Anchor frame tobacco chain small object detection method based on optimized YOLOv10

Abstract

The invention belongs to the technical field of tobacco chain small object detection, and particularly discloses an anchor frame tobacco chain small object detection method based on optimized YOLOv10, and the method comprises the following steps: S1, constructing a data set, and customizing a user-defined data set; s2, optimizing the anchor frame; step S3, integrating to a YOLOv10 model system structure; s4, a training program; and S5, evaluating the performance of the optimization model. According to the anchor frame tobacco chain small object detection method based on the optimized YOLOv10, small target detection in the tobacco chain is detected by refining the anchor frame in the YOLOv10 model, and small impurities can be better adapted, so that the detection performance is remarkably improved, the influence of optimization and improvement on the overall performance is minimum, and the detection accuracy is improved. Therefore, the detection method is feasible in real-time industrial application in which the precision is preferentially considered.

Description

Technical Field

[0001] The present invention belongs to the technical field of tobacco chain small object detection, and specifically relates to an anchor frame tobacco chain small object detection method based on optimized YOLOv10. Background Art

[0002] Detecting foreign matter in the tobacco chain is an important quality control task in the tobacco industry, which has a significant impact on the safety and quality of the final product. Foreign materials, such as non-metallic impurities, can compromise consumer safety and cause operational disruptions to online products. In addition, regulatory standards have strict purity level requirements and any deviations may lead to product recalls, legal consequences and reputational damage, which in turn highlights the importance of effective impurity detection. Therefore, maintaining a product free of impurities is not only necessary to meet legal requirements, but is also crucial to maintain consumer trust and ensure the efficiency of the production process.

[0003] Non-metallic impurities, including paper, wood, and plastic fragments, pose unique challenges due to their regular shapes and small sizes. These impurities are often integrated into the complex texture of the tobacco chain, making them difficult to detect using traditional machine vision systems. In addition, traditional vision systems rely heavily on predefined feature extraction algorithms, which have difficulty balancing high detection accuracy and the computational speed required in dynamic industrial environments. In real-time manufacturing environments, although the highest efficiency, the inability to achieve both accuracy and speed to meet the effectiveness of these systems limits their application in industrial environments.

[0004] Recent advances in deep learning-based object detection have led to some promising solutions. Unlike traditional systems, deep learning models, especially convolutional neural networks (CNNs), can automatically learn relevant features from data, allowing them to adapt to a wide range of object types. One such model, YOLO, has gained attention due to its ability to perform fast, real-time object detection without sacrificing accuracy.

[0005] YOLO's architecture is specifically designed to predict multiple bounding boxes and class probabilities that pass through the network, making it well suited for industrial applications that require real-time detection capabilities. But its effectiveness decreases when detecting small objects, especially for detecting some complex biological chains. This limitation occurs because the YOLO model uses predef inline anchor boxes, which provide templates for object detection. These anchor boxes are usually designed based on general datasets such as COCO, which include a wide variety of object sizes, but are not optimized for specialized industrial applications where small or irregularly shaped objects are common detection targets. Therefore, the 10 predefined anchor points cannot be consistent with the unique characteristics of toxic impurities found in tobacco chains, resulting in reduced detection accuracy.

[0006] Therefore, an innovative method needs to be developed in the art to optimize the anchor frame technology in the YOLOv10 model, which is specifically used to detect small-sized non-metallic design objects in tobacco chains, and can effectively solve the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide an anchor frame tobacco chain small object detection method based on optimized YOLOv10. By refining the anchor frame in the YOLOv10 model to detect small targets in the tobacco chain, it can better adapt to small impurities, thereby significantly improving the detection performance, and the optimization and improvement have minimal impact on the overall performance, making the detection method feasible in real-time industrial applications where accuracy is prioritized.

[0008] To achieve the above object, the present invention provides an anchor frame tobacco chain small object detection method based on optimized YOLOv10, comprising the following steps:

[0009] Step S1, data set construction, customizing a custom data set;

[0010] The custom dataset consists of high-resolution images of cut tobacco chains, each containing various non-metallic foreign objects, with each object having a bounding box and corresponding class label;

[0011] Step S2, anchor frame optimization;

[0012] Step S21: Given a set of real bounding boxes (w i ,h i ), where w i and h i are the width and height of different groups of anchor boxes respectively;

[0013] Step S22: normalize the width W and height H by the corresponding image dimension; the objective function of the K-means clustering algorithm is expressed as:

[0014]

[0015] In the formula, μ k , μ j are all center points, corresponding to the optimized anchor box size; and are the width and height of the image corresponding to the center point of the j group of anchor boxes; N is the total number of bounding boxes; w i and h i are the width and height of different groups of anchor boxes respectively;

