A method, system and equipment for identifying superior rice callus

By combining the improved YOLOv8s target detection model with the MobileViT-UNet segmentation network, accurate identification and quantitative determination of rice callus tissue were achieved, solving the problems of low detection accuracy and contamination in existing technologies and meeting the needs of large-scale breeding.

CN123118522APending Publication Date: 2026-11-17HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611313914.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for detecting rice callus tissue suffer from problems such as low detection accuracy, high false negative rate, inability to quantify the quality of the callus, and easy introduction of microbial contamination. They are particularly difficult to accurately separate and identify small targets and overlapping structures.

Method used

An improved YOLOv8s target detection model and a MobileViT-UNet segmentation network are adopted. By combining detection and segmentation branches, decision-level fusion of detection confidence, segmentation confidence, and morphological confidence of candidate target boxes is performed to achieve accurate identification and quantitative determination of rice callus tissue.

Benefits of technology

It effectively distinguishes callus tissues that are similar in appearance but have different qualities, filters out false detections and noisy targets, improves detection accuracy, meets the batch precision sorting needs of large-scale breeding, and avoids microbial contamination introduced by manual sorting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118522A_ABST
    Figure CN123118522A_ABST
Patent Text Reader

Abstract

This application discloses a method, system, and device for identifying superior rice callus tissue, relating to the field of intelligent detection technology in crop tissue culture. The method includes acquiring an image of the rice callus tissue to be identified; inputting the rice callus tissue image into a detection branch and a segmentation branch, respectively. The detection branch employs an improved YOLOv8s target detection model, which includes a backbone network, a neck network, and a detection head. The C2f module in the backbone network is connected to an attention module. A pixel-level segmentation mask is obtained through the segmentation branch, and segmentation confidence and morphological confidence are calculated. The detection confidence, segmentation confidence, and morphological confidence of each candidate target box are fused at the decision level to obtain the comprehensive confidence of each candidate target box. Superior rice callus tissue is determined, and its location information is output. This application can accurately and quickly screen and identify superior rice callus tissue from mixed and stacked rice callus tissue samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent detection technology in crop tissue culture, and in particular to a method, system and equipment for identifying superior callus tissue in rice. Background Technology

[0002] Rice is one of my country's major food crops. Genetic transformation breeding and somatic cell clonal variation breeding are important technical pathways for cultivating high-yielding, stress-resistant, and high-quality new rice varieties. The induction and screening of rice callus tissue is a fundamental preliminary step in these breeding processes. The quality of callus tissue directly affects the efficiency of genetic transformation and the success rate of breeding: superior callus tissue, characterized by firm texture, milky white or pale yellow color, uniform size, and absence of browning and water staining, exhibits significantly higher subsequent differentiation and transformation efficiency than inferior callus tissue. Therefore, standardized initial selection and quality testing of callus tissue are necessary in gene editing, genetic transformation, and large-scale seedling production.

[0003] Currently, the screening of rice callus tissue mainly relies on manual visual sorting. However, manual sorting has the following drawbacks: the sorting standards depend on the personal experience of the sorters, and it is difficult to unify the judgment standards among different people, resulting in large fluctuations in accuracy; the daily processing capacity of a single person is limited, which cannot meet the batch sorting needs of large-scale breeding; and the direct contact between personnel and callus tissue during the sorting process can easily introduce microbial contamination and reduce the survival rate of callus tissue.

[0004] To replace manual sorting, existing technologies have developed auxiliary detection methods based on traditional machine vision. These methods use algorithms such as color thresholding and morphological contour extraction to process callus images and filter for quality. However, these methods rely solely on single-dimensional features such as color and contour, making them poorly adaptable to real-world detection scenarios where callus tissues vary in size, are stacked and mixed, and are subject to varying lighting conditions. In particular, for stacked and overlapping callus samples, threshold-based segmentation methods struggle to accurately separate individual targets, resulting in high false negative and false positive rates. Furthermore, these methods can only perform rough classification based on preset rules and cannot quantitatively determine the quality of callus tissues.

[0005] With the development of deep learning technology, target detection models, represented by the YOLO series, have been gradually applied to the automated detection of crop targets. However, when applying existing target detection models directly to rice callus detection scenarios, the following problems still exist: callus targets are small in size and densely distributed, and existing models lack the ability to extract features from small targets, resulting in limited detection accuracy; for stacked and overlapping callus targets, the bounding box regression bias is large, leading to significant missed detections; in addition, the detection results usually only provide the target location and category, making it difficult to combine the morphological characteristics of the callus to make a comprehensive judgment on its quality, which cannot meet the actual needs of precise quality control of callus in breeding production. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, and device for identifying superior rice callus tissue, which can accurately screen and identify superior rice callus tissue from mixed stacked rice callus tissue samples.

