Tomato fruit maturity detection method based on YOLOv11

Abstract

The invention discloses a tomato fruit maturity detection method based on YOLOv11. Belongs to the technical field of agricultural informatization and intelligence, and particularly relates to the technical field of tomato fruit maturity detection. The tomato maturity detection method solves the technical problem that a tomato maturity detection mode capable of comprehensively optimizing the detection precision, the recall rate and the detection efficiency of a YOLO model does not exist in an existing detection method. The method comprises the following steps: S1, obtaining tomato fruit picture sets with different maturity degrees, angles, shapes and sizes from the actual existence environment of tomato fruits; s2, marking the tomato fruit picture set, and determining the range of tomato fruits with different maturity degrees; s3, constructing a tomato fruit maturity degree detection model based on YOLOv11; s4, training and verifying the tomato fruit maturity detection model by adopting the marked tomato fruit picture set; and S5, performing maturity detection on the tomato fruit picture by adopting the tomato fruit maturity detection model which is verified to be qualified.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural informatization and intelligentization, and specifically relates to the technical field of tomato fruit maturity detection. Background Art

[0002] As a vegetable grown in most countries, tomatoes hold a significant economic importance. Their rich vitamins and various nutrients, combined with their affordability, have led to a significant increase in global tomato production over the past few decades. However, due to their inherent perishability and vulnerability, selecting tomatoes of appropriate maturity and quality is crucial throughout the tomato production process. Traditionally, tomatoes were grown in small, scattered plots and manually identified and harvested. However, as time has progressed, tomatoes are now increasingly cultivated in larger fields. This labor-intensive identification and harvesting process significantly consumes human resources, rendering traditional tomato harvesting methods unsuitable for modern harvesting. Globally, every year, a significant number of tomatoes are harvested at the wrong time due to incorrect maturity assessments, resulting in incalculable economic losses. The huge tomato yield and high harvest loss rate highlight the importance of finding a new, efficient, and accurate tomato maturity identification method to replace traditional manual tomato maturity identification. Under this premise, people are gradually discovering the importance of automated fruit and vegetable maturity detection technology. This new maturity identification method will meet the modern tomato production industry's demand for efficient and precise tomato production, and is also the direction for future breakthroughs in the tomato production industry.

[0003] To achieve automated detection of tomato ripeness, scientists have explored this research direction from various perspectives. With the rise of You Only Look Once (YOLO), more and more researchers are using YOLO for automated crop detection. The Chinese invention patent application, "Method for Detecting Ripeness of Tomatoes Based on an Improved YOLOv5 Network" (Publication No. CN116030456A), improves the YOLOv5 network structure, making the model more lightweight and reducing the number of parameters while maintaining the original detection accuracy. The Chinese invention patent application, "A Method for Detecting Ripeness of Tomatoes Based on an Improved YOLOv11" (Publication No. CN120125908A), designs the C3K2_WTConv module and proposes the YOLOv11-WTConv algorithm. Through its innovative wavelet transform convolutional structure, it enhances the multi-scale feature extraction capability of the target, improving both the mAP50 metric and detection accuracy.

[0004] The aforementioned YOLO model-based methods for detecting tomato ripeness all focus on improving a single performance feature of the existing detection model. There is no tomato ripeness detection method that can comprehensively optimize the YOLO model's detection accuracy, recall rate, and detection efficiency. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a tomato fruit maturity detection method based on YOLOv11, the method comprising the following steps: S1. Obtain a collection of tomato fruit images of different maturity, angles, shapes, and sizes from the environment in which the tomato fruits actually exist; S2. Label the tomato fruit image collection and determine the range of tomato fruits of different maturity levels; S3. Build a tomato fruit maturity detection model based on YOLOv11; S4. Use the labeled tomato fruit image set to train and validate the tomato fruit maturity detection model; S5. Use a verified tomato fruit maturity detection model to perform maturity detection on the tomato fruit image.

[0006] Furthermore, when annotating the tomato fruit image collection, tomato fruits of different maturity levels were divided into four categories, namely, ripe, unripe, semi-ripe, and rotten fruit.

[0007] Furthermore, the annotated tomato fruit image set was divided into training set, test set and validation set in a ratio of 6:2:2.

