Tomato fruit maturity detection method based on YOLOv11

Abstract

The application discloses a tomato fruit maturity detection method based on YOLOv11, belongs to the technical field of agricultural informatization and intelligentization, and particularly relates to the technical field of tomato fruit maturity detection. The application solves the technical problem that the existing detection methods cannot optimize the tomato maturity detection in the aspects of detection precision-recall rate-detection efficiency from a YOLO model. The method comprises the following steps: S1, obtaining a tomato fruit picture set with different maturity, angle, shape and size from an actual environment of the tomato fruits; S2, labeling the tomato fruit picture set to determine the range of tomato fruits with different maturity; S3, constructing a tomato fruit maturity detection model based on YOLOv11; S4, training and verifying the tomato fruit maturity detection model by using the labeled tomato fruit picture set; and S5, detecting the maturity of the tomato fruit picture by using the verified tomato fruit maturity detection model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of agricultural informatization and intelligent technology, and particularly relates to the field of tomato fruit maturity detection technology. BACKGROUND

[0002] As a vegetable planted in most countries, tomato has an important economic position. Due to its rich vitamins and various nutrients and affordable price, the global tomato yield has increased substantially in the past few decades. Due to the perishable and easily damaged nature of tomatoes, it is particularly important to select tomatoes with appropriate maturity and good quality in the entire tomato production process. In the traditional tomato planting process, tomatoes are usually planted in small and scattered plots and are identified and picked by manual labor. With the passage of time, tomatoes are now more often planted in a large field, and the identification and picking workload is huge, which seriously consumes human resources, making the traditional tomato picking method not suitable for modern tomato picking process. Globally, a large number of tomatoes are picked at the wrong time every year due to incorrect maturity assessment, resulting in immeasurable economic losses. The huge yield of tomatoes and the high picking loss rate highlight the importance of seeking a new, efficient, and accurate tomato maturity identification method to replace the traditional manual tomato maturity identification. Under this premise, people gradually realize the importance of fruit and vegetable maturity automatic detection technology. This new maturity identification method will meet the needs of modern tomato production industry for tomato production efficiency and precision, and is also the direction for future breakthroughs in the tomato production industry.

[0003] In order to realize the automatic detection of tomato maturity, scientists have explored this research direction from various angles. With the rise of You Only Look Once (YOLO), more and more researchers use YOLO for automatic detection of crops. Chinese invention patent application “Detection method for maturity of string tomatoes based on improved YOLOv5 network” (publication number: CN116030456A) improves the YOLOv5 network structure, making the model more lightweight, reducing the number of parameters, and basically maintaining the original detection accuracy. Chinese invention patent application “Tomato maturity detection method based on improved yolo11” (publication number: CN120125908A) designs a C3K2_WTConv module and proposes a YOLOv11-WTConv algorithm. Through the innovative wavelet transform convolution structure, the multi-scale feature extraction capability of the target is enhanced, improving the mAP50 index and the detection accuracy.

[0004] The above tomato maturity detection method based on the YOLO model focuses on improving a single performance of an existing detection model, and there is no tomato maturity detection method that can optimize the detection accuracy-recall rate-detection efficiency of the YOLO model. SUMMARY

[0005] To solve the above technical problems, the present application provides a tomato fruit maturity detection method based on YOLOv11, which comprises the following steps:

[0006] S1, obtaining a set of tomato fruit pictures of different maturity, angle, shape and size from the actual environment of the tomato fruit;

[0007] S2, labeling the set of tomato fruit pictures to determine the range of tomato fruits of different maturity;

[0008] S3, constructing a tomato fruit maturity detection model based on YOLOv11;

[0009] S4, training and verifying the tomato fruit maturity detection model using the labeled set of tomato fruit pictures;

[0010] S5, using the verified tomato fruit maturity detection model to detect the maturity of the tomato fruit pictures.

[0011] Further, when labeling the set of tomato fruit pictures, the tomato fruits of different maturity are divided into four types, namely mature, unripe, semi-mature and rotten fruit.

