurate assessment. This solves the technical problems of existing technologies, such as limited accuracy, low efficiency, and single-dimensionality in seed quality evaluation. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention.

[0039] Figure 2 This is a structural block diagram of the present invention.

[0040] Figure 3 This is a schematic diagram of the structure of the seed recognition model in this invention.

[0041] 1-Image acquisition module; 2-Processor module; 3-Encoder module; 4-Display module. Detailed Implementation

[0042] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0043] Example 1: This embodiment presents a multi-dimensional seed quality evaluation method, such as... Figure 1 As shown, it includes the following steps: S1. Acquire optical and X-ray images of the seed.

[0044] The seed optical image is collected by a traditional high-definition camera, and the image is a multi-seed optical image; the seed X-ray image is collected by an agricultural X-ray machine, and the image is a multi-seed X-ray image. The two types of images are respectively subjected to noise filtering and pretreatment to obtain standardized optical images and X-ray images as input sources for subsequent analysis.

[0045] S2. Segmentation of single-seed optical images and X-ray images.

[0046] S21. The seed optical image is identified and segmented by a semantic segmentation network to obtain a plurality of single-seed optical images.

[0047] The collected seed optical image is a multi-seed optical image, the plurality of seeds in the multi-seed optical image are identified and located by a semantic segmentation network, and the identified seeds are segmented to generate a plurality of single-seed optical images. The semantic segmentation network in this embodiment preferably adopts a pixel-level segmentation model based on a U-net convolutional neural network. The model extracts multi-scale image features through an encoder, combines a decoder for high-resolution recovery, and uses a Softmax layer to output the class label of each pixel, thereby achieving accurate identification and segmentation of the seed region. The model has good edge detection capability and spatial detail preservation capability, and can accurately extract the boundary profile of each seed in a complex background.

[0048] S22. A plurality of single-seed X-ray images are generated by an association algorithm from the seed X-ray image and the plurality of single-seed optical images.

[0049] The seed X-ray image collected by the X-ray machine is a multi-seed X-ray image, a plurality of seeds in the multi-seed X-ray image are identified and located, and the identified seeds are segmented to obtain a plurality of single-seed X-ray images. In this embodiment, the seed X-ray image and the plurality of single-seed optical images are combined, the seed positioning information is transmitted by an association algorithm, and a plurality of single-seed X-ray images are generated from the multi-seed X-ray image. In this embodiment, the association algorithm preferably adopts a method combining ORB feature extraction and matching with RANSAC affine transformation estimation to complete spatial mapping and positioning projection between different modal images, ensure one-to-one correspondence between the single X-ray image and the target optical image, and ensure the accuracy of subsequent parameter calculation and vitality index evaluation.

[0050] S3. Calculate the seed appearance feature parameters from the single-seed optical image and the seed internal structure parameters from the single-seed X-ray image.

[0051] S31. Extract the seed basic geometric and structural features from the single-seed optical image, calculate the area, defect area, size, aspect ratio of the seed in the single-seed optical image, and identify the seed color.

[0052] As a preferred scheme of the embodiment, according to the calculation of the appearance feature parameters of the single-seed optical image, a set appearance feature recognition network can be used for calculation, which can extract the basic geometric and structural features of the seed, calculate the area and defect area of the single seed, identify the color of the seed, and use the minimum circumscribed matrix and the minimum circumscribed circle of OpenCV to calculate the size and aspect ratio of the seed.

[0053] S32. The embryo and endosperm regions are separated and extracted from the single-seed X-ray image based on the image gray-scale analysis method, the area ratio of the embryo and endosperm is calculated, and the transparency feature inside the seed is obtained by the mean value and distribution method of the overall gray value of the image.

[0054] As a preferred scheme of the embodiment, according to the calculation of the internal structure parameters of the single-seed X-ray image, a set internal structure recognition network can be used for calculation, which calculates the area ratio of the embryo and endosperm by analyzing the embryo and endosperm regions in the single-seed X-ray image, and determines the median internal transparency by analyzing the gray value.

