Seed germination segmentation method, seed germination rate and germination rate identification method

Abstract

The invention relates to a seed germination segmentation method, a seed germination rate and a germination rate identification method, which belong to the field of image recognition technology and solve the problem that the existing YOLOv8-seg has low accuracy in segmenting small target seed germination. The seed germination image is input into VT-YOLOv8-seg for feature extraction, and the seed germination image is segmented; in the Backbone of VT-YOLOv8-seg, the seed germination image is input into a multi-level downgraded convolution layer for feature extraction, and then input into a dynamic snake convolution feature extraction module for dimensionality reduction, and outputs a seed germination feature map after dimensionality reduction; the seed germination feature map after dimensionality reduction is input into a downgraded convolution layer for feature extraction, and then input into a feature extraction module, a multi-level convolution layer and a feature extraction module in sequence for feature extraction, and then input into a feature optimization pooling layer for pooling, and outputs a seed germination feature map after pooling.

Description

Technical Field

[0001] The invention relates to the technical field of image recognition, and in particular to a seed germination segmentation method, a seed germination rate and a germination rate recognition method thereof. Background Art

[0002] Traditional seed quality assessment methods, especially seed vigor testing, are time-consuming and laborious. Repeatable germination rate testing requires manual statistics and measurement of relevant indicators every day during the actual test process, which often limits the efficiency and scale of the test, and is prone to introduce subjective errors, resulting in inconsistent statistical results between different personnel and poor repeatability. Therefore, an objective, repeatable, fast, economical and reliable measurement method is needed. Most seed tests developed by ISTA (International Seed Testing Association) are manually evaluated using different standardized procedures for different crops. We need a faster and more efficient seed germination detection method.

[0003] Machine vision technology has developed rapidly and has been widely used in the agricultural field. Researchers have proposed a seed quality detection method based on machine vision technology. GERMINATOR (automatic scoring model) is an automatic scoring model for Arabidopsis and Brassica seeds. It can evaluate germination based on the color contrast between the radicle and the seed coat in a single image, fit the germination curve, and extract and visualize various germination parameters. However, for seeds of other species, a lot of manual experiments are still required to set appropriate thresholds, which is still a very error-prone and cumbersome step. Importantly, the system is most likely to fail due to changes in lighting or partial occlusion of seeds. SVIS (seed vigor imaging model) is used to automatically evaluate the viability of three-day-old soybean and corn seeds, using the RGB (red, green, and blue) pixel values ​​of digital scanned images to calculate the length of the seeds. When using a scanner instead of a camera, the lighting settings are standardized, which undoubtedly improves the performance of the system. When using a camera or even a mobile phone to obtain images, there is no standardized standard that will cause errors in the results; on the other hand, this method requires researchers to be present throughout the germination experiment, use a flatbed scanner for manual imaging, and then evaluate the seeds. Medeiros et al. combined image mapping and linear discriminant analysis models to predict seed germination. Awty-Carroll et al. used KNN (K nearest neighbor model) to improve the scoring accuracy of mangosteen seeds. There is also a RSGES (rice seed germination evaluation model) based on artificial neural network classifier, which is used to evaluate the germination status of Thai rice seeds. The digital images used in the model are taken with ordinary digital cameras or mobile phone cameras, which are very easy for farmers to handle. RSGES includes 6 main processing modules: 1) image acquisition, 2) image preprocessing, 3) feature extraction, 4) germination evaluation, 5) result display, and 6) germination verification. The seeds of Thai rice variety cp-8111 were tested in Bangkok and Chiang Mai, Thailand. RSGES extracted 18 features: 3 color features, 7 morphological features, and 8 texture features, with a final false positive rate of 7.6%. However, the classic machine learning classifier is based on manual feature extraction, with slow training speed and low accuracy.

[0004] In recent years, deep learning has developed rapidly and has been widely used in the field of agriculture, especially in the field of seed detection, recognition and germination. Compared with other first-level object detectors such as SSD (Solid State Drive), the YOLO (You Only Look Once) family has been continuously updated and improved in detection performance. The first-level convolutional neural detector represented by YOLO is faster in detection and has more important lightweight features. For example, WSVAS (Wheat Seed Vitality Assessment System) uses a lightweight convolutional neural network YOLOv4 (v4 version of YOLO) to quickly and accurately detect wheat seed germination, with a mAP (average detection accuracy) of 97.59% and a recall of 97.35%. SGR-YOLO (YOLO combined with convolutional neural network and region proposal network) is a new method for detecting wild rice germination rate using YOLO through deep learning technology. SGR-YOLO has a discriminant mAP of 94% and 98.2% for wild rice seed germination in hydroponic boxes and culture dishes, respectively. It can be seen that the YOLO series of algorithms can detect seed germination very well and have achieved good accuracy, but they only consider the task of detecting germination rate, and do not use instance segmentation methods to segment the embryo, and further explore the expression of seed vitality in the bud area. SeedAI (a seed germination prediction model based on a two-stage deep learning framework) uses Mask-RCNN (masked region convolutional neural network model) for seed segmentation in the first stage, and the RoI (region of interest) is used as input and input into the proposed CNN (convolutional neural network model) in the second stage for germination prediction. The accuracy of this model is 88%. SeedGerm (a cost-effective phenotyping platform) is used for automatic seed imaging and machine learning-based crop seed germination phenotyping. The system combines deep learning and machine learning methods, and uses five crops, tomatoes, peppers, brassicas, barley, and corn, as test materials to establish a large-scale germination scoring method, using a hybrid method of DL (deep learning) and ML (machine learning) to divide background (i.e. filter paper) and foreground (i.e. seeds) pixels. Both of these two germination detection systems based on the two-stage algorithm are highly complex and require high-power hardware. Although the above two systems use deep learning semantic segmentation technology, they only separate the background from the seeds themselves, but fail to separate the seeds from the embryos and further analyze the seed vitality.

