Deep learning-based intelligent rice seed screening technical method and control system

Abstract

The invention discloses an intelligent rice seed screening technical method based on deep learning and a control system, and the method comprises the steps: collecting continuous frame RGB images, rolling on an inclined collection plate, of each batch of rice seeds in a preset time period through a high-speed camera, and analyzing the continuous frame RGB image information of all batches of rice seeds; extracting image contour information and an image pixel ratio of each batch of rice seeds, judging the batch with the optimal representation of the rice seeds, and recording the batch as a target sample; the method comprises the following steps: acquiring spectral data of a target sample through a hyperspectral imaging system, preprocessing the spectral data and extracting a spectral marking signal; constructing a convolutional neural network model, inputting the spectral marking signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and outputting an adjustment result; according to the method, the characterization information of the rice seeds is accurately obtained, and batch intelligent screening of the rice seeds is carried out, so that the screening efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent seed screening methods, and in particular to a deep learning-based intelligent rice seed screening technical method and a control system. Background Art

[0002] Intelligent screening technology has a wide range of applications in rice breeding, molecular research and other fields. For example, it can be used for rapid identification of rice varieties, disease resistance assessment, and quantitative analysis of seed quality. The application of these technologies helps to accelerate the breeding process, improve crop yield and quality, and is of great significance to food safety and sustainable agricultural development.

[0003] Deep learning-based intelligent rice seed screening technology usually involves the use of hyperspectral imaging, image processing, and advanced artificial intelligence technologies such as convolutional neural networks (CNNs). These technologies can extract key spectral and morphological features from rice seeds for rapid and accurate classification and screening of seeds of different varieties, and provide a powerful tool for breeding by constructing a large-scale dataset of Chinese cultivated rice populations and using deep learning models to predict phenotypic values ​​at different planting locations.

[0004] In existing research, in order to more clearly study the external structural integrity and internal characteristics of rice seeds, technicians have conducted in-depth research and analysis on rice seeds using image acquisition related equipment and combined with spectral technology; especially in terms of seed growth metabolism; the better vitality of rice seeds provides better genetic resources and nutrients required for growth; however, in the process of detecting the external structure of rice seeds, it is difficult to conduct a more accurate study on the characteristics of batch rice seeds. Only the current rice seeds can be sorted and analyzed separately, and a small amount of samples is not accurate enough for planting with high demand, and more samples are needed for statistics. The actual operation is also time-consuming and labor-intensive. Summary of the invention

[0005] The purpose of the present invention is to provide a rice seed intelligent screening technology method and control system based on deep learning to solve the following technical problems: How to achieve accurate acquisition of rice seed characterization information and conduct batch intelligent screening of rice seeds to improve screening efficiency.

[0006] The purpose of the present invention can be achieved through the following technical solutions: A rice seed intelligent screening technology method based on deep learning, the method includes: S1. Obtain several batches of rice seeds, use a high-speed camera to collect continuous frame RGB images of rice seeds of each batch rolling on an inclined collection plate within a preset period of time, and analyze the continuous frame RGB image information of all batches of rice seeds; S2, extracting image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and recording the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; S3, collecting spectral data of the target sample through a hyperspectral imaging system, preprocessing the spectral data and extracting spectral marker signals; S4. Construct a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

[0007] Preferably, in step S1: S11, obtaining a set of continuous frame RGB image pixels of rice seeds from different batches within the same time period; S12, extracting multiple continuous frame RGB image pixel sets of the same frame rate from each batch and arranging them in time sequence; S13. Analyze the rolling characteristics of all batches of rice seeds based on the time-series arranged RGB image pixel sets.

[0008] Preferably, in step S2: S21, obtaining the contour sizes of abnormal regions of multiple batches of images and pixel ratio parameter values ​​of multiple batches of seed images corresponding to the rolling feature analysis of each batch of rice seeds; S22, analyzing the contour sizes of abnormal regions of multiple batches of images and the acquisition characterization coefficients of pixel ratio parameter values ​​of multiple batches of seed images; S23, comparing the characterization coefficient with a preset standard threshold, analyzing and obtaining a batch characterization coefficient data set that meets the requirements, and recording the batch quantity; S24. Record the data set of batch characterization coefficients that meet the requirements as training samples and input them into the neural network model for training. Take the batch of characterization coefficients corresponding to the obtained optimal simulation parameters as output and record them as target samples.