[0055] The order of the above-mentioned steps of calculating the appearance feature parameters of the seed and the steps of calculating the internal structure parameters of the seed can be exchanged, or they can be performed simultaneously, and finally the appearance feature parameters and the internal structure parameters are obtained and subsequent calculation is performed.

[0056] S4. The single-seed optical image and the X-ray image are input into the seed vigor recognition model, and various seed vigor indexes are output.

[0057] The seed vigor recognition model is constructed, as shown in Figure 3 The specific structure comprises two input layers, the two input layers are connected to the first branch and the second branch respectively, the first branch and the second branch are jointly connected to a fusion layer, the fusion layer is connected to an output layer through a full connection layer, and the first branch and the second branch each comprise three layers of convolution layers, a spatial attention mechanism, a reshaping layer, a 3D convolution layer and a time attention mechanism connected in sequence.

[0058] Input layer: Two input layers are used to input the single-seed optical image and the single-seed X-ray image respectively, and the input layer is input as a video frame sequence. It is assumed that each video protection T frame, and the size of each frame is HxWxC (heightxwidthxchannel number).

[0059] Convolution layer (LF Layer): including three layers of convolution layers connected in sequence, the convolution layer uses a 2D local feature extraction layer (Local Feature Layer) to extract local spatial features; the 2D local feature extraction layer is composed of Con2D+BatchNorm+ReLU activation function.

[0060] Spatial Attention: Add spatial attention mechanism after the convolutional layer to calculate the attention weight of each feature map. Specifically, a 1x1 convolutional layer is used to produce a spatial attention map, which is multiplied with the convolutional feature map to emphasize important regions.

[0061] Reshape Layer: Change the dimension of the input and adjust the shape of the input data to the specified shape.

[0062] Conv3D Layer: Extract features in the time dimension. Specifically, it directly processes video sequences to extract time features. As an alternative, a Bi-LSTM (Bidirectional Long Short-Term Memory) network can also be used to replace it, which models the sequence of features at each time step.

[0063] Temporal Attention: After time feature extraction, apply temporal attention mechanism to calculate the attention weight of each time step. Multiply the temporal attention map with the temporal feature to emphasize important time steps.

[0064] Fusion Layer: Fuse the features of the two branches.

[0065] Fully Connected Layer: Flatten the extracted features and pass them through a fully connected layer for classification or regression. Use Dropout layer to prevent overfitting.

[0066] Output Layer: Output the vitality index according to the task requirements. In this scheme, the vitality index is the vitality level or vitality score.

[0067] The seed vitality recognition model needs to be trained in advance, and the training process includes: 1. Data preparation, collect a large number of conveyor belt seed image set video data, and label each seed in the image. The labeled content is the seed quality level or quality index, which is used as the seed training label. The labeled images are divided into training set, validation set and test set according to the proportion. Each of the above data sets should contain the optical image and corresponding X-ray image of the same seed.

[0068] 2. Model training, input seed optical image and X-ray image samples into two branches for multi-modal learning during training. Joint optimization strategy is used to train two branches, so that the features extracted by the two branches can accurately predict the vitality index of the seed after fusion, thereby significantly improving the accuracy and robustness of the model in evaluating the vitality index of the seed.

[0069] 3. Loss function, choose Adam optimization algorithm to optimize network parameters.

[0070] 4. The training strategy can adopt batch training, set appropriate batch size and learning rate adjustment strategy, and use data augmentation techniques such as random cropping, rotation, flipping and scaling to increase data diversity and improve network generalization ability.

[0071] The single-seed optical image and the single-seed X-ray image are input into the trained vitality recognition model, and the vitality index of the seed is output.

[0072] S5. According to the seed appearance characteristic parameters, internal structure parameters and vitality index, a seed quality score is obtained.

[0073] As a preferred scheme of the embodiment, a quality evaluation model is established, which obtains a corresponding matching value according to the input seed appearance characteristic parameters, internal structure parameters and vitality index, and calculates a quality grade according to the matching value.