[0005] YOLOv8 (v8 version of YOLO) is a popular real-time object detection algorithm that utilizes deep neural networks for object classification and localization. YOLOv8 is general purpose and is capable of handling a variety of visual tasks, including object detection, segmentation, pose estimation, tracking, and classification. It is an improved version of the YOLO series, which is widely recognized for its speed and accuracy in object detection tasks. Developed by the same authors as YOLOv5 (v5 version of YOLO) and inspired by the architecture of previous YOLO versions, YOLOv8 marks an iteration of refinement and improvement in real-time object detection and various related computer vision tasks. In YOLOv8, CSPDarknet53 (Cross-Stage Partially Connected Darknet53 Model) is used as the backbone for feature extraction, and the combination of SPPF layers (Spatial Pyramid Pooling) and PANet (Path Aggregation Network) helps merge feature maps of different scales, making it more robust in object detection of different object sizes. YOLOv8 uses an anchor-free approach and directly predicts bounding boxes and object classes without using anchor boxes as references. This allows for the flexibility to detect objects of different sizes, shapes, and irregular shapes. The anchor-free approach in YOLOv8 has been shown to be effective in producing accurate predictions.

[0006] In order to meet the needs of image segmentation, such as processing multi-scale information, refining boundaries and improving segmentation accuracy, Ultralytics (R&D team) continued the fast and accurate target detection capabilities of the YOLO series, and improved and optimized it on this basis, and designed YOLOv8-seg (YOLO's v8 version - segmentation), which is specifically used for image segmentation tasks. In the prior art, Chinese patent CN118505995A discloses "Understory Wide-angle Single Tree Image Segmentation Method and System Based on Improved YOLOv8-Seg", which improves the accuracy of understory single tree image segmentation under complex forest backgrounds by enhancing the segmentation ability of the model. However, seeds are small target objects, and the ratio of the actual area of ​​the seeds to the actual surface area of ​​the embryo is very different. For example, the diameter of corn seeds is about 5 to 10 mm, and the length may be about 10 to 20 mm. The ratio of the actual area of ​​the corn seeds to the actual area of ​​the embryo is about 1500:50. If the existing YOLOv8-seg is directly applied to seed germination segmentation, a large number of problems of missed detection or false detection of seed germination will occur, which greatly reduces the accuracy of seed germination segmentation. If the existing YOLOv8-seg seed germination segmentation results are directly applied to agricultural planting, it will directly reduce food production.

[0007] In summary, the existing YOLOv8-seg has low accuracy in segmenting the seed germination of small targets. Summary of the invention

[0008] The present invention solves the problem that the existing YOLOv8-seg has low accuracy in segmenting the seed germination of small targets.

[0009] The seed germination and segmentation method of the present invention is specifically as follows:

[0010] The seed germination image is input into the Backbone, Neck and Head of VT-YOLOv8-seg in sequence for feature extraction and segmentation of the seed germination image;

[0011] In the Backbone of VT-YOLOv8-seg, the seed germination image is input into the multi-level downconvolution layer for feature extraction, and then input into the dynamic snake convolution feature extraction module for dimensionality reduction, and the seed germination feature map after dimensionality reduction is output;

[0012] The seed germination feature map after dimensionality reduction is input into the downgraded convolution layer for feature extraction, and then sequentially input into the feature extraction module, the multi-level convolution layer and the feature extraction module for feature extraction, and then input into the feature optimization pooling layer for pooling, and the pooled seed germination feature map is output.

[0013] Furthermore, in one embodiment of the present invention, the downgraded convolution layer is specifically:

[0014] The image is divided into two paths. One path is input into the multi-stage convolutional layer to change its size. The image after the size change is input into the convolutional layer to change its number of channels, and the image after the number of channels change is output.

[0015] After the other image is added to the image with the changed number of channels, batch normalization and activation function are performed in sequence to output the feature map.

[0016] Furthermore, in one embodiment of the present invention, the feature optimization pooling layer is specifically:

[0017] After the image passes through the convolution layer, it is input into the adaptive pooling for multi-scale feature extraction, and then into the multi-head attention mechanism to capture the global context information, and then the feature map is output. The feature map is input into the multi-head attention mechanism and spatial pyramid pooling in turn, and the fused feature map is output.

[0018] Further, in one embodiment of the present invention, the feature map is sequentially input into the multi-head attention mechanism and the spatial pyramid pooling, and the fused feature map is output, specifically:

[0019] The feature map is divided into two paths, one of which is input into the multi-level maximum pooling, and each feature map after maximum pooling is divided into two paths again. One feature map after multi-level maximum pooling is connected with the other feature map after maximum pooling and the other feature map, and then input into the convolution layer for fusion, and the fused feature map is output.