[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for identifying superior rice callus, comprising: acquiring an image of the rice callus to be identified; inputting the rice callus image into a detection branch and a segmentation branch respectively; wherein, the detection branch adopts an improved YOLOv8s target detection model, the improved YOLOv8s target detection model includes a backbone network, a neck network, and a detection head, the C2f module in the backbone network is connected to an attention module, the neck network adopts a bidirectional feature pyramid network to perform bidirectional weighted fusion of the multi-scale features output by the backbone network with learnable weights, and the detection head outputs candidate targets based on the fused multi-scale features. The segmentation branch uses the MobileViT-UNet segmentation network to obtain a pixel-level segmentation mask for the rice callus image. Based on this mask, the segmentation confidence and morphological confidence of each candidate target box are calculated. The detection confidence, segmentation confidence, and morphological confidence of each candidate target box are then fused at the decision level to obtain a comprehensive confidence score. Candidate target boxes whose comprehensive confidence scores meet preset conditions are identified as superior rice callus, and their location information in the rice callus image is output.

[0008] In one embodiment, the backbone network includes multiple sequentially connected C2f modules, with each C2f module's output connected in series with an attention module, which is a SimAM parameterless attention module. The SimAM parameterless attention module independently calculates the energy value of each pixel within each channel of the feature map output by the C2f module, generating an attention weight map. This attention weight map is then multiplied pixel-by-pixel with the feature map output by the C2f module to obtain an attention-enhanced feature map. The formula for calculating the energy value is: In the formula, Let i be the energy value of the pixel in the i-th row and j-th column. Let be the pixel value of the pixel in the i-th row and j-th column of a single channel. This is the global mean of a single channel. For the global variance of a single channel, This is the smoothing constant.

[0009] In one embodiment, the bidirectional feature pyramid network is a BiFPN network, which includes a top-down branch and a bottom-up branch. The top-down branch upsamples large-scale features stepwise and then fuses them with features of adjacent scales. The bottom-up branch downsamples the small-scale features output by the top-down branch stepwise and then fuses them with features of adjacent scales a second time. The fusion of each branch uses a weighted fusion with learnable weights, and the formula for the weighted fusion is: In the formula, This is the feature map output after weighted fusion. For the i-th feature map participating in the fusion, Let be the learnable weights corresponding to the i-th feature map. To prevent constants with a denominator of zero.

[0010] In one embodiment, the backbone network of the improved YOLOv8s object detection model is further connected to an adaptive upsampling module, which is a CARAFE adaptive upsampling module. The CARAFE adaptive upsampling module predicts an upsampling convolution kernel that is adaptive to the low-resolution feature map based on the content of the input low-resolution feature map, and performs weighted recombination on the neighboring pixels of the low-resolution feature map based on the predicted upsampling convolution kernel to output a high-resolution feature map. The high-resolution feature map is used as a supplementary feature input to the bidirectional feature pyramid network to participate in the weighted fusion.

[0011] In one embodiment, obtaining the pixel-level segmentation mask of the rice callus image through the segmentation branch includes: concatenating the multi-scale features output by the neck network of the detection branch, and performing convolution processing on the concatenated features to unify the channel size, thereby obtaining a detection fusion feature map; inputting the rice callus image into the encoder of the MobileViT-UNet segmentation network to obtain an encoded segmentation feature map; performing convolutional weighted fusion of the detection fusion feature map and the encoded segmentation feature map to obtain a comprehensive fusion feature map; inputting the comprehensive fusion feature map into the decoder of the MobileViT-UNet segmentation network, upsampling it to the resolution of the rice callus image, and then outputting the pixel-level segmentation mask through an activation function.

[0012] In one embodiment, the segmentation confidence is calculated as follows: the average or maximum value of the pixel-by-pixel binary classification probability values ​​corresponding to the segmentation mask within each candidate target box region is taken to obtain the segmentation confidence of the candidate target box; the morphological confidence is calculated as follows: the morphological geometric parameters of the binary segmentation mask corresponding to each candidate target box are extracted, the morphological geometric parameters include at least the contour regularity parameter and the overlap adhesion parameter of the binary segmentation mask; the morphological geometric parameters are normalized to obtain the morphological confidence with a value range of 0 to 1.

[0013] In one embodiment, the calculation formula for the decision-level fusion is: In the formula, The overall confidence level is... The detection confidence level is... The segmentation confidence level. The morphological confidence level is defined as follows: the preset condition is that the overall confidence level is greater than or equal to a preset threshold.

[0014] In one embodiment, the improved YOLOv8s object detection model employs a composite bounding box loss function during the training phase. The composite bounding box loss function is as follows: In the formula, The loss value of the composite bounding box loss function. For the intersection-union ratio loss value, To normalize the Wasserstein distance loss value, The weighting coefficients range from 0 to 1; the normalized Wasserstein distance loss is used to provide an effective gradient when the intersection-union ratio of stacked overlapping targets approaches zero.

[0015] Secondly, this application provides a system for identifying superior rice callus tissue, comprising: an image acquisition module for acquiring images of rice callus tissue to be identified; a detection module employing an improved YOLOv8s target detection model, wherein the improved YOLOv8s target detection model includes a backbone network, a neck network, and a detection head, wherein the C2f module in the backbone network is connected to an attention module, the neck network employs a bidirectional feature pyramid network to perform bidirectional weighted fusion of the multi-scale features output by the backbone network with learnable weights, and the detection head outputs candidate target boxes and the detection confidence scores corresponding to each candidate target box based on the fused multi-scale features; segmentation. The system comprises the following modules: a module employing the MobileViT-UNet segmentation network to obtain a pixel-level segmentation mask for the rice callus image; a confidence calculation module to calculate the segmentation confidence and morphological confidence of each candidate bounding box based on the pixel-level segmentation mask; a decision fusion module to perform decision-level fusion of the detection confidence, segmentation confidence, and morphological confidence of each candidate bounding box to obtain a comprehensive confidence of each candidate bounding box; and a judgment output module to determine the target corresponding to the candidate bounding box whose comprehensive confidence meets a preset condition as a superior rice callus and output the position information of the target in the rice callus image.