[0008] Furthermore, the tomato fruit maturity detection model is improved on the existing YOLOv11 model. The specific improvements are as follows: Introducing the improved scSE module into the YOLOv11 model; Introducing the ADown module into the YOLOv11 model; The Gold-YOLO network is used to replace the neck network part in the YOLOv11 model.

[0009] Furthermore, the improved scSE module specifically improves the scSE module by adding a 1×1 convolution module after the output of the scSE module.

[0010] Furthermore, the improved scSE module is introduced into the YOLOv11 model as follows: the second convolution module starting from the input of the backbone network of the YOLOv11 model is replaced by the improved scSE module.

[0011] Furthermore, the ADown module is introduced into the YOLOv11 model as follows: the fourth convolution module starting from the input of the YOLOv11 model backbone network is replaced by the ADown module.

[0012] Furthermore, when using the annotated tomato fruit image set to train the tomato fruit maturity detection model, feature extraction is performed through the backbone network, and then the multi-scale features output by the backbone network are fused through the neck network. The detection head module in the head network then generates candidate prediction boxes for abnormal types. The candidate prediction boxes are continuously optimized according to the loss function, and the training model obtains weights that are closer to the labels.

[0013] Furthermore, the loss function includes a positioning loss function and a classification loss function.

[0014] The beneficial effects of the method of the present invention are: The existing YOLOv11 model has low accuracy in identifying and classifying tomato fruit maturity, which will seriously affect tomato yield and quality, thereby affecting the local agricultural economy. To solve the above technical problems, the present invention improves the existing YOLOv11n. The proposed improved detection model SGA-YOLOv11, through experimental data, shows that SGA-YOLOv11 has high recognition accuracy and can accurately identify the maturity of different types of common tomato fruits. It can better detect the location of tomato fruits in different environments and the condition of tomato fruits and maturity, thereby solving the technical problem of low accuracy of the existing YOLOv11n in detecting and identifying the maturity of tomato fruits of different types and sizes.

[0015] The tomato fruit ripeness detection method described in this paper introduces an improved scSE module to address the high computational complexity and complexity of the original YOLOv11 model. This module adds a convolution to the original scSE module. This allows the scSE module to learn how to more effectively integrate information from channel-wise and spatial-wise attention, rather than simply adding the results of two attention branches. The convolution kernel weights can be understood as learning which attention branch provides more reliable and important information at each spatial location and channel.

[0016] The tomato fruit ripeness detection method described in this paper introduces an ADown module, which helps the model capture image features at a higher level while reducing computational complexity. The improved scSE module and ADown module form a complementary design, establishing a cross-level collaborative optimization mechanism. The improved scSE module significantly enhances detail perception through attention-guided feature calibration and refinement, while the ADown module compresses deep feature dimensions through structural reparameterization, reducing computational resource consumption while maintaining feature integrity. The synergistic effect of these two modules significantly improves recall by 6.1% while maintaining high precision (mAP50 increased by 2.0% and mAP50-95 increased by 1.2%). The overall parameter count is reduced by 3.9%, achieving a comprehensive optimization of detection precision, recall, and efficiency. This dual-module design achieves breakthroughs in feature quality, computational efficiency, and multi-scale capabilities. The synergistic effect is reflected in the collaboration of shallow and deep modules, resulting in a detection system with higher precision, stronger generalization, and a smaller footprint.

[0017] The tomato fruit maturity detection method described in this paper incorporates the gather-and-distribute mechanism proposed by Gold-YOLO, further improving the ability to fuse multi-scale features. By using a unified module to collect and fuse information at each scale, and then distributing the fused features to different layers, this method avoids the information loss inherent in traditional FPN architectures while enhancing the information fusion capabilities of the neck network without significantly increasing latency. Each of the three modules plays an irreplaceable role: the improved scSE module suppresses noise, ADown protects details, and Gold-YOLO acts as an intelligent dispatching center, dynamically combining the features processed by the first two to achieve an optimal state. Gold-YOLO forms a powerful synergistic effect with the first two improvements, jointly driving an overall improvement in model performance. Experimental results validate this synergistic effect.