[0074] The quality evaluation model is established according to the set quality grade calculation method. Specifically, a multi-stage matching range and a corresponding matching value are set for the input data, and a final quality grade is calculated according to the matching value, such as a weighted sum, and the calculated data is converted into a corresponding quality grade. The input seed appearance characteristic parameters, internal structure parameters and vitality index are input into the quality evaluation model and converted into corresponding matching values, and the corresponding quality grade is calculated and converted according to the matching values. The present application evaluates the seed quality from multiple dimensions such as seed external morphology and internal structure, and vitality index, has more evaluation dimensions and more comprehensive evaluation angles, and more accurately reflects the quality of the seed.

[0075] The embodiment also includes a multi-dimensional seed quality evaluation device for implementing the above-mentioned multi-dimensional seed quality evaluation method, as shown in Figure 2 The device includes: An image acquisition module acquires seed optical images and X-ray images respectively; A processor module identifies and locates the seeds in the acquired images, divides them into single-seed optical images and X-ray images, calculates the appearance characteristic parameters and internal structure parameters according to the single-seed optical images and X-ray images respectively, inputs the single-seed optical images and X-ray images into the vitality recognition model to obtain the seed vitality index, and inputs the seed appearance characteristic parameters, internal structure parameters and vitality index into the quality evaluation model to output the seed quality grade.

[0076] An encoder module acquires the conveying distance of the conveying belt and sends it to the processor module to trigger the image acquisition module to work according to the conveying distance of the conveying belt; A display module obtains the calculation results of the processing module and displays the seed appearance characteristic parameters, internal structure parameters, vitality index and quality grade.

[0077] Specifically, the image acquisition module includes a camera and an X-ray machine. The camera is a traditional high-pixel camera, which is used to shoot the appearance features of the seeds and transmit the optical images to the processor module. The X-ray machine is used to shoot the internal embryo and endosperm of the seeds and transmit the X-ray images to the processor module. The processor module, preferably JestonNano microprocessor, processes the input seed optical images and X-ray images, which includes four parts.

[0078] The first part is to obtain single seed optical images and single seed X-ray images. The semantic segmentation network is used to identify and locate the seeds in the collected seed optical images, and the identified seeds are segmented to generate several single seed optical images. The semantic segmentation network uses a pixel-level segmentation model based on a U-Net convolutional neural network. The encoder extracts multi-scale image features, the decoder performs high-resolution recovery, and the Softmax layer outputs the class label of each pixel, thereby realizing accurate identification and segmentation of the seed region. This network has good edge detection ability and spatial detail preservation ability, and can accurately extract the boundary profile of each seed in a complex background. According to the single seed optical image and the seed X-ray image, the seed positioning information is transmitted through the correlation algorithm to generate several single seed X-ray images. The correlation algorithm uses a combination of ORB feature extraction and matching and RANSAC affine transformation estimation to complete the spatial mapping and positioning projection between different modal images, thereby realizing high-precision cross-modal seed matching and sub-image cropping. This process not only improves the positioning accuracy, but also ensures the one-to-one correspondence between the single X-ray image and the corresponding optical image target, ensuring the accuracy of subsequent parameter calculation and vitality evaluation.

[0079] The second part calculates the key parameters of the seeds, including the appearance feature parameters and the internal structure parameters. The calculation of the seed appearance feature parameters is mainly used to extract the basic geometric and structural features from the single seed optical image obtained by segmentation, calculate the area and defect area of the seed in the single optical image, identify the color of the seed, and use the OpenCV minimum circumscribed matrix and minimum circumscribed circle to calculate the size and aspect ratio of the seed. The calculation of the internal structure parameters of the seed includes the proportion of embryo and endosperm in the single X-ray image and the transparency of the seed. The image gray-based analysis method is used to separate and extract the embryo and endosperm regions, and then calculate the area proportion of the embryo and endosperm. At the same time, the overall gray value mean and distribution of the image are used to evaluate the transparency characteristics of the seed.