[0016] Thirdly, this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for identifying superior rice callus tissue.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, system, and device for identifying superior rice callus. By fusing the detection confidence of candidate target boxes, the segmentation confidence obtained based on pixel-level segmentation masks, and the morphological confidence at the decision level, and using the comprehensive confidence as the criterion for judging superiority, this application comprehensively judges candidate targets from three dimensions: target detection results, pixel-level segmentation results, and morphological quantification results. It can effectively distinguish callus tissues that are similar in appearance but have different qualities, filter out falsely detected culture medium areas, browned callus, and noisy false targets, and solve the problem that the existing technology cannot quantify the quality of callus tissues.

[0018] Furthermore, the C2f module in the backbone network of the improved YOLOv8s target detection model is connected to an attention module, which can enhance the feature expression of callus tissue and suppress background interference such as culture medium, thereby improving the model's feature extraction capability in scenarios with small callus tissue targets and complex backgrounds. In addition, the neck network adopts a bidirectional feature pyramid network to perform bidirectional weighted fusion of multi-scale features with learnable weights, realizing the bidirectional transmission and fusion of deep semantic information and shallow detail information, strengthening the feature expression of small-sized callus tissue and stacked overlapping callus tissue, and solving the problems of low accuracy in small target recognition and high false negative rate of overlapping targets in existing target detection models.

[0019] Furthermore, this application automatically locates, identifies, and quantifies superior callus tissue in mixed stacked samples using a model. The judgment criteria are uniform and unaffected by the subjective experience of sorting personnel. The identification process does not require direct contact between personnel and callus tissue, avoiding the problems of microbial contamination and reduced callus survival rate introduced by manual sorting. Moreover, the identification results can be combined with sorting devices to achieve automated batch sorting, meeting the batch precision sorting requirements of large-scale breeding and solving the problems of inconsistent standards, low efficiency, and easy secondary contamination caused by manual sorting. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an application environment diagram of a method for identifying superior rice callus in one embodiment of this application; Figure 2 This is a flowchart illustrating a method for identifying superior rice callus in one embodiment of this application. Figure 3 This is a schematic diagram of the structure of the improved YOLOv8s target detection model in one embodiment of this application; Figure 4 This is a flowchart illustrating the feature-level and decision-level dual-fusion architecture in one embodiment of this application; Figure 5 This is a graph showing the change of the loss function during the training process of a traditional YOLOv8s model in one embodiment of this application. Figure 6 This is a graph showing the change of evaluation metrics during the training process of a traditional YOLOv8s model in one embodiment of this application. Figure 7This is a graph showing the change in the loss function during the training process of an improved YOLOv8s model in one embodiment of this application. Figure 8 This is a graph showing the change of evaluation metrics during the training process of an improved YOLOv8s model in one embodiment of this application. Figure 9 This is a schematic diagram of the functional modules of a rice callus identification system provided in another embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In one exemplary embodiment, the method for identifying superior rice callus provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. Data storage system 103 can store data that server 102 needs to process, such as historical image data of rice callus tissue, labeled sample data, and trained model parameter data. Data storage system 103 can be set up independently, integrated into server 102, or deployed in the cloud or on other servers.

[0024] Terminal 101 is set up at the rice callus culture or sorting site to acquire images of the rice callus to be identified, for example, by taking pictures of the callus in the culture dish using an image acquisition device. Terminal 101 can send the acquired rice callus images to server 102. After receiving the rice callus images, server 102 inputs them into the detection branch and the segmentation branch respectively. Based on the improved YOLOv8s target detection model, candidate target boxes and detection confidence scores are obtained. Based on the MobileViT-UNet segmentation network, pixel-level segmentation masks are obtained and segmentation confidence scores and morphological confidence scores are calculated. Then, the above three types of confidence scores are fused at the decision level to determine the superior rice callus. Server 102 can feed back the determination results and location information of the superior rice callus to terminal 101 for display, or for the sorting device (e.g., a robotic arm or robotic gripper) connected to terminal 101 to perform sorting operations on the superior rice callus based on the location information.

[0025] Furthermore, in some embodiments, the method for identifying superior rice callus tissue can also be implemented independently by the server 102 or the terminal 101. For example, the improved YOLOv8s target detection model is a lightweight model that can be deployed in the embedded processing unit of the terminal 101. The terminal 101 can directly perform recognition on the collected rice callus tissue images, which is suitable for field breeding sites without network access or with high real-time requirements. Alternatively, the server 102 can obtain the rice callus tissue image to be identified from the data storage system 103, perform the recognition, and store the recognition result in the data storage system 103.