[0009] Preferably, the characterization coefficient is obtained in step S22 as follows: By formula Calculate the first Characterization coefficient of rice seeds in batches ; in, The total number of continuous frame images of rice seeds rolling on the inclined acquisition plate within a preset period of time for the same batch of detection; ∈ ; is the preset function; For the Batch No. Abnormal area structure value of consecutive frame images; For the Batch No. The seed image pixel ratio parameter of consecutive frame images; For the The standard value of abnormal area structure of batches of continuous frame images; For the The standard value of the seed image pixel ratio parameter of batches of continuous frame images; For the Batch No. The standard deviation of the seed image pixel ratio parameter of consecutive frame images; is a first weight coefficient preset according to the influence of the contour of the abnormal area of ​​the current image, A second weight coefficient is preset according to the influence of the seed image pixel ratio parameter value of the current image; and , Both are greater than 0.

[0010] Preferably, in step S23, the batch characterization coefficients meeting the requirements are obtained and the batch quantity is recorded in the following manner: The characterization coefficient of the current batch of rice seeds The standard characterization coefficient threshold interval obtained by machine simulation calculation in advance For comparison: like ∈ , it is judged that the characterization of the current batch of rice seeds is normal, and the characterization coefficient of the current rice seeds and the number of batches are recorded; like < , it is judged that the current batch of rice seeds is abnormally small; like > , it is judged that the characteristics of the current batch of rice seeds are abnormally large.

[0011] Preferably, in step S3: The radiation information of the target sample under the preset light wave is obtained by continuous spectral scanning and recorded as the original spectral data; Perform spectral wavelength correction and noise reduction pre-processing on the original spectral data; Obtain the preprocessed spectral data to extract characteristic wavelengths and calculate using the random Senli model: By formula Calculate the features At the node Importance ;in, For each feature Gini index; , They are random forest nodes Two new nodes after branching , of index; And judge: If the feature In the decision tree Nodes that appear in the preset node set Among them, then: ; For the current characteristic wavelength In random forests, it is assumed that Based on the situation of the decision tree, the importance of the current characteristic wavelength is evaluated: ; The importance scores of all characteristic wavelengths calculated are normalized, sorted, and the order is output as a spectral marker signal.

[0012] Preferably, the process of identifying the contour of the abnormal area of ​​the image is: The continuous frame RGB rice seed images rolling on the tilted acquisition plate are grayed out; The Canny algorithm is used to extract the contour of the abnormal area of ​​the image within the range of rice seeds in the current image; Calculate the number of contours of abnormal regions of the image to form a first set A; The ratio of the integrated area of ​​the contours of the abnormal regions of the images of rice accumulated at the same time point in different batches to the number of contours of the corresponding abnormal regions of the images is obtained as an element to form a second set B; Compare the elements x in the second set B with the standard set D one by one: like , then it is determined that the rice seeds of the element x corresponding to the abnormal area contour of the image at the corresponding time point of the current batch have normal quality; , it is determined that there is a problem with the quality of rice seeds in the contour of the abnormal area of ​​the image corresponding to the element x at the corresponding time point of the current batch.

[0013] Preferably, the method further comprises adjusting the inclined collecting plate with different inclination angles, and analyzing the rolling speed of the rice seeds on the inclined collecting plate to determine the best rolling state of the seeds.

[0014] A rice seed intelligent screening technology control system based on deep learning, the system includes: An information collection module is used to obtain several batches of rice seeds, collect continuous frame RGB images of rice seeds rolling on the inclined collection plate within a preset period of time for each batch using a high-speed camera, and analyze the continuous frame RGB image information of all batches of rice seeds; A sample selection module is used to extract image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and record the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; An information extraction module is used to collect spectral data of a target sample through a hyperspectral imaging system, preprocess the spectral data and extract spectral marker signals; The model building module is used to build a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

[0015] Beneficial effects of the present invention: (1) The present invention uses a high-speed camera to capture the continuous frame RGB image acquisition process of rolling rice seeds on a tilted acquisition plate, thereby ensuring sufficient and batched acquisition of images of seeds at different angles and states; then extracts the image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed characterization, and records the current best batch of rice seeds as the target sample, and the image contour information includes the contour of the abnormal image area; selects the batch with the best characterization as the target sample, which is convenient for subsequent spectral analysis.