[0012] Further, the labeled set of tomato fruit pictures is divided into a training set, a test set and a verification set according to a ratio of 6:2:2.

[0013] Further, the tomato fruit maturity detection model is improved based on the existing YOLOv11 model, and the specific improvements are:

[0014] The improved scSE module is introduced into the YOLOv11 model;

[0015] The ADown module is introduced into the YOLOv11 model;

[0016] The Gold-YOLO network is used to replace the neck network part of the YOLOv11 model.

[0017] Further, the improvement of the improved scSE module is that a 1x1 convolution module is added after the output of the scSE module.

[0018] Further, the improved scSE module is introduced into the YOLOv11 model, specifically: the second convolution module of the backbone network part of the YOLOv11 model is replaced by the improved scSE module.

[0019] Further, the ADown module is introduced into the YOLOv11 model, specifically: the fourth convolution module of the backbone network part of the YOLOv11 model is replaced by the ADown module.

[0020] Further, when the tomato fruit ripeness detection model is trained by using the labeled tomato fruit picture set, feature extraction is performed through the backbone network, then multi-scale features output by the backbone network are fused through the neck network, and then an abnormal type candidate prediction box is generated by a detection head module in the head network, the candidate prediction box is continuously optimized according to a loss function, and the model is trained to obtain weights closer to labels.

[0021] Further, the loss function comprises a positioning loss function and a classification loss function.

[0022] The method has the following beneficial effects:

[0023] The existing YOLOv11 model has low recognition and classification accuracy for tomato fruit ripeness, which seriously affects the yield and quality of tomatoes, thereby affecting the local agricultural economy, in order to solve the above technical problems, the existing YOLOv11n is improved, the improved detection model SGA-YOLOv11 is proposed, and experimental data show that the SGA-YOLOv11 has high recognition accuracy, can accurately recognize the ripeness of different types of common tomato fruits, and can accurately detect the position of the tomato fruit in different environments and the ripeness of the tomato fruit, thereby solving the technical problem of low detection and recognition accuracy of the existing YOLOv11n for different types and sizes of tomato fruit ripeness.

[0024] The tomato fruit ripeness detection method has the following beneficial effects: in order to solve the problem that the original YOLOv11 model may have a large amount of calculation and high model complexity after improvement, an improved scSE module is introduced, in the modification of the module, a convolution is added to the original scSE module, so that the original scSE module is no longer simply added to the results of the two attention branches, but can learn how to more effectively fuse the information from the channel attention and the spatial attention. The weight of the convolution kernel can be understood as learning which attention branch provides more reliable and important information on each spatial position and channel.

[0025] The tomato fruit maturity detection method provided by the present application introduces an ADown module, which helps the model to capture the characteristics of the image at a higher level, reduces the amount of calculation, and the improved scSE module+ADown module is a complementary design, which builds a cross-level collaborative optimization mechanism. The improved scSE module significantly improves the detail perception ability through attention-guided feature calibration and refinement, and the ADown module reduces the consumption of computing resources while maintaining the integrity of the features through structural reparameterization technology. Under the synergistic effect of the two, the model maintains high precision (mAP50 improves 2.0%, mAP50-95 improves 1.2%), while the recall rate is greatly improved by 6.1%, and the overall parameter amount is reduced by 3.9%, realizing the all-round optimization of detection accuracy-recall rate-efficiency. The double-module design realizes the triangular breakthrough of feature quality-computing efficiency-multiscale capability, and the synergistic effect is that through the cooperation of the shallow module and the deep module, a higher precision, stronger generalization and smaller volume detection system is created.