[0026] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, industrial control equipment, embedded inspection equipment, and automated sorting equipment equipped with image acquisition devices. The server 102 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0027] In one exemplary embodiment, such as Figure 2 As shown, a method for identifying superior callus tissue in rice is provided. This method can be implemented by a terminal, a server, or through interaction between the terminal and the server; this application does not limit this approach. Figure 2 As shown, the method includes the following steps: S1, Obtain an image of the rice callus tissue to be identified.

[0028] S2, input the rice callus images into the detection branch and the segmentation branch respectively.

[0029] S3 obtains a pixel-level segmentation mask through segmentation branches and calculates segmentation confidence and morphological confidence.

[0030] S4, decision-level fusion yields comprehensive confidence level.

[0031] S5 identifies superior rice callus tissue and outputs its location information.

[0032] The following detailed explanation will illustrate the implementation process of each of the above steps.

[0033] Regarding step S1 above, in specific implementation, an image acquisition device set up at the rice callus culture or sorting site can be used to photograph the mixed and stacked rice callus in the culture dish to obtain an image of the rice callus to be identified. The image acquisition device can be an industrial camera or other equipment with image acquisition capabilities; this application does not limit the specific device used.

[0034] After obtaining images of rice callus tissue, the rice callus tissue images can be processed. Preprocessing: Images of rice callus tissue By scaling and pixel normalization, a standardized image is obtained. In this embodiment, the scaling specifically refers to: scaling the rice callus tissue image... The three-channel image is interpolated and scaled to 640×640 pixels; pixel normalization specifically involves mapping pixel values ​​from the 0~255 range to the 0~1 range.

[0035] Preprocessed standardized image Input the subsequent detection branch and segmentation branch respectively.

[0036] In specific implementation of step S2 above, the detection branch adopts an improved YOLOv8s target detection model. Figure 3 This is a schematic diagram of the structure of the improved YOLOv8s target detection model provided in the embodiments of this application, as shown below. Figure 3 As shown, the improved YOLOv8s object detection model includes a backbone network, a neck network, and a head. Specifically: The backbone network consists of a first-layer convolutional module and multiple sequentially connected C2f modules. Each C2f module's output is connected in series with an attention module. In this embodiment, the attention module can be a SimAM parameterless attention module, and the C2f module has four levels, from level 1 to level 4.

[0037] The first convolutional module processes the normalized image. Perform convolution and downsampling to extract basic texture and color features, and output basic feature maps. In this embodiment, the standardized image is... Perform 6×6 convolution with a stride of 2 for downsampling.

[0038] Each C2f module splits the input features into two residual convolution branches and then merges the branch outputs through channel concatenation. The cascaded attention module applies attention weights to the feature maps output by the C2f module. The attention module is used to enhance the feature expression of callus tissue and suppress background interference such as culture medium.

[0039] Taking Level 1 as an example: The Level 1 C2f module will generate the basic feature map The stream is split into two residual convolutional branches, and the outputs of the branches are concatenated and fused to obtain an intermediate feature map. The cascaded SimAM parameterless attention module focuses on intermediate feature maps. The energy value of each pixel is calculated independently in each channel to generate an attention weight map, and the attention weight map is then compared with the intermediate feature map. Pixel-wise multiplication yields the attention-enhanced feature map. Attention-enhanced feature maps Perform downsampling to obtain the downsampled feature map. Input to the second-level C2f module. In this embodiment, the attention-enhanced feature map is... Perform a downsampling operation with a step size of 2. The energy value is calculated using the following formula: ; In the above formula, intermediate feature map The energy value of the pixel in the i-th row and j-th column of any channel. Let be the pixel value of the pixel in the i-th row and j-th column of any channel. The global mean of any channel. Let be the global variance of any channel, and λ be the smoothing constant.

[0040] The data flow of the second-level C2f module is the same as that of the first level, and it outputs a downsampled feature map. This will not be elaborated upon here. The second-level C2f module also outputs small-scale feature maps to the neck network. .

[0041] Sampling feature maps below level 3 C2f module As input, after undergoing the same data flow as in level 1, the output is a downsampled feature map. The third-level C2f module also outputs mesoscale feature maps to the neck network. .

[0042] Sampling feature maps below level 4 C2f module As input, after undergoing the same data flow as in level 1, the output is a downsampled feature map. The fourth-level C2f module also outputs large-scale feature maps to the neck network. .

[0043] Optionally, the backbone network is also connected to an adaptive upsampling module. This module predicts upsampling convolutional kernels based on the content of the input low-resolution feature map, and performs weighted recombination on the neighboring pixels of the low-resolution feature map based on the predicted upsampling convolutional kernels to output a high-resolution feature map. In this embodiment, the adaptive upsampling module can be a CARAFE adaptive upsampling module, with each of the 3rd and 4th stages connected to a CARAFE upsampling branch: the CARAFE upsampling branch of the 3rd stage upsampling feature map... As input, output a high-resolution supplementary feature map. ; Subsampling feature map of CARAFE upsampling branch at level 4 As input, output a high-resolution supplementary feature map. The adaptive upsampling module can retain more detailed information during the upsampling process, improving the feature quality of small and overlapping targets.

[0044] The Neck network employs a bidirectional feature pyramid network to perform bidirectional weighted fusion of multi-scale features output by the backbone network with learnable weights.