[0080] The third part calculates the seed vigor index; as an embodiment, a seed vigor recognition model is constructed, the single-seed optical image and the X-ray image are input into the vigor recognition model, and the seed vigor index is output, the vigor index is a vigor score and a vigor grade. As another embodiment, a seed vigor recognition model is constructed, the single-seed optical image and the X-ray image are input into the vigor recognition model, and the seed vigor recognition model extracts multi-level deep features representing the seed vigor from the single-seed optical image and the X-ray image, including seed endosperm texture clarity, embryo structure visibility, and seed internal density distribution characteristics.

[0081] The fourth part calculates the seed quality grade; a quality evaluation model is constructed, the seed appearance feature parameters, the seed internal structure parameters, and the seed vigor index calculated in the previous three parts are input into the quality evaluation model, and the seed quality grade is output, so as to realize comprehensive evaluation of the seed quality.

[0082] The device includes a conveying belt for conveying seeds, the conveying belt conveys the seeds to below the camera and the X-ray machine in the image acquisition chamber; it should be noted that the conveying belt speed should not be too fast, and the seeds should be dispersed as much as possible to reduce the situation of being close to or overlapping each other, so as to avoid affecting the subsequent seeds. The camera and the X-ray machine are used to capture and collect seed optical images and X-ray images. An encoder module is installed on the conveying belt, the encoder module collects the conveying distance of the conveying belt, and the collected information is input into the processor module, and the processor module triggers the camera and the X-ray machine to take pictures according to the conveying distance of the conveying belt; at the same time, the processor module outputs motor control signals for controlling the conveying belt speed; specifically, the processor module obtains the encoder information or outputs signals to the motor through the GPIO port. The display module is connected with the processor module; specifically, the display module outputs the calculation results to the display module for display through the HDMI interface, including displaying the seed appearance feature parameters, the internal structure parameters, the vigor index, and the quality grade, and directly displaying the multi-dimensional quality evaluation information of each seed.

[0083] Embodiment 2 This embodiment gives another embodiment of a multi-dimensional seed quality evaluation method, which specifically includes the following steps: S1. Collecting seed optical images and X-ray images.

[0084] The seed optical images are collected by a traditional high-definition camera, and the images are multi-seed optical images; the seed X-ray images are collected by an agricultural X-ray machine, and the images are multi-seed X-ray images. The two types of images are subjected to noise filtering and pretreatment respectively, and standardized optical images and X-ray images are obtained as input sources for subsequent analysis.

[0085] S2. Segmentation to obtain single-seed optical images and X-ray images.

[0086] S21. The seed optical image is identified and segmented by a semantic segmentation network to obtain a plurality of single-seed optical images.

[0087] The collected seed optical image is a multi-seed optical image. The semantic segmentation network is used to identify and locate the seeds in the multi-seed optical image, and the identified seeds are segmented to generate a plurality of single-seed optical images. In this embodiment, the semantic segmentation network preferably uses a pixel-level segmentation model based on a U-net convolutional neural network. The model extracts multi-scale image features through an encoder, combines a decoder for high-resolution recovery, and uses a Softmax layer to output the class label of each pixel, thereby achieving accurate identification and segmentation of the seed region. This model has good edge detection capability and spatial detail preservation capability, and can accurately extract the boundary profile of each seed in a complex background.

[0088] S22. A plurality of single-seed X-ray images are generated by associating the seed X-ray image with the plurality of single-seed optical images.

[0089] The seed X-ray image collected by the X-ray machine is a multi-seed X-ray image. The seeds in the multi-seed X-ray image are identified and located, and the identified seeds are segmented to obtain a plurality of single-seed X-ray images. In this embodiment, the seed X-ray image is combined with a plurality of single-seed optical images, and the seed positioning information is transmitted by an association algorithm to generate a plurality of single-seed X-ray images from the multi-seed X-ray image. In this embodiment, the association algorithm preferably uses a method combining ORB feature extraction and matching with RANSAC affine transformation estimation to complete the spatial mapping and positioning projection between different modal images, ensuring a one-to-one correspondence between the single X-ray image and the target optical image, and ensuring the accuracy of subsequent parameter calculation and vitality index evaluation.