[0016] (2) The present invention calculates the characterization coefficient to analyze the distribution characteristics of the contour size of the abnormal area of ​​the image and the pixel ratio parameter value of the seed image of multiple batches of rice seeds, and then determines whether the characterization of the rice seeds is normal; the potential regularities or anomalies in the data set can be identified through statistics and calculations; by comparing the characterization coefficient with a preset standard threshold, the batch characterization coefficient data set that meets the requirements is analyzed and the batch number is recorded; the characterization coefficient obtained by analysis is compared with the preset standard threshold to evaluate which batches of seed images meet the predetermined quality standards; batches that are consistent and meet the requirements can be screened out, and the number of these batches is recorded for further analysis.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 This is a step diagram of the rice seed intelligent screening technology method based on deep learning of the present invention; Figure 2 This is a step diagram of the rice seed image information collection method in step S1 of the present invention; Figure 3 A schematic diagram of a method for obtaining a target sample of rice seeds in step S2 of the present invention; Figure 4 This is a module diagram of the control system of the rice seed intelligent screening technology based on deep learning of the present invention; Figure 5 It is a schematic diagram of the arrangement of the rice seed image information acquisition equipment of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] In existing research, in order to more clearly study the external structural integrity and internal characteristics of rice seeds, technicians have conducted in-depth research and analysis on rice seeds using image acquisition related equipment and combined with spectral technology; especially in terms of seed growth metabolism; the better vitality of rice seeds provides better genetic resources and nutrients required for growth; however, in the process of detecting the external structure of rice seeds, it is difficult to conduct a more accurate study on the characteristics of batch rice seeds. Only the current rice seeds can be sorted and analyzed separately, and a small amount of samples is not accurate enough for planting with high demand, and more samples are needed for statistics. The actual operation is also time-consuming and labor-intensive.

[0022] To solve the above technical problems, please refer to Figure 1 and Figure 5 As shown, the present invention is a rice seed intelligent screening technology method based on deep learning, and the method includes: S1. Obtain several batches of rice seeds, use a high-speed camera to collect continuous frame RGB images of rice seeds of each batch rolling on an inclined collection plate within a preset period of time, and analyze the continuous frame RGB image information of all batches of rice seeds; S2, extracting image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and recording the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; S3, collecting spectral data of the target sample through a hyperspectral imaging system, preprocessing the spectral data and extracting spectral marker signals; S4. Construct a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

[0023] In the above technical scheme, rice seeds are generally oval or oblate, with a smooth surface and luster; and the external structure of the seeds includes husks and rice grains; the rice seeds scattered under the initial driving force will roll on a smooth plane with a suitable inclined angle, and the fuller the seeds are, the faster they roll on the corresponding inclined plane; in order to realize batch collection of rice seed surface information, especially the surface information of rice seeds at different angles can be fully obtained in a dynamic rolling state, compared with the acquisition of static and one-sided surface information, obtaining different dynamic surface information is conducive to comprehensively obtaining the state of seed surface characteristics, and then judging the surface shape quality of the seeds.

[0024] Based on the above considerations, the design plans to collect rice seed image information of different frames through high-speed cameras to complete the different dynamic surface information of rice seeds. For details, please refer to Figure 5 As shown, one of the embodiments is to obtain the characterization information by means of continuous frame RGB images rolling on an inclined acquisition plate; the second embodiment also considers obtaining by means of two acquisition plates at different angles, wherein the first acquisition plate is fixed on a plane carrier by an adjusting rod and is inclined relative to the plane carrier, and the second acquisition plate is arranged vertically relative to the carrier plane, and one end is connected to the downwardly inclined end of the first acquisition plate; the rice seeds are collected by rolling the image downward through the first acquisition plate, and different frames of vertically falling rice seeds are collected by another high-speed camera with the vertical second acquisition plate as the background as an auxiliary reference, and the analysis method is the same as that of the first embodiment.

[0025] In order to realize an embodiment of the above-mentioned deep learning-based intelligent screening technology method for rice seeds, the first step is: first obtain several batches of rice seeds, use a high-speed camera to collect continuous frame RGB images of rice seeds rolling on an inclined acquisition board within a preset period of time for each batch, and analyze the continuous frame RGB image information of all batches of rice seeds. By using a high-speed camera to capture the continuous frame RGB images of rice seeds rolling on an inclined acquisition board, it is ensured that images of seeds at different angles and states are fully and batched; by image information analysis, it is ensured that the collected continuous frame images are processed, and the appearance characteristics of the seeds, such as color, shape and surface texture, are analyzed to identify possible abnormalities or defects.