[0045] The bidirectional feature pyramid network includes a top-down branch and a bottom-up branch: the top-down branch upsamples large-scale features step by step and then fuses them with features of adjacent scales; the bottom-up branch downsamples the small-scale features output by the top-down branch step by step and then fuses them with features of adjacent scales a second time. In this embodiment, the bidirectional feature pyramid network can be selected as a BiFPN network.

[0046] Specifically, top-down branches: for large-scale feature maps Upsampling to mesoscale feature maps after unifying channels For the same size, the upsampled large-scale and medium-scale feature maps and high-resolution supplementary feature maps Weighted fusion is performed to obtain an intermediate mesoscale feature map; the intermediate mesoscale feature map is then upsampled to match the small-scale feature map. For the same size, the upsampled mesoscale and small-scale feature maps and high-resolution supplementary feature maps Weighted fusion is performed to obtain intermediate small-scale feature maps. In this embodiment, CARAFE upsampling is used for upsampling the intermediate mesoscale feature map.

[0047] Bottom-up branch: for intermediate small-scale feature maps Perform convolutional downsampling to the mesoscale, then weight and fuse the downsampled small-scale features with the intermediate mesoscale feature maps obtained from the top-down branch to obtain the mesoscale enhanced feature map. ;Mesoscale enhanced feature maps Perform convolutional downsampling to a large scale, then combine the downsampled mid-scale features with the large-scale feature map. Weighted fusion is performed to obtain large-scale enhanced feature maps. Intermediate small-scale feature map This refers to small-scale enhanced feature maps. .

[0048] The weighted fusion of each branch is performed using the following formula: ; In the formula, This is the feature map output after weighted fusion. For the i-th input feature map participating in the fusion, The learnable weights are the input feature maps of the i-th path. To prevent constants with a denominator of zero.

[0049] Small-scale augmented feature maps output by the neck network Mesoscale Enhanced Feature Map and large-scale enhanced feature maps It is divided into two paths: the first path inputs the detection head; the second path inputs subsequent feature-level fusion.

[0050] Through bidirectional cross-scale connectivity and weighted fusion, the neck network can enhance the feature representation of small-sized callus and overlapping callus, thereby reducing the false negative rate.

[0051] The detection head outputs candidate bounding boxes and the detection confidence scores for each candidate bounding box based on the multi-scale features fused from the neck network. Specifically, small-scale enhanced feature maps Mesoscale Enhanced Feature Map and large-scale enhanced feature maps The same branching process is performed in parallel within the detection head: each feature map is simultaneously split into a classification branch and a regression branch.

[0052] The classification branch performs convolution on the feature map and outputs the class confidence score for each grid cell after activation, thus obtaining the detection confidence score for high-quality callus tissue. In this embodiment, the classification branch uses 3 layers of convolution to compress the channels into 2-dimensional category channels, and the activation function can be the Sigmoid function.

[0053] The regression branch performs convolution on the feature map, outputs the offset coordinates of each grid, and maps the offset coordinates to the rice callus image. The size of the target area is determined to generate candidate bounding boxes. In this embodiment, the regression branch uses a 3-layer convolution, and the mapped candidate bounding boxes are located on the 640×640 pixel rice callus image I.

[0054] During the inference phase, all candidate bounding boxes undergo redundancy filtering, and the output carries the detection confidence score. The set of candidate target boxes. In this embodiment, redundancy filtering uses non-maximum suppression (NMS): based on detection confidence. Sort the candidate bounding boxes, retain the candidate bounding boxes with the highest detection confidence, and remove the candidate bounding boxes whose intersection-union ratio with the retained candidate bounding boxes is greater than the preset intersection-union ratio threshold. Repeat the above process until all candidate bounding boxes have been processed.

[0055] During the training phase, the regression branch uses a composite bounding box loss function to optimize the bounding box localization accuracy and uses distributed focus loss (DFL) to finely correct small offsets in the bounding box, as detailed in the model training example below.

[0056] In specific implementation of step S3 above, the segmentation branch adopts the MobileViT-UNet segmentation network, which includes an encoder and a decoder. Figure 4 This is a flowchart illustrating the feature-level and decision-level dual-fusion architecture provided in an embodiment of this application. Figure 4 As shown, this embodiment adopts a dual-fusion architecture that combines feature-level fusion and decision-level fusion.

[0057] In the feature-level fusion stage: Small-scale augmented feature maps output by the neck network of the detection branch Mesoscale Enhanced Feature Map and large-scale enhanced feature maps Channel concatenation is performed, and the concatenated features are then convolved to unify the channel size, resulting in a detection fusion feature map. In this embodiment, the features after channel splicing are processed by 1×1 convolution.

[0058] Standardized image The encoder of the MobileViT-UNet segmentation network is input, and it extracts pixel-level detail features such as shallow edges and textures through convolution and Transformer downsampling, outputting an encoded segmentation feature map. .

[0059] Detection fusion feature map With encoding segmentation feature map Perform convolutional weighted fusion to obtain a comprehensive fused feature map. .