[0020] The seed germination rate identification method of the present invention adopts any of the seed germination segmentation methods described in the above methods to segment the seed germination and identify the seed germination rate.

[0021] The seed germination rate identification method of the present invention realizes the identification of the seed germination rate based on the seed vitality analysis index and any seed germination segmentation method described in the above methods.

[0022] Furthermore, in one embodiment of the present invention, the seed vitality analysis index is specifically:

[0023] Based on the seed germination image segmented by VT-YOLOv8-seg, the seed germination mask image is obtained, and the actual area of ​​the embryo is obtained based on the seed germination mask image. The ratio of the actual area of ​​the embryo to the seed germination time is defined as the seed vitality analysis index.

[0024] The corn seed germination rate identification method of the present invention is implemented based on the seed germination rate identification method described in the above method.

[0025] The corn seed germination rate identification method of the present invention is implemented based on any of the seed germination rate identification methods described above.

[0026] The present invention solves the problem that the existing YOLOv8-seg has low accuracy in segmenting the seed germination of small targets. Specific beneficial effects include:

[0027] 1. The seed germination segmentation method described in the present invention, the existing YOLOv8-seg has a low accuracy in segmenting the seed germination of small targets, which will seriously affect the yield of agricultural products. In order to solve the above technical problems, the present invention improves the existing YOLOv8-seg and designs VT-YOLOv8-seg. Experimental data show that the VT-YOLOv8-seg can accurately segment small target seeds (such as corn seeds) and embryos, thereby solving the technical problem of low accuracy of the existing YOLOv8-seg segmentation of small target seeds;

[0028] 2. The seed germination segmentation method of the present invention can reduce the feature image by the downgraded convolution layer, which can play a filtering role and reduce the features of information redundancy, thereby retaining key information. At the same time, a smaller feature map means that the subsequent layers in the network need to process fewer parameters, thereby speeding up the training and reasoning process. And halving the size of the feature map by the downgraded convolution layer does not affect the extraction of features of small target seeds and embryos;

[0029] 3. The seed germination segmentation method described in the present invention improves the segmentation accuracy of seed germination for small targets in the Backbone of YOLOv8-seg. Considering that the multi-head attention mechanism can capture global context information, it helps the model to make more accurate judgments in complex scenarios. That is, the combination of the multi-head attention mechanism and the spatial pyramid pooling is intended to achieve accurate judgment of the seed germination of small targets. However, in actual operation, it is found that if the multi-head attention mechanism and the spatial pyramid pooling are simply combined, the maximum pooling layer in the spatial pyramid pooling will lose a lot of detailed information on seed germination, that is, it is impossible to capture the overall information of the seed germination of the small target. In order to solve the technical problem, the present invention introduces adaptive pooling to form a feature optimization pooling layer, thereby overcoming the technical difficulties and solving the technical problems described in the present invention;

[0030] 4. In the seed germination segmentation method of the present invention, the feature optimization pooling layer performs pooling operations of different sizes on the feature map and integrates the results to obtain a fixed-size output. This method can effectively handle targets with varying sizes, thereby improving the generalization ability and robustness of VT-YOLOv8-seg;

[0031] 5. The seed germination rate identification method of the present invention can realize the identification of seed germination rate based on seed vitality analysis index and VT-YOLOv8-seg;

[0032] The seed germination segmentation method of the present invention can be applied to any small target seeds, and the ratio of the actual area of ​​the seeds to the actual area of ​​the embryos is very different, and accurate segmentation of seeds and embryos can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0034] Figure 1 It is a VT-YOLOv8-seg structure diagram described in Implementation Method 1;

[0035] Figure 2It is a structural diagram of the downgraded convolutional layer described in Implementation Method 2;

[0036] Figure 3 It is a feature optimization pooling layer structure diagram described in Implementation Method 3;

[0037] Figure 4 is the experimental configuration diagram described in Implementation Example 7;

[0038] Figure 5 It is a performance evaluation diagram of VT-YOLOv8-seg in the training and verification process described in Implementation Method 7;

[0039] Figure 6 The existing YOLOv8-seg corn germ recognition image described in Implementation Example 7;

[0040] Figure 7 The VT-YOLOv8-seg corn germ recognition image described in Implementation Mode 7;

[0041] Figure 8 The existing YOLOv8-seg corn germ recognition image described in Implementation Example 7;

[0042] Fig. 9 The VT-YOLOv8-seg corn germ recognition image described in Implementation Mode 7;

[0043] Fig.10 The existing YOLOv8-seg corn germ recognition image described in Implementation Example 7;

[0044] Fig.11 The VT-YOLOv8-seg corn germ recognition image described in Implementation Mode 7;

[0045] Fig.12 The existing YOLOv8-seg corn germ recognition image described in Implementation Example 7;

[0046] Fig.13 The VT-YOLOv8-seg corn germ recognition image described in Implementation Mode 7;

[0047] Fig.14 7 is a segmentation result diagram of the five algorithms described in Implementation Mode 7;

[0048] Fig.15 is a curve diagram of embryo growth potential of corn seeds at different aging gradients according to the seventh embodiment;