[0018] This study provides a theoretical basis and technical support for the study of tomato fruit maturity. Building on this research approach and technical foundation, other topics can be extended, not only focusing on tomato fruit maturity, but also on various crop fruits using similar research methods, which has practical significance for the development of smart agriculture.

[0019] The dataset used in this paper was obtained from the actual environment in which tomato fruits exist. Compared with the tomato leaf disease dataset in the traditional laboratory environment, this dataset is much closer to our actual production and life, and is more beneficial for us to train a deep learning model to detect the maturity of tomato fruits. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 These are images of four types of tomato fruits under different environments: ripe, semi-ripe, unripe, and rotten; Figure 2 This is a structural diagram of a tomato fruit maturity detection model based on YOLOv11 in an embodiment of the present invention; Figure 3 Schematic diagram of the recognition results of the tomato fruit maturity detection model based on YOLOv11 for images of four types of tomato fruits: ripe, unripe, semi-ripe, and rotten; Figure 4 The performance evaluation chart of the tomato fruit maturity detection model based on YOLOv11 during the training and validation process. DETAILED DESCRIPTION

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides a method for detecting tomato fruit maturity based on YOLOv11, the method comprising the following steps: S1. Obtain a collection of tomato fruit images of different maturity, angles, shapes, and sizes from the environment in which the tomato fruits actually exist; S2. Label the tomato fruit image collection and determine the range of tomato fruits of different maturity levels; S3. Build a tomato fruit maturity detection model based on YOLOv11; S4. Use the labeled tomato fruit image set to train and validate the tomato fruit maturity detection model; S5. Use a verified tomato fruit maturity detection model to perform maturity detection on the tomato fruit image.

[0023] Example 2 This embodiment is a further limitation of embodiment 1. When annotating the tomato fruit image set, in order to enhance the generalization and robustness of the model and to increase the richness of this dataset, this embodiment uses a total of 1,200 real-life and online images. The original images are divided into four categories, namely, ripe, semi-ripe, immature, and rotten tomato fruits. The tomato fruit images are as follows: Figure 1As shown in the figure, the dataset was ultimately divided into training, test, and validation sets in a 6:2:2 ratio. The image pixels were uniformly set to 640×640, resulting in a training set of 720 images, a test set of 240 images, and a validation set of 240 images. The images were annotated using LabelImg to create the dataset, with ripe fruit labeled "Ripe," semi-ripe fruit labeled "Half-Ripe," unripe fruit labeled "Unripe," and bad fruit labeled "Bad."

[0024] Example 3 This embodiment further limits the embodiment 1. The tomato fruit maturity detection model is improved on the existing YOLOv11 model (hereinafter referred to as SGA-YOLOv11). The specific improvements are as follows: Introduce the improved scSE module (hereinafter referred to as scSEc) into the YOLOv11 model; Introducing the ADown module into the YOLOv11 model; The Gold-YOLO network is used to replace the neck network part in the YOLOv11 model.

[0025] The scSEc module improves the scSE module by adding a 1×1 convolution module after the output of the scSE module and replacing the second convolution module from the input of the YOLOv11 model backbone network with the scSEc module.

[0026] The specific steps of introducing the ADown module into the YOLOv11 model are as follows: the fourth convolution module starting from the input of the YOLOv11 model backbone network is replaced by the ADown module.

[0027] The schematic diagram of the improved SGA-YOLOv11 model is as follows Figure 2 shown.

[0028] The innovation of the scSEc module is demonstrated by the following: Experimental results before the improvement showed accuracy of 0.798, recall of 0.732, map50 of 0.781, map50-95 of 0.623, and a parameter count of 3.1M (experimental data is provided below). After inserting the 1×1 convolutional module, experimental results showed accuracy of 0.807, recall of 0.732, map50 of 0.779, map50 of 0.656, and a parameter count of 2.58M. These experimental results can be understood as adding a convolution layer after the addition, which is equivalent to introducing a nonlinear feature fusion and transformation layer. This convolutional layer performs further nonlinear transformation and adjustment on the fused attention-weighted features. It suppresses noise or inconsistent responses that may be introduced during the fusion process and emphasizes truly important features. This is equivalent to a more refined fine-tuning of the feature map.