[0060] Integrate and fuse feature maps The input to the decoder of the MobileViT-UNet segmentation network is upsampled and restored to the image of rice callus tissue. After determining the resolution, a pixel-level segmentation mask M is output through an activation function. In this embodiment, the decoder uses CARAFE upsampling, and the activation function can be a sigmoid function.

[0061] The encoder of the MobileViT-UNet segmentation network continuously preserves shallow edges and texture details during downsampling, and focuses more on contour and pixel-level boundary information compared to the backbone network of the detection branch. Integrating such fine-grained coding features into the detection features can improve the feature recognition of small calluses and adhesion boundaries, and alleviate the problems of weakened features and blurred boundaries of small targets.

[0062] After obtaining the pixel-level segmentation mask M, the segmentation confidence score corresponding to each candidate target box is calculated based on the pixel-level segmentation mask M. and morphological confidence .in: Segmentation confidence Characterizes the proportion of foreground pixels within the candidate target bounding box and / or mask integrity. Optionally, segmentation confidence. The calculation method is as follows: statistically analyze the pixel-by-pixel binary classification probability values ​​corresponding to the pixel-level segmentation mask M within each candidate target box region to obtain the segmentation confidence of the candidate target box. In this embodiment, the mean or maximum value of the pixel-by-pixel binary classification probability values ​​is calculated.

[0063] Morphological confidence Characterizes the contour regularity and / or overlap / adhesion of the pixel-level segmentation mask M within the candidate target bounding box region. Optionally, morphological confidence. The calculation method is as follows: extract the morphological geometric parameters of the binary segmentation mask corresponding to each candidate target box, and normalize the morphological geometric parameters to obtain the morphological confidence score with a value range of 0 to 1. In this embodiment, the morphological geometric parameters include at least the contour regularity parameter and the overlap and adhesion parameter of the binary segmentation mask.

[0064] Regarding step S4 above, in specific implementation, the detection confidence of each candidate target box is... Segmentation confidence and morphological confidence Decision-level fusion is performed to obtain the overall confidence level of each candidate target box. In this embodiment, the calculation formula for decision-level fusion is: ; In the above formula, To assess the overall confidence level, To test the confidence level, To segment confidence levels, For morphological confidence.

[0065] By integrating decision-level fusion, the evidence from three dimensions—target detection results, pixel-level segmentation results, and morphological quantification results—is combined. Compared to methods that rely solely on detection confidence, this approach effectively filters out falsely detected culture medium areas, browned callus, and noisy false targets, thereby improving the accuracy of identifying superior callus tissue.

[0066] Regarding step S5 above, in specific implementation, the overall confidence level will be considered. Candidate target boxes that meet preset conditions are identified as high-quality rice callus tissue, and the target in the rice callus tissue image is output. Location information within. In this embodiment, the preset condition is the comprehensive confidence level. Greater than or equal to a preset threshold, the preset threshold can be selected as 0.6; if : Retain the bounding box and mask image corresponding to the target, and determine it as excellent rice callus tissue; if Remove the target, filtering out falsely detected culture medium, browned callus, and noisy spurious targets. Location information is the coordinates of the bounding box in the rice callus image I.

[0067] The output location information can be sent to a terminal for display or to a sorting device connected to the terminal. The sorting device then performs sorting operations on superior rice callus tissue based on the location information, thereby replacing manual selection and achieving standardization and unmanned operation of callus tissue detection and sorting. In this embodiment, the sorting device can be a robotic arm or a robotic gripper.

[0068] In one exemplary embodiment, the training process of the improved YOLOv8s object detection model is described: Training sample preparation: Rice callus images were collected and labeled. The labeled images were divided into training, validation, and test sets. Data augmentation was performed on the images in the training set to expand the sample size. In this embodiment, callus images of the main rice variety (e.g., Nipponbare) were collected. Labeling tools (e.g., Labelme) were used to mark the target bounding boxes and categories of superior callus. The labeling criteria for superior callus were: firm texture, milky white or light yellow color, and no browning or water staining. The labeled images were divided into training, validation, and test sets in a 7:2:1 ratio. Random rotation, flipping, brightness perturbation, and mosaic stitching enhancement were performed on the images in the training set to expand the sample size to 2000 images.

[0069] Training process: Images from the training set are preprocessed and then input into the improved YOLOv8s object detection model. The loss between the prediction results output by the detection head and the labeled results is calculated, and the weights of each layer of the network are updated through backpropagation until the model converges. The losses include composite bounding box loss, classification loss, and distributed focal loss (DFL): the composite bounding box loss is used to optimize the bounding box localization accuracy, the classification loss is used to distinguish between good callus tissue and the background, and the distributed focal loss is used to finely correct small offsets in the bounding box. All three are backpropagated simultaneously.

[0070] The composite bounding box loss function is: ; In the above formula, The loss value of the composite bounding box loss function. For the intersection-union ratio loss value, The normalized Wasserstein distance loss is the value of α, which is a weighting coefficient ranging from 0 to 1. The intersection-union ratio (IUU) loss is used to optimize the bounding box localization bias of regular targets; the normalized Wasserstein distance loss models the bounding box as a two-dimensional Gaussian distribution and measures the normalized Wasserstein distance between the distributions. It is used to provide effective gradients when the IUU of stacked, overlapping small callus targets approaches zero, thus solving the gradient vanishing problem.