[0016] Step S3, integrating into the YOLOv10 model architecture;

[0017] Utilize multiple multi-technique detector heads across different scales to capture objects of different sizes, each working on feature maps at different resolutions;

[0018] Step S4, training procedure;

[0019] Step S41, the Adam optimizer is used, the learning rate is 0.001, and the optimization problem of the training model is expressed as follows:

[0020]

[0021] In the formula, θ * is the optimal value of parameter θ; is the loss function; f(x i ; θ) is the prediction function of the model; y i is the actual observed value; N is the total number of data points;

[0022] Step S42, calculate the mean average precision (AP) of each class, and average mAP over all classes and IoU thresholds; where mAP is the mean average precision, providing a comprehensive method for model accuracy in different object scales and detection scenarios;

[0023] The average precision mAP is used to evaluate the performance of the model. The calculation method is as follows:

[0024]

[0025] Where, IoUthresholds is the preset threshold; Classes and class are both classes;

[0026] Step S5, evaluating the performance of the optimization model;

[0027] Calculate key indicators, including precision P, recall R and F1 score, which are defined as follows:

[0028]

[0029] In the formula, TP is a positive sample, FP is a positive negative sample, and FN is a negative positive sample.

[0030] Preferably, step S1 specifically includes collecting a plurality of test samples to ensure different representations of lighting conditions, tobacco textures and object types, and accurately reflecting the diversity of real-world tobacco processing environments.

[0031] Preferably, in step S1, the non-metallic foreign matter includes tiny fragments of plastic, paper, and wood; wherein the bounding box deletes the spatial dimension of the prefix experimental object, and the class label specifies the type of material, and the material type includes plastic and paper.

[0032] Preferably, step S3 specifically includes: the YOLOv10 model processes feature maps of three scales, using p j represents the predicted bounding box;

[0033] p j =(x j ,y j , w j ,h j );

[0034] In the formula, (x j ,y j ) is the center coordinate of the predicted bounding box, (w j ,h j ) are the width and height of the predicted bounding box, scaled relative to the input image size.

[0035] Preferably, at each detection scale, the YOLOv10 model predicts the bounding box and the corresponding object presence confidence score, as well as the class probability of each detected object; the architecture uses a semi-partial loss function to optimize the following predictions, specifically,

[0036] L = Lloc + Lconf + Lcls;

[0037] Where L is the total loss function of the target recognition system, which penalizes the design between the predicted ground truth bounding box coordinates; Lconf is the centroid prediction probability loss, which penalizes the wrong object and ensures that the model accurately identifies whether an object exists in the predicted bounding box; Lcls is the target category loss, which penalizes incorrect classification using cross entropy loss; Lloc is the positioning loss.

[0038] Preferably, in step S3, a combination of smoothing loss and joint loss is used to ensure improved overlap between the predicted bounding box regression and the actual box; double cross entropy is used to calculate the loss, in which the model learns to distinguish between foreground objects and background areas; the Lcls component indicates that the YOLOv10 model can detect the subject and correctly assign the subject to the appropriate category, which includes different types of impurities and background.

[0039] The present invention adopts the above-mentioned anchor frame tobacco chain small object detection method based on optimized YOLOv10, and the beneficial effects are as follows:

[0040] (1) The present invention utilizes a K-means clustering algorithm to redefine the size of the anchor box to accommodate the specific characteristics of foreign matter in the tobacco chain; this optimization improves the performance of the YOLOv10 model by better fitting the unique scale and aspect ratio of small impurities, allowing the YOLOv10 model to more effectively detect small objects in complex real industrial environments;

[0041] (2) The tobacco chain dataset constructed using Vacuum shows that the detection method of the present invention has significant improvements in detection accuracy compared to the baseline YOLOv10 model, especially in small object recall and precision;

[0042] (3) The method in this paper can be applied to a wider range of object detection tasks, including those with different object types and environmental conditions, which will provide further insights into the versatility of anchor optimization across different application domains;

[0043] (4) The present invention develops a customized tobacco line dataset, which provides a valuable resource for future research in similar industrial fields; and the robustness of the method in the present invention can be evaluated in more complex industrial environments, where changes in object appearance, lighting, and occlusion are more obvious.

[0044] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a precision-recall curve diagram of a baseline YOLOv10 model and an optimized YOLOv10 model in an embodiment of a method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 of the present invention;

[0046] Figure 2 This is a diagram of the detection results of two types of non-metallic foreign objects by the optimized YOLOv10 model in an embodiment of a method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 of the present invention; wherein (a) is a bright object, and (b) is an object similar in color to the tobacco chain but larger in size. DETAILED DESCRIPTION

[0047] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0048] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0049] Example

[0050] A method for detecting small objects in tobacco chains based on anchor frame optimization YOLOv10, comprising the following steps:

[0051] Step S1: Dataset construction, customizing a custom dataset.

