Seed germination information acquisition method based on deep learning

Abstract

The invention discloses a seed germination information acquisition method based on deep learning, and belongs to the technical field of germination information detection.The method comprises the steps that seeds are cultivated; acquiring images of the seeds in different germination states at the initial stage, the middle stage and the later stage; marking the obtained image; dividing the marked data, and performing image enhancement on the divided data; the YOLOv7-large model is improved, and training and verification are carried out on the YOLOv7-large model; detecting the trained YOLOv7-large model, and carrying out numbering, screenshot and storage on the detected germinated seed target; and processing the stored germinated seeds, and converting corresponding bud length data based on a processing result in combination with the shooting distance of the camera. According to the method, the seeds in different germination stages are rapidly detected through the improved YOLOv7-large model, and the germination rate and the corresponding bud length of the seeds are calculated according to the detection result, so that the cost is remarkably reduced, the working efficiency is greatly improved, and errors caused by subjective factors are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of germination information detection, and in particular to a method for acquiring seed germination information based on deep learning. Background Art

[0002] The germination rate of seeds is an important indicator for evaluating seed vitality and growth potential, and is of great significance to fields such as plant breeding and agricultural production. Each batch of seeds needs to be tested for germination in the laboratory before being sold or used to evaluate whether the batch of seeds meets the sales standards. A higher germination rate is an important measure of seed quality, which usually indicates that the seeds can germinate more quickly and show stronger environmental adaptability. In contrast, seeds with a lower germination rate often face problems such as slow germination and weak growth, and are easily restricted by environmental factors, resulting in reduced yields of agricultural products. When the length of the coleus exceeds 2 mm after seed germination, the seeds are generally considered to have germinated.

[0003] Traditional detection methods have the following defects:

[0004] (1) Traditional seed germination rate detection methods mainly rely on manual observation and manual operation;

[0005] (2) Traditional evaluation methods not only consume a lot of manpower and time, but also have difficulty in accurately quantifying the seed germination rate and are easily affected by the external environment. They are also highly subjective and inefficient.

[0006] (3) When detecting the germination rate of high-density seeds, the existing detection technology is easily affected by the radicle entanglement, resulting in a decrease in accuracy and prone to false detection and missed detection.

[0007] (4) The existing detection technology is applicable to a small number of varieties and is only suitable for the germination rate detection of some varieties of seeds. In addition, when faced with seeds of different sizes, the detection rate of the detection technology is not high.

[0008] (5) Existing detection technologies only provide limited information after detection, including germination rate results, and lack bud length information and corresponding bud image number recording functions.

[0009] Currently, automated detection methods based on image processing and machine learning are receiving widespread attention. These methods use computer vision technology to monitor and analyze the seed germination process, which not only significantly reduces human errors, but also processes large amounts of data in a short period of time, greatly improving detection efficiency. However, despite the great potential of these emerging technologies, there is still room for further improvement in the objectivity, repeatability, and generalization of detection. Therefore, there is an urgent need for a more objective, repeatable, and efficient detection method to replace traditional manual detection methods and simple detection techniques. Summary of the invention

[0010] The purpose of the present invention is to provide a seed germination information acquisition method based on deep learning. The seeds at different germination stages are quickly detected through an improved YOLOv7-large model, and the germination rate of the seeds and the corresponding sprout length are calculated accordingly, which effectively solves the problem of consuming a lot of manpower and material resources in traditional manual measurement, not only significantly reduces the cost, but also greatly improves the work efficiency and reduces the errors caused by subjective factors.