[0049] Fig.16 This is a graph showing changes in corn seeds over seven days as described in Implementation Example 7. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe various embodiments of the present invention in conjunction with the accompanying drawings. The embodiments described with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0051] Embodiment 1: The seed germination and segmentation method described in this embodiment is specifically as follows:

[0052] The seed germination image is input into the Backbone, Neck and Head of VT-YOLOv8-seg in sequence for feature extraction and segmentation of the seed germination image;

[0053] In the Backbone of VT-YOLOv8-seg, the seed germination image is input into the multi-level downconvolution layer for feature extraction, and then input into the dynamic snake convolution feature extraction module for dimensionality reduction, and the seed germination feature map after dimensionality reduction is output;

[0054] The seed germination feature map after dimensionality reduction is input into the downgraded convolution layer for feature extraction, and then sequentially input into the feature extraction module, the multi-level convolution layer and the feature extraction module for feature extraction, and then input into the feature optimization pooling layer for pooling, and the pooled seed germination feature map is output.

[0055] In the prior art, since YOLOv8-seg cannot accurately segment the seed germination of small targets, the application of seed germination of small targets segmented by the existing YOLOv8-seg in the agricultural industry cannot maximize the benefits.

[0056] To solve the above technical problems, Figure 1 As shown, this embodiment designs VT-YOLOv8-seg (v8 version of VY-YOLO - segmentation), and segments the seed germination of the small target based on VT-YOLOv8-seg, specifically:

[0057] The seed germination image of the small target is input into the Backbone, Neck and Head of VT-YOLOv8-seg in sequence for feature extraction to achieve segmentation of the seed germination image of the small target;

[0058] In the Backbone of VT-YOLOv8-seg, the seed germination image of the small target is input into the two-level downgraded convolution layer for feature extraction, and then input into the dynamic snake convolution feature extraction module for dimension reduction, and the seed germination feature map after dimension reduction is output. The seed germination feature map after dimension reduction is input into the downgraded convolution layer for feature extraction, and then the feature extraction module outputs the seed germination feature map of the first size. The seed germination feature map of the first size is subjected to feature extraction by the convolution layer and the feature extraction module, and the seed germination feature map of the second size is output. The seed germination feature map of the second size is subjected to feature extraction by the convolution layer and the feature extraction module, and then input into the feature optimization pooling layer for pooling, and the pooled seed germination feature map is output.

[0059] In the Neck of VT-YOLOv8-seg, the pooled seed germination feature map is upsampled and connected with the seed germination feature map of the second size, and then they are jointly input into the feature extraction module for feature extraction, and a seed germination feature map of the third size is output. The seed germination feature map of the third size is upsampled and connected with the seed germination feature map of the first size, and then they are jointly input into the feature extraction module for feature extraction, and a seed germination feature map of the fourth size is output. The seed germination feature map of the fourth size is input into the convolutional layer, and then it is connected with the seed germination feature map of the third size, and the feature extraction module outputs a seed germination feature map of the fifth size. After the seed germination feature map of the fifth size passes through the convolutional layer, it is connected with the pooled seed germination feature map, and then it passes through the feature extraction module again to output a seed germination feature map of the sixth size.

[0060] In the Head of VT-YOLOv8-seg, the seed germination feature map of the fourth size, the seed germination feature map of the fifth size, and the seed germination feature map of the sixth size are input into the segmentation head together for feature segmentation.

[0061] In order to enhance the network's segmentation effect on curled and slender buds, the dynamic snake convolution feature extraction module is integrated into the Backbone of VT-YOLOv8-seg, allowing some 3×3 convolution kernels to adjust their shapes to adapt to the irregular structure of weeds, while maintaining the stability of the convolution structure and reducing the deviation, and constructing a convolution kernel with strong perception of irregular curves. In view of the good performance of the dynamic snake convolution feature extraction module in segmenting tubular structures, it adapts to the slender and twisted local structural characteristics, enhances the perception of geometric structures, and constructs a dynamic convolution layer with strong perception of irregular curves. Through technologies such as quantization and offset, the amount of calculation and memory usage can be greatly reduced, thereby accelerating the inference process.

[0062] Explain the application of depthwise separable convolution in the dynamic snake convolution feature extraction module in extracting local features of edge irregularities, assuming standard two-dimensional convolution coordinates , center coordinates The original 3×3 convolution kernel is , expressed as:

[0063] ; (1)

[0064] in, is the horizontal coordinate of the convolution kernel in the network, is the vertical coordinate of the convolution kernel in the network.

[0065] By introducing deformation offset , the convolution kernel becomes more flexible and focuses on the irregular edges of the tubular germ. The standard kernel is linearized in both axes and expanded to a kernel of size 9. Taking the axial direction as an example, Each grid position in is denoted as ,in =0,1,2,3,4 represents the horizontal distance from the center grid. Each grid position The selection is a cumulative process. Initially, the position of the grid away from the center depends on the position of the previous grid: Relative to Add a deformation offset Therefore, the offset needs to be accumulated to ensure that the kernel conforms to the linear structural form.