[0029] The innovative nature of the scSEc+ADown approach lies in its complementary design. By embedding the scSEc module to enhance shallow feature representation capabilities and simultaneously introducing the ADown module for efficient downsampling, the two modules establish a cross-layer collaborative optimization mechanism. The scSEc module significantly improves detail perception through attention-guided feature calibration and refinement, while the ADown module compresses deep feature dimensions through structural reparameterization, reducing computational resources while maintaining feature integrity. The synergistic effect of these two approaches significantly improves recall by 6.1% while maintaining high precision (mAP50 increases by 2.0% and mAP50-95 by 1.2%). The overall parameter count is reduced by 3.9%, achieving a comprehensive optimization of detection precision, recall, and efficiency. This dual-module design achieves breakthroughs in feature quality, computational efficiency, and multi-scale capabilities. The synergistic effect is reflected in the collaboration between the shallow and deep modules, resulting in a detection system with higher accuracy, stronger generalization, and a smaller footprint.

[0030] The innovation of scSEc+ADown+Gold-YOLO is reflected in the following: The core of Gold-YOLO is the GD mechanism, which essentially dynamically fuses multi-scale features. However, its effectiveness is highly dependent on the quality of the input features, which is precisely where the value of the first two improvements lies. The first-layer scSEc module acts like a precision filter, filtering out noise from the original features while enhancing key features. This results in cleaner features passed to the backbone network, allowing subsequent convolutions to more efficiently extract semantic information.

[0031] The Adown module in the fifth layer, replacing standard downsampling, is a clever trick. Traditional downsampling tends to lose information about small objects, while the Adown module's structure preserves more detailed features. When these detailed features enter the Gold-YOLO neck, the GD mechanism better integrates details of different scales. For example, after the features of small objects are preserved by Adown in the shallow layers, GD can accurately combine them with semantic information in deeper layers.

[0032] Each of the three modules plays an irreplaceable role: the scSEc module suppresses noise, the Adown module protects details, and Gold-YOLO acts like an intelligent scheduling center, dynamically combining the features processed by the first two to the optimal state.

[0033] Therefore, the core purpose of incorporating Gold-YOLO is to build a significantly superior 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 significantly enhances the model's (particularly the Neck) ability to perceive multi-scale objects, integrate features, and achieve localization accuracy. Ultimately, while maintaining high accuracy, it significantly improves recall (finding more objects) and average precision under strict standards (better localization), achieving a comprehensive performance leap. Gold-YOLO forms a powerful synergistic effect with the first two improvements, jointly driving a comprehensive improvement in model performance. Experimental results perfectly validate the powerful synergy of these three improvements.

[0034] The following is a schematic diagram of the recognition results of the four types of tomato fruits: ripe, unripe, semi-ripe, and rotten fruit using the SGA-YOLOv11 model in this embodiment. Figure 3 As shown, label 0 indicates ripe fruit, label 1 indicates unripe fruit, label 2 indicates half-ripe fruit, and label 3 indicates bad fruit.

[0035] Example 4 This embodiment is a further limitation of Example 1. The labeled pictures in the training set are input into the SGA-YOLOv11 network for training. First, 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. The detection head module then generates a candidate prediction box for the abnormal type. The candidate prediction box is continuously optimized according to the loss function, and the training model obtains a weight that is closer to the label. The loss function of the present invention 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. After constraining the bounding box by center distance, aspect ratio, etc., the overlapping area, intersection-over-union ratio and other parameters between the prediction box and the real box are measured, and the corresponding matching degree is obtained.

[0036] The training parameters are input image size of 640×640, epoch of 250, learning rate of 0.01, IOU of 0.7, momentum of 0.937, weight decay of 0.0005, and batch size of 16.

[0037] Example 5 This embodiment further limits embodiment 1. The effectiveness of the SGA-YOLOv11 network in the present invention is verified through specific experiments.

[0038] In this experiment, the main indicators of the model used for detection are precision (Presion), recall rate (Recall), average precision (mAP) indicators mAP50 (%) and mAP50-95 (%) and parameter amount Params.