[0026] For details, please refer to Figure 2 As shown, in step S1: S11, obtaining a set of continuous frame RGB image pixels of rice seeds from different batches within the same time period; S12, extracting multiple continuous frame RGB image pixel sets of the same frame rate from each batch and arranging them in time sequence; S13. Analyze the rolling characteristics of all batches of rice seeds based on the time-series arranged RGB image pixel sets.

[0027] In the above technical scheme, since it is necessary to analyze different batches of rice seeds in the same period during the selection and breeding process of rice seeds, and the adaptability of rice seeds in different regions to the local climate environment is also different, it is necessary to select and analyze the rice seed characterization information in combination with the local climate. First, by batching and analyzing the rice seeds, a continuous frame RGB image pixel set in the same time period is obtained; wherein the number of frames here can be pre-designed according to the high-speed camera, and the designed number of frames is adaptively set according to the rolling speed and rolling state of the rice seeds in the reference sample on the collection board at a set tilt angle; then, extract multiple continuous frame RGB image pixel sets of the same frame rate of each batch and arrange them in time sequence; finally, analyze the rolling characteristics of all batches of rice seeds according to the RGB image pixel sets arranged in time sequence, because the rolling characteristics of the current rice seeds are relatively simple, according to the image arrangement of different time sequences, the motion information of the rice seeds in the continuous frame time can be obtained by using the optical flow method or the inter-frame difference method, and the rolling motion state and rolling motion frequency of the current rice seeds at the set frame rate are analyzed to obtain the rolling characteristics.

[0028] As an embodiment of the present invention, the second is: then extracting the image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed characterization, and recording the current best batch of rice seeds as the target sample, the image contour information includes the contour of the abnormal image area; obtaining the continuous frame RGB image information of all batches of rice seeds in step S1; wherein the continuous frame RGB image information refers to a series of images arranged in time sequence, and each frame of the image contains information of three color channels of red (R), green (G) and blue (B); these images are usually used in video sequences, in which each frame is a continuous copy of the previous frame and may contain differences due to object movement, camera movement or environmental changes; and extracting the image contour information and the brightness values ​​of the three color channels of red, green and blue to analyze the image pixel ratio to determine the characterization information of all batches of rice seeds, and analyzing and recording the batch of rice seeds with the best characterization information as the target sample, the image contour information includes the contour of the abnormal image area; through image processing technology, extracting the contour information of the rice seeds, including the contour of the abnormal area, which helps to identify irregular shapes or damages of seeds. The pixel ratio of the seed image is compared through the pixel ratio analysis process to determine the fullness and uniformity of the seeds; finally, the best batch is determined, and the batch with the best characterization is selected as the target sample based on contour information and pixel ratio analysis to facilitate subsequent spectral analysis.

[0029] As an embodiment of the present invention, based on the above-mentioned image processing technology, the contour information of rice seeds is identified and a method for identifying the contour of abnormal regions is performed. The specific process of the contour identification of abnormal regions of images set in this embodiment is as follows: The continuous frame RGB rice seed images rolling on the tilted acquisition plate are grayed out; The Canny algorithm is used to extract the contour of the abnormal area of ​​the image within the range of rice seeds in the current image; Calculate the number of contours of abnormal regions of the image to form a first set A; The ratio of the integrated area of ​​the contours of the abnormal regions of the images of rice accumulated at the same time point in different batches to the number of contours of the corresponding abnormal regions of the images is obtained as an element to form a second set B; Compare the elements x in the second set B with the standard set D one by one: like , then it is determined that the rice seeds of the element x corresponding to the abnormal area contour of the image at the corresponding time point of the current batch have normal quality; , it is determined that there is a problem with the quality of rice seeds in the contour of the abnormal area of ​​the image corresponding to the element x at the corresponding time point of the current batch.

[0030] In the above technical scheme, for the characterization information of rice seeds, the seed area of ​​the surface seeds during the rolling process and the size of the identified abnormal seed area can reflect the current seed quality of rice seeds, specifically by converting the collected RGB color image into a grayscale image to reduce the complexity of image processing while retaining the main feature information of the seeds; then the Canny edge detection algorithm is used to identify the edges of the seeds in the image; and based on the Canny edge detection algorithm, the edges of possible damage, lesions or other abnormal features on the rice seeds can be highlighted; the Canny edge detection algorithm can effectively detect the edges in the image while suppressing noise, and then calculate the contour of the abnormal area to count each image The number of abnormal regions in the image is calculated, and the number of abnormal region contours in all images is collected to form a set A for subsequent analysis; and further, for each batch of rice seeds, the comprehensive area of ​​the abnormal regions in all images is calculated; the comprehensive area is compared with the number of corresponding abnormal region contours to obtain a ratio, which reflects the density of abnormal features in a certain area, and the ratios of all batches are collected to form a set B for evaluating the quality of different batches of rice seeds; finally, the set B is analyzed, and a standard set is predefined first, which includes the standard ratio range corresponding to normal and abnormal rice seed quality; each element x in the set B is compared with the standard value in the standard set D to evaluate the seed quality. The above systematically evaluates the quality status of rice seeds in different batches and at different time points, providing a scientific basis for seed screening and quality control.