[0090] S3. Calculate the seed appearance feature parameters from the single-seed optical image and the seed internal structure parameters from the single-seed X-ray image.

[0091] S31. Extract the seed basic geometric and structural features from the single-seed optical image, calculate the area, defect area, size, aspect ratio of the seed in the single-seed optical image, and identify the seed color.

[0092] As a preferred scheme of this embodiment, the calculation of the seed appearance feature parameters from the single-seed optical image can be performed using a set appearance feature recognition network. This appearance feature recognition network can extract the seed basic geometric and structural features, calculate the area and defect area of a single seed, and identify the seed color. The OpenCV minimum bounding rectangle and minimum bounding circle are used to calculate the size and aspect ratio of the seed.

[0093] S32. An image grayscale-based analysis method is used to separate and extract the embryo and endosperm regions from a single seed X-ray image, calculate the area ratio of the embryo and endosperm, and obtain the transparency characteristics inside the seed by the mean and distribution method of the overall image grayscale value.

[0094] As a preferred embodiment, the internal structural parameters of the seed can be calculated based on the X-ray image of a single seed. This calculation can be performed using a pre-defined internal structure recognition network. This internal structure recognition network calculates the area ratio of the germ and endosperm by analyzing the germ and endosperm regions in the X-ray image of a single seed, and determines the median internal transparency by analyzing the grayscale values.

[0095] The order of the steps for calculating seed appearance parameters and seed internal structure parameters can be reversed or performed simultaneously. The ultimate goal is to obtain the appearance parameters and internal structure parameters and to perform subsequent calculations.

[0096] S4. Input single-seed optical images and X-ray images into the seed vigor recognition model, and output various seed vigor indices.

[0097] Construct a seed viability identification model, such as Figure 3 As shown, its specific structure includes two input layers, which are connected to the first branch and the second branch respectively. The first and second branches are connected to the fusion layer. The fusion layer is connected to the output layer through a fully connected layer. The first and second branches each include three convolutional layers, a spatial attention mechanism, a reshaping layer, a 3D convolutional layer, and a temporal attention mechanism connected in sequence.

[0098] Single-seed optical and X-ray images are input into the seed vigor recognition model, and the extracted vigor indicators include multi-level deep features used to characterize seed vigor.

[0099] This scheme extracts multi-level depth features characterizing seed vigor from single-seed optical and X-ray images using a seed vigor recognition model. These multi-level depth features are combinations of image features that reflect seed vigor status from different levels and dimensions. Examples include the clarity of seed endosperm texture, the visibility of the germ structure, and the distribution characteristics of seed internal density. In this scheme, the seed vigor recognition model is a pre-defined computational model that includes calculation steps for acquiring each depth feature. It takes a single-seed optical and X-ray image as input and outputs the calculated results.

[0100] S5. Obtain a seed quality score based on seed appearance characteristics, internal structure parameters, and vigor indicators.

[0101] As a preferred embodiment, a quality evaluation model is established. The quality evaluation model obtains corresponding matching values ​​based on the input seed appearance feature parameters, internal structure parameters, and vitality index, and calculates the quality level based on the matching values.

[0102] According to the quality grade calculation mode, the quality evaluation model is established, and the matching range and the corresponding matching value of the input data are set in multiple stages, and the final quality grade is calculated according to the matching value, such as the weighted sum mode, and the corresponding quality grade is converted according to the calculated data. The seed appearance feature parameters, internal structure parameters and vitality indexes are input into the quality evaluation model, and the corresponding matching values are converted, and the corresponding quality grade is calculated and converted. The present application evaluates the seed quality from multiple dimensions such as seed external morphology and internal structure, and vitality index, and the evaluation dimension is more, the evaluation angle is more comprehensive, and the seed quality is more accurately reflected.