[0066] The change in the axis direction is:

[0067] ; (2)

[0068] The change in the axis direction is:

[0069] ; (3)

[0070] Due to deformation offset Usually decimal, while coordinates are usually in integer form, so bilinear interpolation is used, expressed as:

[0071] ; (4)

[0072] here, Indicates the decimal place in equations (2) and (3). Enumerate all integer spatial positions. is a bilinear interpolation kernel, which can be decomposed into two one-dimensional kernels:

[0073] ; (5)

[0074] in, , They are respectively Decompose to axis, The coordinate components on the axis, , They are Decompose to axis, The coordinate components on the axis, is a one-dimensional kernel.

[0075] Due to the two-dimensional changes, the dynamic snake convolutional feature extraction module covers a 9×9 range during deformation. The dynamic snake convolutional feature extraction module better adapts to the slender tubular structure based on the dynamic structure, thereby better perceiving the key features.

[0076] Therefore, the seed germination segmentation method described in this embodiment realizes the accurate segmentation of seed germination of small targets based on VT-YOLOv8-seg, and the improved VT-YOLOv8-seg improves the segmentation accuracy of tiny germs and curved germs, and increases the flexibility of the convolutional layer to adapt to germs of various shapes.

[0077] Embodiment 2: This embodiment further limits the seed germination segmentation method described in Embodiment 1. The degraded convolution layer is specifically:

[0078] The image is divided into two paths. One path is input into the multi-stage convolution layer to change its size. The image after the size change is input into the convolution layer to change its number of channels, and the image after the number of channels change is output.

[0079] After the other image is added to the image with the changed number of channels, batch normalization and activation function are performed in sequence to output the feature map.

[0080] This implementation scheme designs a downgraded convolutional layer, such as Figure 2 As shown, the first convolutional layer:

[0081] Convolution kernel is 3: Using a 3x3 convolution kernel can capture local features in the image.

[0082] Stride of 2: A stride of 2 means that the convolution operation moves 2 pixels at a time on the input feature map, which causes the size of the output feature map to be halved (in height and width).

[0083] Padding of 1: Add 1 extra pixel (usually zero padding) around the border of the input feature map to maintain the spatial dimension of the output feature map (relative to the case without padding). However, since the stride is 2, the size of the output feature map will still be halved, but padding helps to preserve the boundary information.

[0084] The second convolutional layer:

[0085] Similar to the first convolutional layer, this layer also uses a 3x3 convolution kernel, a stride of 2, and a padding of 1. Since the stride is still 2, this layer further reduces the size of the feature map by half.

[0086] After these two convolutional layers, the size of the feature map becomes 1 / 4 of the original size (in height and width).

[0087] The third convolutional layer:

[0088] Convolution kernels with a kernel size of 1:1x1 are often used to change the number of channels (depth) of a feature map without changing its spatial dimension.

[0089] Stride of 1 and padding of 0: A stride of 1 and no padding means that this convolutional layer is applied pixel by pixel and does not change the spatial size of the feature map. This layer is used to reduce or increase the number of channels in the feature map to accommodate the needs of subsequent layers.

[0090] Skip connections:

[0091] Skip connections add the original data directly to the output of the module. This helps solve the vanishing and exploding gradient problems in deep neural networks and allows the network to learn the identity mapping (i.e. input equals output) as a special case.

[0092] In this module, a skip connection adds the input feature map to the output of the third convolutional layer.

[0093] In general, this module reduces the spatial size of the feature map through two convolutional layers and adjusts the number of channels through a 1x1 convolutional layer. The skip connection helps to preserve the original information and facilitates the back propagation of the gradient.

[0094] Finally, after batch normalization and activation function processing, the feature map is output.

[0095] The downgraded convolutional layer described in this embodiment can effectively reduce the size of the feature map, thereby reducing the amount of calculation of subsequent layers. This is particularly important for large-scale image data and complex network structures. By reducing the size of the feature map, the computational efficiency of the YOLOv8 network can be improved, and the speed of training and reasoning can be accelerated.

[0096] Embodiment 3: This embodiment further limits the seed germination segmentation method described in Embodiment 1. The feature optimization pooling layer is specifically:

[0097] After the image passes through the convolution layer, it is input into the adaptive pooling for multi-scale feature extraction, and then into the multi-head attention mechanism to capture the global context information, and then the feature map is output. The feature map is input into the multi-head attention mechanism and spatial pyramid pooling in turn, and the fused feature map is output.

[0098] In this implementation, the feature map is sequentially input into the multi-head attention mechanism and the spatial pyramid pooling, and the fused feature map is output, specifically:

[0099] The feature map is divided into two paths, one of which is input into the multi-level maximum pooling, and each feature map after maximum pooling is divided into two paths again. One feature map after multi-level maximum pooling is connected with the other feature map after maximum pooling and the other feature map, and then input into the convolution layer for fusion, and the fused feature map is output.