[0071] like Figures 5 to 8 As shown, Figure 5 This is a graph showing the change in the loss function during the training process of a traditional YOLOv8s model. Figure 6 This is a graph showing the changes in evaluation metrics during the training process of a traditional YOLOv8s model. Figure 7 To improve the loss function change curve during YOLOv8s model training, Figure 8 To improve the evaluation metric change curve during the YOLOv8s model training process, the following measures are taken: Figure 5 and Figure 7 The horizontal axis represents the number of training rounds, and the vertical axis represents the loss value. Figure 6 and Figure 8 The horizontal axis represents the number of training epochs, and the vertical axis represents the metric value. Using the above training method, both the traditional YOLOv8s model and the improved YOLOv8s model of this application were trained for 100 epochs and tested on the same test set. The results are shown in the table below: As can be seen, compared with the traditional YOLOv8s model, the improved YOLOv8s model of this application has increased the mAP50 from 80.02% to 84.96%, the mAP50-95 from 62.74% to 65.18%, and the recall rate from 68.72% to 83.49%, indicating that the model improvement of this application effectively reduces the false negative rate of stacked, overlapping and small-sized callus tissues.

[0072] In one exemplary embodiment, Figure 9 This is a schematic diagram of the functional modules of the rice callus identification system provided in an embodiment of this application. Figure 9 As shown, the system includes: an image acquisition module 201, a detection module 202, a segmentation module 203, a confidence calculation module 204, a decision fusion module 205, and a decision output module 206.

[0073] Image acquisition module 201 is used to acquire image I of rice callus tissue to be identified; Detection module 202 employs an improved YOLOv8s object detection model. This improved model includes a backbone network, a neck network, and a detection head. The C2f module in the backbone network is connected to an attention module. The neck network uses a bidirectional feature pyramid network to perform bidirectional weighted fusion with learnable weights on the multi-scale features output by the backbone network. The detection head outputs candidate bounding boxes and the corresponding detection confidence scores for each candidate bounding box based on the fused multi-scale features. ; The segmentation module 203 uses the MobileViT-UNet segmentation network to obtain the pixel-level segmentation mask M of the rice callus image I; The confidence calculation module 204 is used to calculate the segmentation confidence of each candidate target box based on the pixel-level segmentation mask M. and morphological confidence ; Decision fusion module 205 is used to combine the detection confidence scores of each candidate target box. The segmentation confidence level and the morphological confidence Decision-level fusion is performed to obtain the overall confidence level of each candidate target box. ; The determination output module 206 is used to output the comprehensive confidence level. The target corresponding to the candidate target box that meets the preset conditions is determined to be a superior rice callus, and the position information of the target in the rice callus image I is output.

[0074] It should be noted that the specific implementation methods of each module mentioned above correspond one-to-one with the implementation methods of steps S1 to S5 in the aforementioned method embodiments. For details, please refer to the relevant descriptions of steps S1 to S5, which will not be repeated here. The above system can be deployed on... Figure 1 The terminal 101 in the application environment shown can also be deployed on the server 102, but this application embodiment does not limit this.

[0075] In one exemplary embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for identifying superior rice callus tissue in the foregoing method embodiment, namely steps S1 to S5.

[0076] The memory may include volatile memory, non-volatile memory, or a combination thereof, for storing computer programs and data such as rice callus images and model parameters; the processor may include one or more of a central processing unit (CPU), a graphics processing unit (GPU), and an embedded neural network processor (NPU). The memory and the processor may be connected via a bus or other means, which is not limited in this embodiment.

[0077] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the method for identifying superior rice callus tissue in the foregoing method embodiments, namely steps S1 to S5.

[0078] Computer-readable storage media may include read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, USB flash drives, portable hard drives, and other media capable of storing program code.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying superior rice callus, characterized in that, include: Obtain images of rice callus tissue to be identified; The rice callus images are input into a detection branch and a segmentation branch, respectively. The detection branch employs an improved YOLOv8s object detection model, which includes a backbone network, a neck network, and a detection head. The C2f module in the backbone network is connected to an attention module. The neck network uses a bidirectional feature pyramid network to perform bidirectional weighted fusion of the multi-scale features output by the backbone network with learnable weights. The detection head outputs candidate bounding boxes and the corresponding detection confidence scores for each candidate bounding box based on the fused multi-scale features. The segmentation branch uses the MobileViT-UNet segmentation network. The pixel-level segmentation mask of the rice callus image is obtained through the segmentation branch, and the segmentation confidence and morphological confidence of each candidate target box are calculated based on the pixel-level segmentation mask. The detection confidence, segmentation confidence, and morphological confidence of each candidate target box are fused at the decision level to obtain the comprehensive confidence of each candidate target box; The target corresponding to the candidate target box whose comprehensive confidence meets the preset conditions is determined as excellent rice callus, and the position information of the target in the rice callus image is output.