[0031] For details, please refer to Figure 3 As shown, in step S2, the method for obtaining the batch of rice seeds with the best characterization information and obtaining the target sample is as follows: S21, obtaining the contour sizes of abnormal regions of multiple batches of images and pixel ratio parameter values ​​of multiple batches of seed images corresponding to the rolling feature analysis of each batch of rice seeds; S22, analyzing the contour sizes of abnormal regions of multiple batches of images and the acquisition characterization coefficients of pixel ratio parameter values ​​of multiple batches of seed images; S23, comparing the characterization coefficient with a preset standard threshold, analyzing and obtaining a batch characterization coefficient data set that meets the requirements, and recording the batch quantity; S24. Record the data set of batch characterization coefficients that meet the requirements as training samples and input them into the neural network model for training. Take the batch of characterization coefficients corresponding to the obtained optimal simulation parameters as output and record them as target samples.

[0032] In the above technical scheme, in this embodiment, the rolling feature analysis of each batch of rice seeds is first obtained, and image processing technology, such as image segmentation, feature extraction, etc., is generally applied to obtain the contour size of the abnormal area and the pixel ratio parameter value of the seed image in the corresponding output batches of images, and these parameter values ​​are the basic data for subsequent analysis; the process of analyzing the rolling features obtained in step S13 first combines and arranges the images that meet the preset rolling features in the acquired time series images, usually first arranging each detection window according to the order of continuous time, and then normalizing each image, calculating its gradient to obtain the histogram and specifying the weighted projection, which are all preset, and then performing contrast normalization on the units in each overlapping key block, and merging the histogram vectors in all the key blocks to obtain the overall feature vector, which are all processes completed according to the existing image segmentation and feature extraction technology, and will not be described in detail here.

[0033] As an embodiment of the present invention, characterization coefficients are obtained by analyzing the contour sizes of abnormal areas of multiple batches of images and the pixel ratio parameter values ​​of multiple batches of seed images. The characterization coefficients refer to statistics used to measure the characteristics of a data set, which help evaluate the consistency and variability of a data set; and the characterization coefficients are used to analyze the distribution characteristics of the contour sizes of abnormal areas of images of multiple batches of rice seeds and the pixel ratio parameter values ​​of seed images, thereby determining whether the characterization of rice seeds is normal; potential regularities or anomalies in the data set can be identified through statistics and calculations.

[0034] Specifically, the characterization coefficient is obtained in step S22 as follows: By formula Calculate the first Characterization coefficient of rice seeds in batches ; in, The total number of continuous frame images of rice seeds rolling on the inclined acquisition plate within a preset period of time for the same batch of detection; ∈ ; is the preset function; For the Batch No. Abnormal area structure value of consecutive frame images; For the Batch No. The seed image pixel ratio parameter of consecutive frame images; For the The standard value of abnormal area structure of batches of continuous frame images; For the The standard value of the seed image pixel ratio parameter of batches of continuous frame images; For the Batch No. The standard deviation of the seed image pixel ratio parameter of consecutive frame images; is a first weight coefficient preset according to the influence of the contour of the abnormal area of ​​the current image, A second weight coefficient is preset according to the influence of the seed image pixel ratio parameter value of the current image; and , Both are greater than 0.

[0035] In the above technical solution, the rice seed characterization coefficient is obtained by calculating the formula ; And according to the distribution of the contour size of the abnormal area of ​​the image and the pixel ratio parameter value of the seed image, the positive and negative effects on the seed characterization are analyzed. It is necessary to make specific judgments based on the seed resources of the preferred rice seeds and / or the different needs of sowing. Generally speaking, ensuring the integrity of the seed characterization structure and the relative uniformity of the seed surface color distribution can better ensure that the nutrition locked in the rice grains is not lost due to the influence of the external environment, affecting seed storage; of course, for sowing, the darker the color distribution is, the thicker the seed surface is, which is not conducive to the early germination of the seeds. Therefore, it is necessary to set a seed characterization information that is relatively in line with the specific requirements according to specific needs.