[0100] In the prior art, the existing YOLOv8-seg is directly applied to the seed germination segmentation of small targets, which will cause a large number of missed detections or false detections of seed germination of small targets, greatly reducing the accuracy of seed germination segmentation of small targets. If the existing YOLOv8-seg seed germination segmentation results of small targets are directly applied to the planting of the agricultural industry, it will directly reduce food production. In order to solve the above technical problems, this embodiment improves the spatial pyramid pooling in the Backbone of YOLOv8-seg, considering that the multi-head attention mechanism can capture global context information, which helps the model make more accurate judgments in complex scenarios. That is, the intention is to achieve accurate judgment of the seed germination of small targets through the combination of multi-head attention mechanism and spatial pyramid pooling. However, in actual operation, it is found that if the multi-head attention mechanism and spatial pyramid pooling are simply combined, the maximum pooling layer in the spatial pyramid pooling will lose a lot of detailed information of the seed germination of the small target, that is, it is impossible to capture the overall information of the seed germination of the small target. In order to solve the above technical problems, this embodiment forms a feature optimization pooling layer by introducing adaptive pooling. The adaptive pooling layer can obtain more feature information by extracting features using different pooling kernel sizes, which can improve the network's recognition accuracy of small targets. At the same time, the multi-head attention mechanism can better and more obtain feature information at different scales, thereby obtaining global perspective information and reducing the impact of different scales, such as Figure 3 As shown, specifically:

[0101] First, the feature map is processed by a 1*1 convolution layer to reduce the amount of calculation and extract preliminary features. Next, three parallel adaptive pooling layers are used to adaptively adjust the size of the pooling kernel according to the size of the input feature map to reduce the impact of features of different scales. After the splicing operation, the fused feature map is input into the multi-head attention mechanism. The multi-head attention mechanism can combine information from feature maps of different sizes, increase the ability to capture global information, and use the correlation between pyramid feature maps to enhance the detection and segmentation capabilities of objects of different scales. Next, three maximum pooling operations are used to further extract features, which can extract multi-scale feature information and ensure that the output feature map has a fixed spatial dimension, which facilitates subsequent processing. After that, the feature map output by the multi-head attention mechanism and the output feature map after the three maximum pooling layers are spliced ​​to merge them into a larger feature map. This spliced ​​feature map contains feature information from different scales and different positions, providing rich input for subsequent convolutional layers. The last 1*1 convolution layer performs further convolution operations on the spliced ​​feature map to extract higher-level feature information. These feature information can be used for subsequent instance segmentation tasks.

[0102] Therefore, the feature optimization pooling layer described in this embodiment not only overcomes the defects of the prior art, but also improves the recognition accuracy of targets of different scales. It can achieve accurate focus on the seeds of small targets while also achieving accurate focus on the embryos of smaller targets, thereby achieving accurate segmentation of seed germination of small targets. The segmentation result will greatly improve the existing YOLOv8-seg segmentation result of seed germination of small targets, which is beneficial to improving the productivity of the agricultural industry.

[0103] Embodiment 4: The seed germination rate identification method described in this embodiment, wherein the method adopts the seed germination segmentation method described in any one of embodiments 1 to 3 to segment the seed germination and identify the seed germination rate.

[0104] In this embodiment, the seed germination rate identification method comprises the following steps:

[0105] Step S1, conducting seed germination experiments on small targets with four different aging gradients to obtain a seed germination dataset of the small targets, and after polygon annotation of the seed germination dataset of the small targets, dividing the seed germination dataset of the small targets into a training set and a validation set;

[0106] Step S2, inputting the training set obtained in step S1 into the T-YOLOv8-seg described in any one of implementation modes 1 to 3 to train it;

[0107] Step S3, input the verification set obtained in step S1 into T-YOLOv8-seg trained in step S2, perform seed detection and segmentation tasks, and count the seed germination rate of small targets.

[0108] Embodiment 5: The seed germination rate identification method described in this embodiment, wherein the method is based on the seed vitality analysis index and the seed germination segmentation method described in any one of embodiments 1 to 3 to realize the identification of the seed germination rate.

[0109] In this implementation manner, the seed vitality analysis index is specifically:

[0110] Based on the seed germination image segmented by VT-YOLOv8-seg, the seed germination mask image is obtained, and the actual area of ​​the embryo is obtained based on the seed germination mask image. The ratio of the actual area of ​​the embryo to the seed germination time is defined as the seed vitality analysis index.

[0111] In this embodiment, the seed germination rate identification method comprises the following steps:

[0112] Step S1, conducting seed germination experiments on small targets with four different aging gradients to obtain a seed germination dataset of the small targets, and after polygon annotation of the seed germination dataset of the small targets, dividing the seed germination dataset of the small targets into a training set and a validation set;

[0113] Step S2, inputting the training set obtained in step S1 into the T-YOLOv8-seg described in any one of implementation modes 1 to 3 to train it;

[0114] Step S3, input the validation set obtained in step S1 into the T-YOLOv8-seg trained in step S2, and perform the prediction task based on the optimal parameters after VT-YOLOv8-seg training, count the number of mask points and draw them on the graph while drawing the mask image. Calculate the total number of non-zero pixels in the mask, convert the tensor into a Python (programming language) value, and obtain the area of ​​the mask and the predicted result. The mask image marks the seeds and buds separately. The actual area of ​​the germ can be obtained by the number of pixels in the germ area, the actual area of ​​the culture dish, and the number of pixels in the culture dish. The ratio of the actual area of ​​the germ to the germination time is the seed germination rate or seed vitality analysis index.

[0115] Embodiment 6: The corn seed germination rate identification method described in this embodiment is implemented based on the seed germination rate identification method described in embodiment 4.