[0026] The tomato fruit maturity detection method provided by the present application fuses the gather-and-distribute mechanism proposed by Gold-YOLO, further improving the fusion ability of multiscale features. Through a unified module, information of each scale is collected and fused, and then the fused features are distributed to different layers, which not only avoids the inherent information loss of the traditional FPN structure, but also enhances the information fusion ability of the neck network part without significantly increasing the delay; the three modules each have irreplaceable role: the improved scSE module suppresses noise, ADown protects details, and Gold-YOLO dynamically combines the features processed by the former two to the best state, Gold-YOLO and the former two improvement points form a strong synergistic effect, which jointly promotes the overall improvement of the model performance, and the experimental results verify the synergistic effect of the three improvement points.

[0027] The present application provides a certain theoretical basis and technical support for the research on tomato fruit maturity. Based on the research ideas and technical basis, other topics can be extended, not only for tomato fruit maturity, but also for various crops. The similar ideas as the present research can be used for research, which has practical significance for the development of intelligent agriculture.

[0028] The data set used in the present application is obtained in the actual environment of tomato fruit, which is more close to our production and life than the traditional laboratory environment of tomato leaf disease, and is more beneficial for training the deep learning model for detecting the maturity of tomato fruit. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Fig. 1 is an image of four kinds of tomato fruits of different maturity, semi-maturity, immaturity and bad fruit in different environments in an embodiment of the present application;

[0030] Figure 2 Fig. 2 is a structure diagram of a tomato fruit maturity detection model based on YOLOv11 in an embodiment of the present application;

[0031] Figure 3 Fig. 3 is a recognition result diagram of the tomato fruit maturity detection model based on YOLOv11 in an embodiment of the present application on images of four kinds of tomato fruits of different maturity, immaturity, semi-maturity and bad fruit;

[0032] Figure 4 Fig. 4 is a performance evaluation diagram of the tomato fruit maturity detection model based on YOLOv11 in the training and verification process. DETAILED DESCRIPTION

[0033] The technical solutions of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0034] Embodiment 1,

[0035] The present embodiment provides a tomato fruit maturity detection method based on YOLOv11, which comprises the following steps:

[0036] S1, obtaining a set of tomato fruit pictures of different maturity, angle, shape and size from the actual environment of tomato fruits;

[0037] S2, labeling the set of tomato fruit pictures to determine the range of tomato fruits of different maturity;

[0038] S3, constructing a tomato fruit maturity detection model based on YOLOv11;

[0039] S4, training and verifying the tomato fruit maturity detection model using the labeled set of tomato fruit pictures;

[0040] S5, using the verified tomato fruit maturity detection model to detect the maturity of tomato fruit pictures.

[0041] Embodiment 2,

[0042] This embodiment is a further limitation of embodiment 1. When annotating the tomato fruit picture set, in order to strengthen the generalization and robustness of the model, and also to increase the richness of the data set. This embodiment uses a total of 1200 pictures in real scenes and network pictures. The original pictures are divided into four categories, namely mature, half-ripe, unripe and bad fruit, and the tomato fruit pictures are as shown in Figure 1 The data set is finally divided into training set, test set and validation set in the ratio of 6:2:2. The image pixels are uniformly set to 640x640, and finally 720 training sets, 240 test sets and 240 validation sets are obtained. The pictures are annotated by LabelImg to produce the data set, wherein the mature fruit is marked as (Ripe), the half-ripe fruit is marked as (Half-Ripe), the unripe fruit is marked as (Unripe), and the bad fruit is marked as (Bad).

[0043] Embodiment 3,

[0044] This embodiment is a further limitation of embodiment 1. The tomato fruit maturity detection model is improved on the basis of the existing YOLOv11 model (hereinafter referred to as SGA-YOLOv11), and the specific improvements are as follows:

[0045] The improved scSE module (hereinafter referred to as scSEc) is introduced into the YOLOv11 model;

[0046] The ADown module is introduced into the YOLOv11 model;

[0047] The Gold-YOLO network is used to replace the neck network part of the YOLOv11 model.

[0048] The improvement of the scSEc module to the scSE module is that a 1x1 convolution module is added after the output of the scSE module, and the second convolution module of the backbone network part of the YOLOv11 model from the input is replaced by the scSEc module.