[0011] To achieve the above object, the present invention provides a method for obtaining seed germination information based on deep learning, the steps comprising:

[0012] S1. Place the culture dish in a germination box with constant temperature and humidity to cultivate the seeds;

[0013] S2, acquiring images of seeds in different germination states at the early, middle and late stages of germination;

[0014] S3, annotate the acquired image, label the germinated seeds as "Germinated", label the ungerminated seeds as "Seed", generate a label file, and export the file into XML format;

[0015] S4, dividing the labeled data into a training set, a validation set, and a test set, and performing image enhancement on the training set, the validation set, and the test set;

[0016] S5. Improve the YOLOv7-large model, use the image-enhanced training set to train the improved YOLOv7-large model, and use the validation set for validation;

[0017] S6. Use the test set to test the trained YOLOv7-large model, and number and take screenshots of the detected germinating seed targets;

[0018] S7, sequentially performing binarization processing, expansion blurring and skeleton extraction processing on the stored germinated seed image, and recording the obtained skeleton according to the longest distance and converting the corresponding sprout length data in combination with the corresponding camera shooting distance.

[0019] Preferably, step S2 includes: acquiring images of seeds in different germination states after germination for 24 hours, 48 ​​hours and 72 hours, and the relative height between the camera and the culture dish is 37 cm and 15 cm each time the image is acquired.

[0020] Preferably, the improvement of the YOLOv7-large model in step S5 includes: replacing the second layer of the backbone feature extraction network in the YOLOv7-large model with an inner convolutional layer, introducing a display visual center module after the last layer output of the backbone feature extraction network, and introducing a spatial context pyramid module in the enhanced extraction network. The improved YOLOv7-large model includes an input layer, a backbone feature extraction network, an enhanced feature extraction network and a detection head.

[0021] Preferably, the training process includes:

[0022] The improved YOLOv7-large model takes the acquired seed germination image as initial data and enters the backbone feature extraction network. After the inner convolution, the channel is compressed to obtain the initial seed features. After that, the backbone feature extraction network undergoes multiple feature samplings to obtain three effective feature layers of different sizes. The fifth effective feature layer passes through the display visual center module and serves as a new effective feature layer. The three effective feature layers undergo multiple feature fusions in the feature enhancement network to obtain the feature information after three layers of processing. The feature information is processed by the spatial context pyramid module to focus on the features of seed targets of different scales. Finally, the multiple prediction boxes generated in the previous step are matched with the real boxes in the detection head to obtain the regression and classification of the seed targets.

[0023] Preferably, the binarization process in step S7 includes:

[0024]

[0025] In the formula, f(x,y) represents the grayscale value of the input image at the coordinate (x,y), T represents the threshold, and g(x,y) represents the binarized image.

[0026] Preferably, the morphological expansion in step S7 includes:

[0027]

[0028] In the formula, A represents the pixel set of the input binary image, B represents the pixel set of the structural element, (B) z represents the translation of the structural element B at position z in the image, and A⊕B represents the set of images after expansion.

[0029] Preferably, the skeleton extraction process in step S7 includes:

[0030]

[0031] In the formula, A k represents the kth iteration result of the image, S represents the structural element, ⊕ represents the expansion operation, Represents an erosion operation.

[0032] Therefore, the present invention adopts the above-mentioned seed germination information acquisition method based on deep learning, which has the following beneficial effects:

[0033] (1) Use a camera to collect information about seeds in different germination states, and use the improved YOLOv7-large model to quickly detect them, and then calculate the germination rate and corresponding sprout length of the seeds, which is non-destructive;

[0034] (2) It effectively solves the problem of consuming a lot of manpower and material resources in traditional manual measurement, which not only significantly reduces costs, but also greatly improves work efficiency and reduces errors caused by subjective factors;

[0035] (3) It is suitable for seed testing at different growth stages, has a wide range of application scenarios, and can meet a variety of testing needs;

[0036] (4) Ability to quickly evaluate the germination rate and seed vigor of a test batch of seeds.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A method flow chart of an embodiment of a method for acquiring seed germination information based on deep learning according to the present invention;

[0039] Figure 2 This is a diagram of the improved YOLOv7-large model architecture of an embodiment of a method for acquiring seed germination information based on deep learning in the present invention. DETAILED DESCRIPTION

[0040] Example

[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0042] Reference Figure 1 The present invention provides a method for obtaining seed germination information based on deep learning, and provides a method for obtaining seed germination information based on deep learning, the steps comprising:

[0043] S1. Place the culture dish in a germination box with constant temperature and humidity to cultivate the seeds.