[0103] The specific embodiments described herein merely exemplify the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

[0104] Although the terms such as image acquisition module, processor module, encoder module and display module are used more frequently herein, the possibility of using other terms is not excluded. The use of these terms is only to facilitate the description and explanation of the essence of the present application; any additional limitation is contrary to the spirit of the present application.

Claims

1. A multi-dimensional seed quality evaluation method, characterized in that, Includes the following steps: Acquire optical and X-ray images of the seeds; Segmentation to obtain single-seed optical images and X-ray images; Calculate seed appearance feature parameters based on single seed optical images, and calculate seed internal structure parameters based on single seed X-ray images; Single-seed optical and X-ray images are input into the seed viability recognition model, which outputs various seed viability indices. Seed quality scores are obtained based on seed appearance characteristics, internal structure parameters, and vigor indicators.

2. The multi-dimensional seed quality evaluation method according to claim 1, characterized in that: Seed optical images are identified and segmented using a semantic segmentation network to obtain several single seed optical images.

3. The multi-dimensional seed quality evaluation method according to claim 2, characterized in that: A seed X-ray image is generated from a seed optical image and several seed X-ray images using an association algorithm.

4. A multi-dimensional seed quality evaluation method according to claim 1, 2, or 3, characterized in that, The calculation of seed appearance feature parameters based on a single seed optical image includes: Extract the basic geometric and structural features of seeds from single-seed optical images, calculate the area, defect area, size, aspect ratio of seeds in single-seed optical images, and identify seed color.

5. The multi-dimensional seed quality evaluation method according to claim 4, characterized in that: The calculation of seed internal structure parameters based on a single seed X-ray image includes: An image grayscale-based analysis method was used to separate and extract the embryo and endosperm regions from a single seed X-ray image, calculate the area ratio of the embryo and endosperm, and obtain the transparency characteristics inside the seed by the mean and distribution method of the overall image grayscale value.

6. A multi-dimensional seed quality evaluation method according to claim 1, 2, or 3, characterized in that: The seed viability recognition model includes two input layers, which are connected to the first branch and the second branch respectively. The first and second branches are connected to the fusion layer. The fusion layer is connected to the output layer through a fully connected layer. Both the first and second branches include three convolutional layers, a spatial attention mechanism, a reshaping layer, a 3D convolutional layer, and a temporal attention mechanism connected in sequence.

7. A multi-dimensional seed quality evaluation method according to claim 1, 2, or 3, characterized in that: Single-seed optical and X-ray images are input into the seed vigor recognition model, and the extracted vigor indicators include multi-level deep features used to characterize seed vigor.

8. The multi-dimensional seed quality evaluation method according to claim 1, characterized in that, The seed quality score is obtained based on seed appearance characteristics, internal structure parameters, and vigor indicators, including: A quality evaluation model is established. The model obtains corresponding matching values ​​based on the input seed appearance feature parameters, internal structure parameters, and vitality index, and calculates the quality level based on the matching values.

9. A multi-dimensional seed quality evaluation device, implementing the method described in any one of claims 1-8, characterized in that, include: The image acquisition module acquires optical images and X-ray images of the seed, respectively; The processor module identifies and locates seeds in the acquired images, segments them into single-seed optical images and X-ray images, calculates appearance feature parameters and internal structure parameters based on the single-seed optical images and X-ray images respectively, inputs the single-seed optical images and X-ray images into the vigor recognition model to obtain seed vigor index, inputs the seed appearance feature parameters, internal structure parameters and vigor index into the quality assessment model, and outputs the seed quality grade.

10. A multi-dimensional seed quality evaluation device according to claim 9, characterized in that, Also includes: The encoder module collects the conveyor belt distance and sends it to the processor module to trigger the image acquisition module to work based on the conveyor belt distance. The display module acquires the calculation results from the processing module and displays the seed's appearance characteristics, internal structure parameters, vigor index, and quality grade.

Images