[0049] The introduction of the ADown module into the YOLOv11 model is that the fourth convolution module of the backbone network part of the YOLOv11 model from the input is replaced by the ADown module.

[0050] The schematic diagram of the improved SGA-YOLOv11 model is as shown in Figure 2 .

[0051] The innovation of the scSEc module is embodied in that: the experimental results before improvement are accuracy 0.798, recall rate 0.732, map50 0.781, map50-95 0.623, and parameter quantity 3.1M (the experimental data are provided below), and the experimental results after inserting a 1*1 convolution module are accuracy 0.807, recall rate 0.732, map50 0.779, map50-95 0.656, and parameter quantity 2.58M. For such experimental results, it can be understood that adding a convolution layer after addition is equivalent to introducing a nonlinear feature fusion and conversion layer. The convolution layer can further nonlinearly transform and adjust the fused attention weighted features. It can suppress the noise or inconsistent response that may be introduced in the fusion process and emphasize the truly important features. It is equivalent to fine-tuning the feature map once again.

[0052] The innovation of scSEc+ADown is embodied in that: it is a functionally complementary design, which enhances the shallow feature representation ability by embedding the scSEc module, and realizes efficient downsampling by introducing the ADown module, and constructs a cross-level collaborative optimization mechanism. The scSEc module significantly improves the detail perception ability through attention-guided feature calibration and refinement; the ADown module compresses the deep feature dimension through structure reparameterization technology, reduces the calculation resource consumption while maintaining the feature integrity. Under the synergistic action of the two, the model maintains high precision (mAP50 improves by 2.0%, mAP50-95 improves by 1.2%) while the recall rate is greatly improved by 6.1%, and the overall parameter quantity is reduced by 3.9%, realizing the all-round optimization of detection accuracy-recall rate-efficiency. The double-module design realizes the triangular breakthrough of feature quality-computing efficiency-multiscale ability, and its synergistic effect is embodied in that: through the cooperation of the shallow module and the deep module, a higher precision, stronger generalization, and smaller volume detection system is jointly created.

[0053] The innovation of scSEc+ADown+Gold-YOLO is embodied in that: the core of Gold-YOLO is the GD mechanism, and the essence is dynamic fusion of multiscale features. But its effect is actually highly dependent on the quality of the input features, which is the value of the first two improvement points. The first layer scSEc module is like a precision filter, which filters out the noise in the most original features while enhancing the key features. In this way, the features passed to the backbone network are cleaner, and the subsequent convolution can more efficiently extract semantic information.

[0054] The Adown module of the 5th layer replaces the standard downsampling, which is prone to lose small target information, while the structure of the Adown module can retain more detailed features. When these detailed features enter the Neck of the Gold-YOLO, the GD mechanism can better fuse the detailed features of different scales. For example, the features of small targets are retained by the ADown in the shallow layer, and the GD can accurately combine them with the semantic information in the deep layer.

[0055] The three modules each have irreplaceable functions: the scSEc module suppresses noise, the Adown module protects details, and the Gold-YOLO acts as an intelligent dispatch center to dynamically combine the features processed by the first two modules to the best state.

[0056] Therefore, the core role of the Gold-YOLO is to construct a significantly better feature pyramid based on the high-quality features provided by the first two improvements through its dynamic adaptive fusion mechanism (GD) and training optimization mechanism (MAF). This pyramid greatly improves the model's (especially the Neck part) perception ability, feature integration ability, and positioning accuracy for multi-scale targets, ultimately significantly improving the recall rate (finding more targets) and the average precision under strict standards (positioning more accurately) while maintaining high precision, achieving a comprehensive performance leap. The Gold-YOLO forms a strong synergistic effect with the first two improvements, jointly promoting the comprehensive improvement of the model's performance. The experimental results perfectly verify the strong synergistic effect of the three improvements.