[0044] Taking cabbage seeds as an example, soak them in 75% ethanol for 1 minute, rinse them with distilled water, and then place them on filter paper with a diameter of 9 cm and a culture dish of the same size. Then, place the culture dish in a germination box with constant temperature and humidity to cultivate the seeds.

[0045] S2. Obtain images of seeds in different germination states at the early, middle and late stages of germination.

[0046] Specifically, a 20-megapixel industrial area array camera MV-CS200-10UC was used to obtain images of seeds in different germination states at 24 hours, 48 ​​hours, and 72 hours of germination, and the relative height between the camera and the culture dish was 37 cm and 15 cm each time the image was obtained.

[0047] S3. Label the acquired images using Labelimg1.8.6 labeling software. The germinated seeds are labeled as “Germinated”, the ungerminated seeds are labeled as “Seed”, and a label file is generated. The file is exported to XML format.

[0048] S4. Divide the labeled data into training set, validation set and test set in a ratio of 4:1:1, and perform image enhancement on the training set, validation set and test set. The training set is used for model training and optimization; the test set and validation set are used to evaluate the performance and generalization ability of the model respectively.

[0049] It should be noted that the images contained in the validation set and the test set are completely different from those in the training set.

[0050] S5. Improve the YOLOv7-large model, use the image-enhanced training set to train the improved YOLOv7-large model, and use the validation set for verification.

[0051] Specifically, the improvements to the YOLOv7-large model include: replacing the second layer of the backbone feature extraction network in the YOLOv7-large model with an inner convolutional layer to increase the spatial specificity of the features, introducing the explicit visual center module (EVC) after the output of the last layer of the backbone feature extraction network to achieve feature adjustment within the centralized layer, achieve a more comprehensive and accurate representation of seed features, and improve independent recognition of seeds. Introducing the spatial context pyramid module (SCP) in the enhanced extraction network, in response to the problem of incomplete recognition of small-sized seeds, the spatial context pyramid module further enhances the detail features of small seeds by learning the global spatial context of each level. The improved YOLOv7-large model includes an input layer, a backbone feature extraction network, an enhanced feature extraction network, and a detection head, such as Figure 2 shown.

[0052] The training process includes:

[0053] The improved YOLOv7-large model enters the acquired seed germination image into the backbone feature extraction network in the form of initial data, and performs channel compression through inner convolution to obtain the initial seed features; then, the backbone feature extraction network undergoes multiple feature samplings to obtain three effective feature layers of different sizes; the fifth effective feature layer improves the large receptive field sensitivity in the feature stream through the display visual center module, and uses the lightweight MLP and encoding capabilities in the module, and serves as a new effective feature layer; the three effective feature layers (two layers of features obtained from the backbone and one layer of features processed by the EVC module) are subjected to multiple feature fusions in the feature enhancement network to obtain the feature information after three layers of processing; the feature information is processed by the SCP module, and the features of seed targets of different scales are given special attention. Finally, the multiple predicted boxes generated in the previous step are matched with the real boxes in the detection head to obtain the regression and classification of the seed targets.

[0054] S6. Use the test set to test the trained YOLOv7-large model, and number and save the detected germinating seed targets.

[0055] S7, sequentially performing binarization processing, expansion blurring and skeleton extraction processing on the stored germinated seed image, and recording the obtained skeleton according to the longest distance and converting the corresponding sprout length data in combination with the corresponding camera shooting distance.

[0056] Binarization is used to convert a grayscale image into an image with only two grayscale values, usually used for object segmentation and contour extraction. It includes:

[0057]

[0058] Where f(x,y) represents the grayscale value of the input image at the coordinate (x,y), T represents the threshold, and g(x,y) represents the binarized image, where the pixel value is either 0 (indicating black) or 1 (indicating white).

[0059] Morphological dilation actually "expands" or "grows" the foreground pixels, thereby increasing the size of the foreground area, including:

[0060]

[0061] Where A represents the pixel set of the input binary image, B represents the pixel set of the structural element, which is usually a small matrix or template, (B) z represents the translation of the structural element B at position z in the image, and A⊕B represents the set of images after expansion.