[0116] Embodiment 7: The method for identifying the germination rate of corn seeds described in this embodiment is implemented based on the method for identifying the germination rate of seeds described in embodiment 5.

[0117] In this embodiment, in order to better illustrate the effects of the corn seed germination rate identification method described in Embodiment 6 and the corn seed germination rate identification method described in Embodiment 7, the experimental results are uniformly compared and illustrated.

[0118] First, in order to better implement the technical solution described in this embodiment, Figure 4 As shown in the figure, the configurations of the experimental settings are CUDA version 11.3, Python version 3.8, PyTorch version 1.12, Python version 3.8, and GPU version RTX3090 (24GB).

[0119] The training configuration includes an input image size of 640×640, a batch size of 16, an initial learning rate of 0.001, 50 iterations, and the Adam (Adaptive Moment Estimation) optimizer.

[0120] VT-YOLOv8-seg was trained on the corn seed aging germination dataset, such as Figure 5 The figure shows the performance indicators in the training and validation process in detail. During the model training process, it can be seen that the validation / localization loss can smoothly converge from the initial value to below 0.4 and infinitely close to 0. The high-precision results of the model are produced without overfitting. The validation / mask loss also shows a significant downward and convergence trend. The reduction of validation / classification loss indicates that the reliability of the model for target classification is gradually enhanced. That is, all loss indicators show a trend of convergence to training during the model validation process, and the final stable value is slightly higher than the training result, indicating that the model has good generalization ability for data outside the training set. In addition, the positioning accuracy, positioning recall rate, mask accuracy and mask recall rate remain above 0.9, and the mean of the average positioning accuracy (50% threshold) and the mean of the average mask accuracy (50% threshold) are both above 0.9. Under more stringent conditions, the mean of the average positioning accuracy 50% to 95% threshold and the mean of the average mask accuracy 50% to 95% threshold are also above 0.7, further confirming that the model has strong segmentation performance and robustness under strict evaluation criteria.

[0121] Therefore, VT-YOLOv8-seg has a low and stable loss value, showing good performance on the corn seed aging and germination dataset, and the model performs well in segmenting the target at different loss thresholds, effectively learning the characteristics of corn seeds and germs without overfitting, and achieving good training results.

[0122] By contrasting the mask, e.g. Figure 6As shown in Figure 1, the existing YOLOv8-seg does not recognize the corn germ on the left. Figure 7 As shown, the VT-YOLOv8-seg described in this embodiment can accurately identify all corn germs, such as Figure 8 As shown in Figure 1, the existing YOLOv8-seg has a duplicate detection problem at the right corn germ, as Fig. 9 As shown, the VT-YOLOv8-seg described in this embodiment does not have the problem of repeated detection, such as Fig.10 As shown in Figure 2, the existing YOLOv8-seg has the problem of missed detection, such as Fig.11 As shown, the VT-YOLOv8-seg described in this embodiment does not have the problem of missed detection, such as Fig.12 As shown, the existing YOLOv8-seg has the problem of repeated detection of seeds and germs, such as Fig.13 As shown, the problem of repeated detection of seeds and embryos by VT-YOLOv8-seg described in this embodiment is greatly reduced.

[0123] The analysis results show that the existing YOLOv8-seg has insufficient convolution flexibility and limited perception of irregular edges. In addition, when corn germs just start to grow, the detection and segmentation effects are not good and there will be omissions. This problem is usually due to insufficient feature fusion, which leads to low small target detection performance. The VT-YOLOv8-seg described in this embodiment improves spatial pyramid pooling and processes multi-scale feature information more efficiently, thereby enhancing the network's perception of small targets. The dynamic snake convolution feature extraction module has a more flexible convolution kernel, has a stronger detection ability for the irregular edges of corn sprouts, improves the accuracy of edge segmentation, and reduces model parameters by degrading convolution layers, reducing redundant information in feature maps. Experiments show that the VT-YOLOv8-seg described in this embodiment can outperform the existing YOLOv8-seg in processing small targets and segmenting curved irregular edges in the detection and segmentation tasks of seeds and germs.

[0124] In addition, if Fig.14 As shown, in order to verify the efficiency and advantages of VT-YOLOv8-seg described in this embodiment, it is compared with Mask RCNN (Mask Region Convolutional Network), YOLOv5-seg (YOLO's v5 version - segmentation), YOLOv7-seg (YOLO's v7 version - segmentation) and YOLOv8-seg. The model is trained using the same dataset and training parameters and tested on the test set. The complete experiment evaluates the positioning accuracy, positioning recall, average precision mean (50% threshold), mask accuracy, mask recall, mask average precision mean (50% threshold) and parameter quantity of each segmentation network.

[0125] The results show the performance comparison of VT-YOLOv8-seg described in this embodiment with other models. In the detection task, the localization accuracy is 97.3%, the localization recall is 97.9%, and the average precision mean (50% threshold) is 98.5%. In the segmentation task, the mask accuracy is 97.2%, the mask recall is 97.7%, the average mask precision mean (50% threshold) is 98.2%, and the parameter count is 3.15%.