2. The method for identifying superior rice callus according to claim 1, characterized in that, The backbone network includes multiple C2f modules connected in sequence. The output of each C2f module is connected in series with an attention module, which is a SimAM parameterless attention module. The SimAM parameterless attention module independently calculates the energy value of each pixel in each channel of the feature map output by the C2f module to generate an attention weight map. The attention weight map is then multiplied pixel by pixel with the feature map output by the C2f module to obtain an attention-enhanced feature map. The formula for calculating the energy value is as follows: In the formula, Let be the energy value of the pixel in the i-th row and j-th column of any channel of the feature map output by any level C2f module. Let be the pixel value of the pixel in the i-th row and j-th column of any channel. The global mean of any one of the channels is given. Let be the global variance of any one of the channels. This is the smoothing constant.

3. The method for identifying superior rice callus according to claim 1, characterized in that, The bidirectional feature pyramid network is a BiFPN network, which includes top-down branches and bottom-up branches; The top-down branch upsamples the large-scale features step by step and then fuses them with the features of adjacent scales. The bottom-up branch downsamples the small-scale features output by the top-down branch step by step and then fuses them with the features of adjacent scales a second time. The fusion of each branch adopts a weighted fusion with learnable weights, and the formula for the weighted fusion is as follows: In the formula, This is the feature map output after weighted fusion. For the i-th feature map participating in the fusion, Let be the learnable weights corresponding to the i-th feature map. To prevent constants with a denominator of zero.

4. The method for identifying superior rice callus according to claim 1 or 3, characterized in that, The backbone network of the improved YOLOv8s object detection model is also connected to an adaptive upsampling module, which is a CARAFE adaptive upsampling module. The CARAFE adaptive upsampling module predicts an upsampling convolution kernel that is adaptive to the low-resolution feature map based on the content of the input low-resolution feature map, and performs weighted recombination on the neighboring pixels of the low-resolution feature map based on the predicted upsampling convolution kernel to output a high-resolution feature map. The high-resolution feature map is used as a supplementary feature input to the bidirectional feature pyramid network to participate in the weighted fusion.

5. The method for identifying superior rice callus according to claim 1, characterized in that, The MobileViT-UNet segmentation network includes an encoder and a decoder; obtaining the pixel-level segmentation mask of the rice callus image through the segmentation branch includes: The multi-scale features output by the neck network of the detection branch are concatenated by channels, and the concatenated features are convolved to unify the channel size, thus obtaining the detection fusion feature map. The rice callus image is input into the encoder of the MobileViT-UNet segmentation network to obtain the encoded segmentation feature map; The detection fusion feature map and the encoding segmentation feature map are convolutionally weighted and fused to obtain a comprehensive fusion feature map; The integrated feature map is input into the decoder of the MobileViT-UNet segmentation network. After being upsampled and restored to the resolution of the rice callus image, the pixel-level segmentation mask is output through an activation function.

6. The method for identifying superior rice callus according to claim 1 or 5, characterized in that, The segmentation confidence is calculated as follows: the average or maximum value of the pixel-by-pixel binary classification probability values ​​corresponding to the segmentation mask within each candidate target box region is taken to obtain the segmentation confidence of the candidate target box. The morphological confidence is calculated by extracting the morphological geometric parameters of the binary segmentation mask corresponding to each candidate target box. The morphological geometric parameters include at least the contour regularity parameter and the overlap and adhesion parameter of the binary segmentation mask. The morphological geometric parameters are normalized to obtain the morphological confidence level, which ranges from 0 to 1.

7. The method for identifying superior rice callus according to claim 1, characterized in that, The calculation formula for the decision-level fusion is as follows: In the formula, The overall confidence level is... The detection confidence level is... The segmentation confidence level. The confidence level of the morphology; The preset condition is that the overall confidence level is greater than or equal to a preset threshold.

8. The method for identifying superior rice callus according to claim 1, characterized in that, The improved YOLOv8s object detection model employs a composite bounding box loss function during the training phase. The composite bounding box loss function is as follows: In the formula, The loss value of the composite bounding box loss function. For the intersection-union ratio loss value, To normalize the Wasserstein distance loss value, The weighting coefficients range from 0 to 1; The normalized Wasserstein distance loss is used to provide an effective gradient when the intersection-union ratio of stacked overlapping targets approaches zero.

9. A system for identifying superior rice callus, characterized in that, include: The image acquisition module is used to acquire images of rice callus tissue to be identified. The detection module employs an improved YOLOv8s object detection model, which includes a backbone network, a neck network, and a detection head. The C2f module in the backbone network is connected to an attention module. The neck network uses a bidirectional feature pyramid network to perform bidirectional weighted fusion of the multi-scale features output by the backbone network with learnable weights. The detection head outputs candidate bounding boxes and the detection confidence scores corresponding to each candidate bounding box based on the fused multi-scale features. The segmentation module uses the MobileViT-UNet segmentation network to obtain a pixel-level segmentation mask for the rice callus image. The confidence calculation module is used to calculate the segmentation confidence and morphological confidence of each candidate target box based on the pixel-level segmentation mask. The decision fusion module is used to perform decision-level fusion of the detection confidence, segmentation confidence and morphological confidence of each candidate target box to obtain the comprehensive confidence of each candidate target box; The determination output module is used to determine the target corresponding to the candidate target box whose comprehensive confidence meets the preset conditions as excellent rice callus, and output the position information of the target in the rice callus image.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the method for identifying superior rice callus as described in any one of claims 1 to 8.