[0062] The purpose of skeleton extraction is to extract a refined version of the object in the image, maintaining the topological structure of the object but reducing its width. It includes:

[0063]

[0064] In the formula, A k represents the kth iteration result of the image, S represents the structural element, ⊕ represents the expansion operation, Represents an erosion operation.

[0065] In order to evaluate the performance of the improved detection algorithm, a comparative experiment was conducted on the mainstream detection algorithms in the current target detection field based on the same equipment, the same data set, and the same parameter settings. The comparison was made by comparing the F1 value, the mean average precision (mAP), and the number of parameters (Params). The specific research results are shown in Table 1, where the content in brackets is the model version or size, DC5, 4scale refers to the version, s refers to the model small size, m refers to the model medium size, l refers to the model large size, and x refers to the model extra large size.

[0066] Table 1 Comparison of results of different detection models

[0067]

[0068]

[0069] First, the experimental models based on the Transformer framework include DETR, Rt-DETR, DINO, and ViDT-swin. From the perspective of data distribution, the AP of these models at the Easy level in the data is generally stronger than that of the convolutional framework model, with the highest being 94.9% in Rt-detr(l). This may be due to the Transformer's powerful long-distance semantic capture ability, which enables it to perform well in the detection of germinating seeds with a large proportion. However, this advantage also limits the model's positioning and recognition of small targets to a certain extent. In the detection of the Hard level, it is generally weaker than the convolutional framework model, with the highest being only 83.2% in ViDT-swin. At the same time, in terms of the F1 index, these models perform generally, with the highest being 85.1% of Rt-detr(x). It can be seen that for data sets of non-uniform size, the performance of the trained model in predicting whether it is a germinating seed is not high. In addition, in terms of the number of model parameters, the model is large in scale, and Rt-detr(l) with the smallest Params value is 52.7M. It is worth mentioning that due to the characteristics of the Transformer framework, a large number of matrix operations are added to the model, which makes the training process slow and often requires a long training time to gradually converge.

[0070] Except for the Transformer framework-based models, the rest are convolutional framework models. In the Easy category detection, YOLOv7(l) has the best AP value of 95.4%; in the Hard category detection, YOLOv8(x) has the best AP value of 91.8%; in terms of mAP and F1 indicators, YOLOv7(l) has the highest values, 82.8% and 88.9% respectively. FasterR-CNN, as a classic two-stage detection model, does not have outstanding values ​​in the Easy and Hard categories. YOLOX has a medium detection level, among which YOLOX(s) has the smallest model parameter of only 8.97M, but the Hard category detection effect is lacking, which is 4.2% lower than YOLOv8(l)'s 91.8%. YOLOv7(l) and YOLOv7(x) have a small difference in mAP, only 0.7%, but a difference of 4.0% in F1 value, and the model parameter size is half as small. Overall, YOLOv7(l) performs better. YOLOv8 combines the various advantages of the YOLO series models. Among them, the improvement of the P5 and P6 feature layers can make good use of various information when the model copes with large-resolution images, which makes the performance of Hard class detection better, and the highest value exceeds YOLOv7(l) by 1.6%. From the perspective of mAP, the difference between the models of s, m, l, and x sizes is small, and the maximum and minimum values ​​differ by only 2.4%. This shows that YOLOv8 has limited further utilization of the data used in this embodiment, and the change in model scale will not cause a significant improvement in the feature extraction effect. The F1 value of YOLOv8(l) is the highest among YOLOv8, but it is still not as high as that of YOLOv7(l), with a difference of 0.4%.

[0071] Combined with the data in the table, YOLOv7(l) showed the best results in Easy AP, mAP, and F1 indicators, which were 95.4%, 92.8%, and 88.9%, respectively. Even in Hard detection, it was only 1.6% different from the highest value. In addition, in terms of the number of model parameters, 37.62M is relatively small compared to most models, and will not put too much burden on the equipment. The improved model achieved the best results in all indicators. Under the model parameters of 45.22M, YOLOv7(l) was 4.2% and 6.2% higher in Easy and Hard detection, respectively, and mAP and F1 were increased by 5.2% and 4.4%, respectively. In particular, the improvement was obvious in the detection of small-sized germinating seeds that were difficult to detect, achieving the highest comprehensive performance in the comparative experiment.