[0057] The recognition result of the image of the mature, immature, semi-mature, and bad fruit of the tomato fruit of the SGA-YOLOv11 model in the embodiment is shown in Figure 3 , wherein label 0 represents mature, label 1 represents immature, label 2 represents semi-mature, and label 3 represents bad fruit.

[0058] Embodiment 4,

[0059] This embodiment is a further limitation of embodiment 1. The labeled pictures in the training set are input into the SGA-YOLOv11 network for training. First, the feature extraction is performed through the backbone network. Then, the multi-scale features output by the backbone network are fused through the neck network to enhance the model's detection ability for multi-scale targets. Then, the detection head module generates abnormal type candidate prediction boxes. According to the loss function, the candidate prediction boxes are continuously optimized, and the model is trained to obtain weights closer to the labels. The loss function of the present application is consistent with the loss function of the basic Yolov11n network, mainly including positioning loss and classification loss. The positioning loss part mainly adopts the CIoU strategy, which measures the overlapping area, intersection ratio, and other parameters between the predicted box and the real box after constraining the center distance, width-height ratio, etc. of the boundary box, and obtains the corresponding matching degree.

[0060] The parameters of training are that the input image size is 640*640, the epoch is 250, the learning rate is 0.01, the iou is set to 0.7, the momentum is set to 0.937, the weight decay is 0.0005, and the batch size is 16.

[0061] Embodiment 5,

[0062] This embodiment is a further limitation of embodiment 1. The effectiveness of the SGA-YOLOv11 network in the application is verified through specific experiments.

[0063] In this experiment, the main indicators of the model used for detection are precision (Presion), recall (Recall), average precision mean (mAP), and the indicators mAP50(%) and mAP50-95(%).

[0064] Comparative experiment: In order to verify the detection performance of the SGA-YOLO network proposed in the application, experiments are carried out under the same test set and compared with other four mainstream target detection algorithms and YOLOv11 with unimproved scSE module. The results are shown in Table 1. As can be seen from Table 1, the SGA-YOLO network proposed in the application achieves the best results in the indicators Precision, mAP50(%) and mAP50-95(%), reaching 82.20%, 80.20% and 67.40 respectively, which are significantly higher than those of other models, and is close to the optimal in the indicator Recall, without excessively increasing the parameter quantity of the model.

[0065] Table 1:

[0066]

[0067] Ablation experiment: After the first improvement point adds the scSEc module, the accuracy is improved by 1.1% compared with the original model, the recall rate is improved by 1.8%, the mAP50 is improved by 0.7%, the mAP50-95 is improved by 1.1%, and the parameter quantity does not change after rounding; After the second improvement point fuses the down-sampling module ADown, the recall rate is improved by 3.3%, the mAP50 is improved by 1.4%, the mAP50-95 is improved by 1.3%, only the accuracy is reduced by 1.4%, but the parameter quantity is reduced by 0.10M; The third improvement point introduces the SGA-YOLO network, the recall rate is improved by 3.2% compared with the original model, the mAP50 is improved by 0.6%, the mAP50-95 is improved by 0.4%, the accuracy is reduced by 5.4% and the parameter quantity is increased by 3.10M. It can be seen that the three improvement points all improve the basic performance of the model without excessive increase in the parameter quantity of the model. The second improvement point reduces the parameter quantity of the basic model to a certain extent. Then we combine the improvement points two by two to observe their influence on the results. It can be seen that after mixing two by two, each index is improved compared with the original model. Finally, combine the three together, the accuracy is improved by 3.4% compared with the original model, the recall rate is improved by 4.6%, the mAP50 and the mAP50-95 are improved by 3% and 2.9% respectively, and the parameter quantity is increased by 2.99M compared with the original model. On the premise of not increasing the complexity of the model, the performance of the model is improved. The results are shown in Table 2.