[0036] In this embodiment, the abnormal region structure values ​​of different continuous frame images of different batches are introduced. And the seed image pixel ratio parameters corresponding to different batches of different continuous frame images ; By design formula Calculate the first Characterization coefficient of rice seeds in batches Realize the calculation of the characterization coefficient and analyze its seed characterization information.

[0037] It needs to be explained that is the standard value obtained under the normal state of the abnormal area structure of the seed image of the current batch; The standard value obtained under normal conditions based on the pixel size of the seed image in the current batch; used as a comparison object for the measured value; usually simulated based on historical experience; standard deviation value of the image pixel ratio parameter The settings are selected based on empirical data and will not be described in detail here; the preset function It is an adjustment function set according to historical data to ensure that the obtained characterization coefficient is within a specific and reasonable range.

[0038] As an embodiment of the present invention, further, in step S23, by comparing the characterization coefficient with a preset standard threshold, analyzing and obtaining a batch characterization coefficient data set that meets the requirements and recording the batch quantity; the above method of obtaining the batch characterization coefficient that meets the requirements and recording the batch quantity is specifically as follows: The characterization coefficient of the current batch of rice seeds The standard characterization coefficient threshold interval obtained by machine simulation calculation in advance For comparison: like ∈ , it is judged that the characterization of the current batch of rice seeds is normal, and the characterization coefficient of the current rice seeds and the number of batches are recorded; like < , it is judged that the current batch of rice seeds is abnormally small; like > , it is judged that the characteristics of the current batch of rice seeds are abnormally large.

[0039] In the above technical solution, the characterization of rice seeds is judged by analyzing the range of the characterization coefficients, and whether it is suitable for internal detection and analysis of rice seeds; a seed batch that meets the requirements for further analysis is obtained; the specific judgment process is to determine whether the characterization coefficients of the current batch of rice seeds are sufficient. The standard characterization coefficient threshold interval obtained by machine simulation calculation in advance Compare the sizes to determine whether the preset standard characterization coefficient threshold interval is met. If so, the characterization of the current batch of rice seeds is judged to be normal, and the current rice seed characterization coefficient and the number of batches are recorded. Otherwise, the characterization is judged to be abnormal.

[0040] The characterization coefficient obtained from the analysis is compared with a preset standard threshold to evaluate which batches of seed images meet the predetermined quality standards; batches that are consistent and meet the requirements can be screened out, and the number of these batches can be recorded for further analysis.

[0041] As an embodiment of the present invention, the third is: then, the spectral data of the target sample is collected through a hyperspectral imaging system, and the spectral data is preprocessed and the spectral marker signal is extracted; by using the hyperspectral imaging system to collect spectral data of the target sample, the reflection or absorption spectrum of the seed at different wavelengths is obtained; according to the spectral data preprocessing process, the collected spectral data is preprocessed, such as removing noise, correcting background, etc., to ensure data quality; by extracting key spectral marker signals from the preprocessed spectral data, it is determined that these signals are related to the intrinsic quality of the seeds (such as nutritional components, moisture content, etc.).

[0042] In the specific step S3: The radiation information of the target sample under the preset light wave is obtained by continuous spectral scanning and recorded as the original spectral data; Perform spectral wavelength correction and noise reduction pre-processing on the original spectral data; Obtain the preprocessed spectral data to extract characteristic wavelengths and calculate them through the random Senli model; The importance scores of all characteristic wavelengths calculated are normalized, sorted, and the order is output as a spectral marker signal.

[0043] In the above technical scheme, since the target sample is usually obtained by continuous spectral scanning in image information of different frames, it is necessary to further screen the image of the target sample. The high-spectral frequency camera can be used for dynamic recognition and capture through real-time adjustment and non-uniform correction, and the radiation information under the preset light wave is recorded as the original spectral data; then the original spectral data is subjected to spectral wavelength correction and noise reduction pre-processing; the correction process specifically includes: the original spectral data collected by the high-spectral frequency camera not only contains the spectral characteristics of the light source, but also contains noise such as weak ambient light and dark current, and cannot truly reflect the reflection spectrum of the sample to be tested.