[0039] Comparative Experiments: To verify the detection performance of the SGA-YOLO network proposed in this paper, experiments were conducted on the same test set to compare it with four other mainstream target detection algorithms and YOLOv11 with the unimproved scSE module. The results are shown in Table 1. By analyzing Table 1, it can be seen that the SGA-YOLO network proposed in this paper achieved the best results in terms of Precision, mAP50(%), and mAP50-95(%), reaching 82.20%, 80.20%, and 67.40 respectively, which are significantly higher than other models. It is also close to the optimal recall rate without increasing the number of model parameters too much.

[0040] Table 1:

[0041] Ablation experiments show that the first improvement, adding the scSEc module, improved accuracy by 1.1%, recall by 1.8%, mAP50 by 0.7%, and mAP50-95 by 1.1% compared to the original model. Rounding up the number of parameters did not change. The second improvement, incorporating the downsampling module ADown, improved recall by 3.3%, mAP50 by 1.4%, and mAP50-95 by 1.3%. Only accuracy decreased by 1.4%, but the number of parameters was reduced by 0.10M. The third improvement, introducing the SGA-YOLO network, improved recall by 3.2%, mAP50 by 0.6%, and mAP50-95 by 0.4%, while reducing accuracy by 5.4% and increasing the number of parameters by 3.10M. It can be seen that all three improvements improved the basic performance of the model without significantly increasing the number of parameters. The second improvement did reduce the number of parameters in the base model to a certain extent. We then combined the improvements in pairs to observe their impact on the results. We found that all metrics improved compared to the original model after pairwise mixing. Finally, combining the three improvements resulted in a 3.4% improvement in accuracy, a 4.6% increase in recall, a 3% increase in mAP50, and a 2.9% increase in mAP50-95. This resulted in a 2.99M increase in parameters compared to the original model, significantly improving model performance without increasing model complexity. The results are shown in Table 2.

[0042] Table 2:

[0043] Furthermore, this embodiment also provides a performance evaluation diagram of SGA-YOLOv11 during training and verification, such as Figure 4As shown. "results" represents the label of the chart summarizing the training results, and "smooth" indicates whether smoothing is applied to the curves on the chart. Smoothing helps reduce noise and makes the overall trend easier to observe. Selecting "smooth" results in a flatter curve; unselecting "smooth" displays the raw, fluctuating data. "train / box_loss" represents the bounding box regression loss (BoundingBox Loss) on the training set. It measures the error between the model's predicted object bounding boxes (position: x, y, w, h) and the ground-truth labeled boxes. Lower values indicate a better model's ability to locate objects on the training set. This is a core metric for the model's learning to locate objects. "train / cls_loss" represents the classification loss on the training set. It measures the error between the model's predicted object categories and the ground-truth class labels. Lower values indicate a better model's ability to identify object categories on the training set. This is a core metric for the model's learning to identify object identities. "train / dfl_loss" represents the distribution focal loss (DFL) on the training set. This is a specific loss function introduced in YOLOv11. It is primarily used to improve the accuracy of bounding box regression, particularly the prediction of bounding box dimensions (width w and height h). It calculates the loss by modeling the predictions as a discrete probability distribution and optimizing its match to the target distribution. Lower values indicate a more accurate distribution of bounding box dimensions. Metrics / precision(B) indicates the proportion of all bounding boxes predicted by the model on the validation set that are truly correct. (B) explicitly states that this metric is specific to the bounding box detection task. Metrics / recall(B) indicates the recall rate on the validation set. For bounding box detection, it indicates the proportion of all real-world objects (positive examples) that the model correctly detects. (B) emphasizes that this is the recall rate for bounding box detection in the object detection task. Higher values indicate fewer false negatives. val / box_loss represents the bounding box regression loss (Bounding Box Loss) on the validation set. It is similar to train / box_loss, but is calculated on the validation set. This is a key indicator for evaluating the model's ability to locate objects on unseen data. It is usually slightly higher than the training loss and is an important basis for determining whether the model is overfitting.val / cls_loss represents the classification loss on the validation set. Similar to train / cls_loss, it is calculated on the validation set. It evaluates the model's ability to identify object categories on unseen data and is a key metric for determining the model's classification generalization ability. val / dfl_loss represents the distribution focal loss on the validation set. Similar to train / dfl_loss, it is calculated on the validation set. It evaluates the model's accuracy in predicting the bounding box size distribution on unseen data. Metrics / mAP50(B) represents the mean average precision (mAP) on the validation set. It is one of the most core single evaluation metrics for object detection models for bounding box detection. It is calculated across different categories. When the intersection over union (IoU) threshold between the predicted box and the ground-truth box is set to 0.5, the average precision (AP) of each category is calculated. The AP values for all categories are then averaged to obtain mAP. (B) emphasizes that this is a detection metric. Higher values (ranging from 0 to 1) indicate better overall detection performance (localization + classification) under the relatively loose IoU threshold of 0.5. This is the primary metric for judging a model's practicality. Metrics / mAP50-95(B) represents the mean average precision (mAP) on the validation set. It is one of the most comprehensive and rigorous core evaluation metrics for object detection models. It calculates the average precision (AP) for different categories at IoU thresholds ranging from 0.5 to 0.95 (with a step size of 0.05 and a total of 10 thresholds: 0.5, 0.55, ..., 0.95). These APs are then averaged across all categories to obtain mAP. (B) emphasizes that this is a metric for detection boxes. It comprehensively evaluates the performance of the model under different positioning strictness. Higher values (ranging from 0 to 1) indicate better overall positioning and classification accuracy.