[0126] Compared with the existing YOLOv8-Seg, in the detection task, the VT-YOLOv8-seg described in this embodiment has an improvement of 5.1%, 8.8% and 6.4%, respectively. In the segmentation task, the improvement is 5.8%, 7.3% and 6.6%, respectively. Compared with YOLOv5-Seg, in the detection task, the VT-YOLOv8-seg described in this embodiment has an improvement of 7.8%, 10.3% and 8.5%, respectively. In the segmentation task, the improvement is 19.1%, 16.3% and 12.8%, respectively. Compared with YOLOv7-Seg, the VT-YOLOv8-seg described in this embodiment has an improvement of 9.6%, 12.9% and 7% in the detection task, and an improvement of 12.2%, 11% and 7.6% in the segmentation task. Compared with Mask RCNN, VT-YOLOv8-seg described in this embodiment improves 5.6%, 2.9% and 1% in the detection task, and improves 4.6%, 8.3% and 6.7% in the segmentation task.

[0127] Furthermore, the actual area of ​​the germ can be further deduced from the number of germ pixels output by VT-YOLOv8-seg, and the growth curve of the germ under different aging gradients of corn seeds can be drawn. Fig.15 As shown, the ordinate is the germ area, in mm 2 ,The horizontal axis is the germination time, in hours, where T0 represents the control group, T1 represents corn seeds with an aging time of 24 hours, T2 represents corn seeds with an aging time of 48 hours, and T3 represents corn seeds with an aging time of 72 hours. It can be seen that the greater the aging time gradient, the flatter the curve and the slower the germination speed. In the comparison of the growth curves of the four groups of seeds, T3 grows the slowest, but in terms of the first germination time, T2 germinates the latest, T0 grows the fastest, and the first germination is also the earliest.

[0128] In order to more intuitively see the growth of corn seed germ under different aging gradients, such as Fig.16 As shown, the embryo growth of corn seeds at 7 days old, with one image taken every 24 hours.

[0129] In summary, the VT-YOLOv8-seg method for segmenting aged corn seeds described in this embodiment detects the germination status of corn seeds with four different aging gradients, and can perform instance segmentation on seeds and embryos. Based on this, the vitality differences of corn seeds under different aging gradients are analyzed, and the germination vitality of aged corn seeds can be accurately evaluated and detected.

[0130] The seed germination segmentation method, seed germination rate and germination rate identification method proposed in the present invention are introduced in detail above. Specific examples are used in this article to illustrate the principle and implementation mode of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation mode and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A seed germination and division method, characterized in that: Specifically: The seed germination image is input into the Backbone, Neck and Head of VT-YOLOv8-seg in sequence for feature extraction and segmentation of the seed germination image; In the Backbone of VT-YOLOv8-seg, the seed germination image is input into the multi-level downconvolution layer for feature extraction, and then input into the dynamic snake convolution feature extraction module for dimensionality reduction, and the seed germination feature map after dimensionality reduction is output; The seed germination feature map after dimensionality reduction is input into the downgraded convolution layer for feature extraction, and then sequentially input into the feature extraction module, the multi-level convolution layer and the feature extraction module for feature extraction, and then input into the feature optimization pooling layer for pooling, and the pooled seed germination feature map is output; The feature optimization pooling layer is specifically: After the image passes through the convolution layer, it is input into the adaptive pooling for multi-scale feature extraction, and then into the multi-head attention mechanism to capture the global context information, and then the feature map is output. The feature map is sequentially input into the multi-head attention mechanism and spatial pyramid pooling, and the fused feature map is output; The feature map is sequentially input into the multi-head attention mechanism and the spatial pyramid pooling, and the fused feature map is output, specifically: The feature map is divided into two paths, one of which is input into the multi-level maximum pooling, and each feature map after maximum pooling is divided into two paths again. One feature map after multi-level maximum pooling is connected with the other feature map after maximum pooling and the other feature map, and then input into the convolution layer for fusion, and the fused feature map is output.

2. The seed germination and division method according to claim 1, characterized in that: The downgraded convolutional layer is specifically: The image is divided into two paths. One path is input into the multi-stage convolution layer to change its size. The image after the size change is input into the convolution layer to change its number of channels, and the image after the number of channels change is output. After the other image is added to the image with the changed number of channels, batch normalization and activation function are performed in sequence to output the feature map.

3. A seed germination rate identification method, characterized in that: The method is to segment seed germination using the seed germination segmentation method described in any one of claims 1-2 to identify the seed germination rate.

4. A seed germination rate identification method, characterized in that: The method is based on the seed vitality analysis index and the seed germination segmentation method described in any one of claims 1-2 to realize the identification of the seed germination rate.

5. The seed germination rate identification method according to claim 4, characterized in that: The seed vitality analysis index is specifically: Based on the seed germination image segmented by VT-YOLOv8-seg, the seed germination mask image is obtained, and the actual area of ​​the embryo is obtained based on the seed germination mask image. The ratio of the actual area of ​​the embryo to the seed germination time is defined as the seed vitality analysis index.

6. A method for identifying corn seed germination rate, characterized in that: The method is implemented based on the seed germination rate identification method described in claim 3.

7. A method for identifying the germination rate of corn seeds, characterized in that: The method is implemented based on the seed germination rate identification method described in any one of claims 4-5.

Images