[0044] Before the sample test, the different reflectance spectra of the test plate and the control plate are collected in advance to eliminate interference. The correction formula is as follows:

[0045] In the formula, is the corrected reflectivity; is the original spectrum; is the reflection spectrum of the test plate; is the reflectance spectrum of the control plate.

[0046] The process of obtaining preprocessed spectral data and extracting characteristic wavelengths is to use smoothing algorithms, multivariate scattering correction, etc. for preprocessing; then construct an image data set and use the random forest algorithm to score the wavelengths; then evaluate them by their importance, and use the sorting selector to select the wavelength with the highest importance as the final pipe laying marking signal to facilitate the next step of analysis.

[0047] The specific method for obtaining the wavelength importance evaluation is: Assume that the collected data sample characteristics ; First calculate the Gini index of each feature:

[0048] In the formula, The value is the number of categories, For the node Medium Category The proportion of random slave nodes The probability that two samples are drawn from the same source and their category labels are inconsistent.

[0049] Also calculated features At the node Importance ; Through the formula:

[0050] in, For each feature Gini index; , They are random forest nodes Two new nodes after branching , of index; And judge: If the feature In the decision tree Nodes that appear in the preset node set Among them, then:

[0051] For the current characteristic wavelength In random forests, it is assumed that Based on the situation of the decision tree, the importance of the current characteristic wavelength is evaluated:

[0052] In the above technical solution, the importance of each feature wavelength is scored through a formula, and the average change in the node split impurity of the current feature in all decision trees of the random forest is analyzed.

[0053] As an implementation mode of the present invention, the fourth is: finally constructing a convolutional neural network model, inputting the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and outputting the adjustment result.

[0054] As an embodiment of the present invention, please refer to Figure 5 As shown, it also includes adjusting the inclined collection plate with different inclination angles, and analyzing the rolling speed of the rice seeds on the inclined collection plate to determine the optimal rolling state of the seeds; the adjustment angle is adjusted by adjusting the adjustment rod, and the first collection plate is rotated to determine different inclination angles.

[0055] See also Figure 4 As shown, the rice seed intelligent screening technology control system based on deep learning includes: An information collection module is used to obtain several batches of rice seeds, collect continuous frame RGB images of rice seeds rolling on the inclined collection plate within a preset period of time for each batch using a high-speed camera, and analyze the continuous frame RGB image information of all batches of rice seeds; A sample selection module is used to extract image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and record the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; An information extraction module is used to collect spectral data of a target sample through a hyperspectral imaging system, preprocess the spectral data and extract spectral marker signals; The model building module is used to build a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

[0056] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0057] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0058] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined in this application, they shall all fall within the protection scope of the present invention.

Claims

1. A rice seed intelligent screening technology method based on deep learning, characterized in that: The method comprises: S1. Obtain several batches of rice seeds, use a high-speed camera to collect continuous frame RGB images of rice seeds of each batch rolling on an inclined collection plate within a preset period of time, and analyze the continuous frame RGB image information of all batches of rice seeds; S2, extracting image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and recording the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; S3, collecting spectral data of the target sample through a hyperspectral imaging system, preprocessing the spectral data and extracting spectral marker signals; S4. Construct a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

2. The rice seed intelligent screening technology method based on deep learning according to claim 1 is characterized in that: In step S1: S11, obtaining a set of continuous frame RGB image pixels of rice seeds from different batches within the same time period; S12, extracting multiple continuous frame RGB image pixel sets of the same frame rate from each batch and arranging them in time sequence; S13. Analyze the rolling characteristics of all batches of rice seeds based on the time-series arranged RGB image pixel sets.

3. The rice seed intelligent screening technology method based on deep learning according to claim 2 is characterized in that: In step S2: S21, obtaining the contour sizes of abnormal regions of multiple batches of images and pixel ratio parameter values ​​of multiple batches of seed images corresponding to the rolling feature analysis of each batch of rice seeds; S22, analyzing the contour sizes of abnormal regions of multiple batches of images and the acquisition characterization coefficients of pixel ratio parameter values ​​of multiple batches of seed images; S23, comparing the characterization coefficient with a preset standard threshold, analyzing and obtaining a batch characterization coefficient data set that meets the requirements, and recording the batch quantity; S24. Record the data set of batch characterization coefficients that meet the requirements as training samples and input them into the neural network model for training. Take the batch of characterization coefficients corresponding to the obtained optimal simulation parameters as output and record them as target samples.