Claims

1. A tomato fruit maturity detection method based on YOLOv11, characterized in that: The method comprises the following steps: S1. Obtain a collection of tomato fruit images of different maturity, angles, shapes, and sizes from the actual environment in which the tomato fruits exist; S2. Label the tomato fruit image collection and determine the range of tomato fruits of different maturity levels; S3. Build a tomato fruit maturity detection model based on YOLOv11; S4. Use the labeled tomato fruit image set to train and validate the tomato fruit maturity detection model; S5. Use a verified tomato fruit maturity detection model to perform maturity detection on the tomato fruit image.

2. The tomato fruit maturity detection method based on YOLOv11 according to claim 1, characterized in that: When annotating the tomato fruit image collection, tomato fruits of different maturity levels are divided into four categories: ripe, unripe, semi-ripe, and rotten fruit.

3. The tomato fruit maturity detection method based on YOLOv11 according to claim 2, characterized in that: The labeled tomato fruit image set is divided into training set, test set and validation set in a ratio of 6:2:

2.

4. The tomato fruit maturity detection method based on YOLOv11 according to claim 2, characterized in that: The tomato fruit maturity detection model is improved on the existing YOLOv11 model. The specific improvements are as follows: Introducing the improved scSE module into the YOLOv11 model; Introducing the ADown module into the YOLOv11 model; The Gold-YOLO network is used to replace the neck network part in the YOLOv11 model.

5. The tomato fruit maturity detection method based on YOLOv11 according to claim 4, characterized in that: The improved scSE module improves the scSE module by adding a 1×1 convolution module after the output of the scSE module.

6. The tomato fruit maturity detection method based on YOLOv11 according to claim 5, characterized in that: The improved scSE module is introduced into the YOLOv11 model as follows: the second convolution module starting from the input of the YOLOv11 model backbone network is replaced by the improved scSE module.

7. The tomato fruit maturity detection method based on YOLOv11 according to claim 6, characterized in that: The specific steps of introducing the ADown module into the YOLOv11 model are as follows: the fourth convolution module starting from the input of the YOLOv11 model backbone network is replaced by the ADown module.

8. The tomato fruit maturity detection method based on YOLOv11 according to claim 7, characterized in that: When using a set of labeled tomato fruit images to train a tomato fruit maturity detection model, the backbone network is used to extract features, and the neck network is used to fuse the multi-scale features output by the backbone network. The detection head module in the head network then generates candidate prediction boxes for abnormal types. The candidate prediction boxes are continuously optimized according to the loss function, and the training model obtains weights that are closer to the labels.

9. The tomato fruit maturity detection method based on YOLOv11 according to claim 8, characterized in that: The loss function includes a positioning loss function and a classification loss function.

Images