[0068] Table 2:

[0069]

[0070] Further, the present embodiment also provides the performance evaluation diagram of SGA-YOLOv11 in the training and verification process, as shown in Figure 4The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary. The chart shows the training results summary.val / cls_loss represents the classification loss on the validation set, similar to train / cls_loss, but the calculation object is the validation set, which evaluates the model's ability to identify object categories on unseen data, and is a key indicator of the model's classification generalization ability. val / dfl_loss represents the distribution focal loss on the validation set, similar to train / dfl_loss, but the calculation object is the validation set, which evaluates the model's accuracy in predicting the size distribution of the bounding box on unseen data. metrics / mAP50(B) represents the mAP (mean average precision) on the validation set, for bounding box detection, which is one of the most core single evaluation indicators for object detection models. It calculates the average precision (AP) for each class when the intersection over union (IoU) threshold of the predicted box and the true box is set to 0.5, and then takes the average of all class APs to get mAP. (B) emphasizes that this is a detection box indicator, and the higher the value (0~1 range), the better the overall detection performance (positioning + classification) of the model under the relatively loose standard of IoU=0.5, which is the primary indicator for judging the practicality of the model. metrics / mAP50-95(B) represents the mAP (mean average precision) on the validation set, which is one of the most comprehensive and strict core evaluation indicators for object detection models. It calculates the average precision (AP) for each class when the IoU threshold is from 0.5 to 0.95 (step size 0.05, a total of 10 thresholds: 0.5, 0.55,..., 0.95), and then takes the average of all class APs to get mAP. (B) emphasizes that this is a detection box indicator, which comprehensively evaluates the performance of the model under different positioning strictness, and the higher the value (0~1 range), the better the positioning accuracy and classification accuracy of the model as a whole.

Claims

1. A tomato fruit ripeness detection method based on YOLOv11, characterized in that, The method comprises the following steps: S1, obtaining a picture set of tomato fruits of different maturity, angle, shape and size from an actual environment in which the tomato fruits exist; S2, labeling the picture set of tomato fruits to determine the range of tomato fruits of different maturity; S3, constructing a tomato fruit maturity detection model based on YOLOv11; the tomato fruit maturity detection model is improved based on an existing YOLOv11 model, and the specific improvements are as follows: introducing an improved scSE module into the YOLOv11 model; introducing an ADown module into the YOLOv11 model; replacing the neck network part in the YOLOv11 model with a neck network of a Gold-YOLO network; the improvement of the improved scSE module is that a 1x1 convolution module is added after the output of the scSE module; introducing the improved scSE module into the YOLOv11 model is that the second convolution module from the input of the backbone network part of the YOLOv11 model is replaced with the improved scSE module; introducing the ADown module into the YOLOv11 model is that the fourth convolution module from the input of the backbone network part of the YOLOv11 model is replaced with the ADown module; S4, training and verifying the tomato fruit maturity detection model by using the labeled picture set of tomato fruits; S5, detecting the maturity of the tomato fruit pictures by using the verified tomato fruit maturity detection model.

2. The YOLOv11-based method for detecting maturity of tomato fruits according to claim 1, characterized in that, When labeling the picture set of tomato fruits, tomato fruits of different maturity are divided into four types, namely mature, unripe, semi-mature and rotten fruit.

3. The YOLOv11-based method for detecting maturity of tomato fruits according to claim 2, characterized in that, The labeled picture set of tomato fruits is divided into a training set, a test set and a verification set according to a ratio of 6:2:

2.

4. The YOLOv11-based method for detecting maturity of tomato fruits according to claim 3, characterized in that, When training the tomato fruit maturity detection model by using the labeled picture set of tomato fruits, feature extraction is performed through the backbone network, then the multi-scale features output by the backbone network are fused through the neck network, then the detection head module in the head network generates abnormal type candidate prediction boxes, the candidate prediction boxes are continuously optimized according to the loss function, and the model is trained to obtain weights closer to the labels.

5. The YOLOv11-based method for detecting maturity of tomato fruits according to claim 4, characterized in that, The loss function comprises a positioning loss function and a classification loss function.

Images