4. The rice seed intelligent screening technology method based on deep learning according to claim 3 is characterized in that: The method for obtaining the characterization coefficient in step S22 is: By formula Calculate the first Characterization coefficient of rice seeds in batches ; in, The total number of continuous frame images of rice seeds rolling on the inclined acquisition plate within a preset period of time for the same batch of detection; ∈ ; is the preset function; For the Batch No. Abnormal area structure value of consecutive frame images; For the Batch No. The seed image pixel ratio parameter of consecutive frame images; For the The standard value of abnormal area structure of batches of continuous frame images; For the The standard value of the seed image pixel ratio parameter of batches of continuous frame images; For the Batch No. The standard deviation of the seed image pixel ratio parameter of consecutive frame images; is a first weight coefficient preset according to the influence of the contour of the abnormal area of ​​the current image, A second weight coefficient is preset according to the influence of the seed image pixel ratio parameter value of the current image; and , Both are greater than 0.

5. The method for intelligent screening of rice seeds based on deep learning according to claim 4, characterized in that: The method of obtaining the batch characterization coefficient that meets the requirements and recording the batch quantity in step S23 is: The characterization coefficient of the current batch of rice seeds The standard characterization coefficient threshold interval obtained by machine simulation calculation in advance For comparison: like ∈ , it is judged that the characterization of the current batch of rice seeds is normal, and the characterization coefficient of the current rice seeds and the number of batches are recorded; like < , it is judged that the current batch of rice seeds is abnormally small; like > , it is judged that the characteristics of the current batch of rice seeds are abnormally large.

6. The rice seed intelligent screening technology method based on deep learning according to claim 1 is characterized in that: In step S3: The radiation information of the target sample under the preset light wave is obtained by continuous spectral scanning and recorded as the original spectral data; Perform spectral wavelength correction and noise reduction pre-processing on the original spectral data; Obtain the preprocessed spectral data to extract characteristic wavelengths and calculate using the random Senli model: By formula Calculate the features At the node Importance ;in, For each feature Gini index; , They are random forest nodes Two new nodes after branching , of index; And judge: If the feature In the decision tree Nodes that appear in the preset node set Among them, then: ； For the current characteristic wavelength In random forests, it is assumed that Based on the situation of the decision tree, the importance of the current characteristic wavelength is evaluated: ； The importance scores of all characteristic wavelengths calculated are normalized, sorted, and the order is output as a spectral marker signal.

7. The rice seed intelligent screening technology method based on deep learning according to claim 1 is characterized in that: The process of identifying the contour of the abnormal area of ​​the image is as follows: The continuous frame RGB rice seed images rolling on the tilted acquisition plate are grayed out; The Canny algorithm is used to extract the contour of the abnormal area of ​​the image within the range of rice seeds in the current image; Calculate the number of contours of abnormal regions of the image to form a first set A; The ratio of the integrated area of ​​the contours of the abnormal regions of the images of rice accumulated at the same time point in different batches to the number of contours of the corresponding abnormal regions of the images is obtained as an element to form a second set B; Compare the elements x in the second set B with the standard set D one by one: like , then it is determined that the rice seeds of the element x corresponding to the abnormal area contour of the image at the corresponding time point of the current batch have normal quality; , it is determined that there is a problem with the quality of rice seeds in the contour of the abnormal area of ​​the image corresponding to the element x at the corresponding time point of the current batch.

8. The rice seed intelligent screening technology method based on deep learning according to claim 1, characterized in that: The method also includes adjusting the inclined collection plate with different inclination angles, and analyzing the rolling speed of the rice seeds on the inclined collection plate to determine the best rolling state of the seeds.

9. A rice seed intelligent screening technology control system based on deep learning, characterized in that: The system comprises: An information collection module is used to obtain several batches of rice seeds, collect continuous frame RGB images of rice seeds rolling on the inclined collection plate within a preset period of time for each batch using a high-speed camera, and analyze the continuous frame RGB image information of all batches of rice seeds; A sample selection module is used to extract image contour information and image pixel ratio of each batch of rice seeds to determine the batch with the best rice seed representation, and record the current best batch of rice seeds as the target sample, wherein the image contour information includes the contour of the abnormal area of ​​the image; An information extraction module is used to collect spectral data of a target sample through a hyperspectral imaging system, preprocess the spectral data and extract spectral marker signals; The model building module is used to build a convolutional neural network model, input the spectral marker signal and the RGB image pixel information of the current frame number of the target sample into the model for training, and output the adjustment result.

Images