[0072] Therefore, a seed germination information acquisition method based on deep learning is proposed. It uses the improved YOLOv7-large model to quickly detect seeds at different germination stages, and calculates the seed germination rate and the corresponding sprout length based on this. It effectively solves the problem of consuming a lot of manpower and material resources in traditional manual measurement. It not only significantly reduces costs, but also greatly improves work efficiency and reduces errors caused by subjective factors.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for obtaining seed germination information based on deep learning, characterized in that the steps include: S1. Place the culture dish in a germination box with constant temperature and humidity to cultivate the seeds; S2, acquiring images of seeds in different germination states at the early, middle and late stages of germination; S3, annotate the acquired image, label the germinated seeds as "Germinated", label the ungerminated seeds as "Seed", generate a label file, and export the file into XML format; S4, dividing the labeled data into a training set, a validation set, and a test set, and performing image enhancement on the training set, the validation set, and the test set; S5. Improve the YOLOv7-large model, use the image-enhanced training set to train the improved YOLOv7-large model, and use the validation set for validation; S6. Use the test set to test the trained YOLOv7-large model, and number and take screenshots of the detected germinating seed targets; S7, sequentially performing binarization processing, expansion blurring and skeleton extraction processing on the stored germinated seed image, and recording the obtained skeleton according to the longest distance and converting the corresponding sprout length data in combination with the corresponding camera shooting distance.

2. A method for acquiring seed germination information based on deep learning according to claim 1, characterized in that: Step S2 includes: acquiring images of seeds at different germination states after germination for 24 hours, 48 ​​hours and 72 hours, and the relative height between the camera and the culture dish is 37 cm and 15 cm each time the image is acquired.

3. The method for obtaining seed germination information based on deep learning according to claim 1, characterized in that: The improvement of the YOLOv7-large model in step S5 includes: replacing the second layer of the backbone feature extraction network in the YOLOv7-large model with an inner convolution layer, introducing a display visual center module after the last layer output of the backbone feature extraction network, and introducing a spatial context pyramid module in the enhanced extraction network. The improved YOLOv7-large model includes an input layer, a backbone feature extraction network, an enhanced feature extraction network and a detection head.

4. The method for obtaining seed germination information based on deep learning according to claim 3, characterized in that: The training process includes: The improved YOLOv7-large model takes the acquired seed germination image as initial data and enters the backbone feature extraction network. After the inner convolution, the channel is compressed to obtain the initial seed features. After that, the backbone feature extraction network undergoes multiple feature samplings to obtain three effective feature layers of different sizes. The fifth effective feature layer passes through the display visual center module and serves as a new effective feature layer. The three effective feature layers undergo multiple feature fusions in the feature enhancement network to obtain the feature information after three layers of processing. The feature information is processed by the spatial context pyramid module to focus on the features of seed targets of different scales. Finally, the multiple prediction boxes generated in the previous step are matched with the real boxes in the detection head to obtain the regression and classification of the seed targets.

5. The method for obtaining seed germination information based on deep learning according to claim 1, characterized in that: The binarization process in step S7 includes: In the formula, f(x,y) represents the grayscale value of the input image at the coordinate (x,y), T represents the threshold, and g(x,y) represents the binarized image.

6. A method for acquiring seed germination information based on deep learning according to claim 5, characterized in that: The morphological dilation in step S7 includes: In the formula, A represents the pixel set of the input binary image, B represents the pixel set of the structural element, (B) z represents the translation of the structural element B at position z in the image, Represents the set of images after dilation.

7. The method for obtaining seed germination information based on deep learning according to claim 5, characterized in that: The skeleton extraction process in step S7 includes: In the formula, A k represents the kth iteration result of the image, S represents the structural element, represents the expansion operation, Represents an erosion operation.

Images