[0052] A total of 5,000 test samples were collected to ensure different representations of lighting conditions, tobacco textures, and object types, accurately reflecting the diversity of real-world tobacco processing environments.

[0053] The custom dataset consists of high-resolution images of cut tobacco chains, each containing various non-metallic foreign objects, including tiny fragments of plastic, paper, wood, and other materials. Each image has an object with a bounding box and a corresponding class label. The bounding box removes the spatial dimensions of the prefix experimental object, and the class label specifies the type of material, which includes plastic and paper.

[0054] Step S2: Anchor box optimization.

[0055] Step S21: Given a set of real bounding boxes (w i ,h i ), w i and h i are the width and height of different groups of anchor boxes respectively.

[0056] Step S22: normalize the width W and height H by the corresponding image dimension. The objective function of the K-means clustering algorithm is expressed as:

[0057]

[0058] In the formula, μ k , μ j are all center points, corresponding to the optimized anchor box size; and are the width and height of the image corresponding to the center point of the j group of anchor boxes; N is the total number of bounding boxes; w i and h i are the width and height of different groups of anchor boxes respectively;

[0059] This example selects k=9 clusters to align with the number of anchor boxes used in the YOLOv10 model, in order to ensure that the model can effectively cover other variations in object size present in our dataset, thereby improving its ability to detect small impurities.

[0060] Step S3: integrate into the YOLOv10 model architecture.

[0061] The optimized anchor boxes can be seamlessly integrated into the YOLOv10 architecture to capture objects of different sizes using multiple multi-technique detection heads across different scales, each working on feature maps at different resolutions, allowing the model to effectively detect small, medium, and large objects.

[0062] The YOLOv10 model processes feature maps at three scales, using p j Represents the predicted bounding box.

[0063] p j =(x j ,yj , w j ,h j ).

[0064] In the formula, (x j ,y j ) is the center coordinate of the predicted bounding box, (w j ,h j ) are the width and height of the predicted bounding box, scaled relative to the input image size.

[0065] At each detection scale, the YOLOv10 model predicts a bounding box and the corresponding confidence score for the presence of an object, as well as the class probability of each detected object. The architecture uses a semi-partial loss function to optimize the following predictions, specifically,

[0066] L=Lloc+Lconf+Lcls.

[0067] Where L is the total loss function of the object recognition system, which penalizes the design between the predicted ground truth bounding box coordinates. Lconf is the centroid prediction probability loss, which penalizes wrong objects and ensures that the model accurately identifies whether an object exists within the predicted bounding box. Lcls is the target class loss, which penalizes incorrect classification using cross entropy loss. Lloc is the positioning loss.

[0068] A combination of smoothing loss and joint loss is used to ensure improved overlap between predicted bounding box regression and actual box. The loss is calculated using double cross entropy, where the model learns to distinguish between foreground objects and background regions. The Lcls component shows that the YOLOv10 model can detect the subjects and correctly assign them to appropriate categories, including different types of impurities and background.

[0069] Step S4: training procedure.

[0070] Step S41, the Adam optimizer is used, the learning rate is 0.001, and the optimization problem of the training model is expressed as follows:

[0071]

[0072] In the formula, θ * is the optimal value of parameter θ. is the loss function. f(x i ; θ) is the prediction function of the model. i is the actual observed value. N is the total number of data points.

[0073] Step S42: Calculate the mean average precision (AP) of each class, and average mAP over all classes and IoU thresholds. mAP is the mean average precision, providing a comprehensive method for model accuracy in different object scales and detection scenarios.

[0074] The average precision mAP is used to evaluate the performance of the model. The calculation method is as follows:

[0075]

[0076] Where, IoUthresholds is the preset threshold. Classes and class are both classes.

[0077] Step S5: Evaluate the performance of the optimization model.

[0078] Calculate key indicators, including precision P, recall R and F1 score, which are defined as follows:

[0079]

[0080] In the formula, TP is a positive sample, FP is a positive negative sample, and FN is a negative positive sample.

[0081] A performance comparison experiment was conducted between the optimized YOLOv10 model of this embodiment and the baseline YOLOv10 model, wherein the baseline model uses a fault anchor box.

[0082] (1) Test equipment:

[0083] The dataset is divided into training, validation and test sets in a ratio of 70%, 15%, and 15%. Training is performed on an NVIDIA TX3090 GPU with 24GB of memory using the Hot Torch deep learning framework. The baseline YOLOv10 model is initialized with the default anchor boxes from the COCOdatas set, while the optimized YOLOv10 model in this example uses anchor boxes generated by K-means clustering of the ground truth bounding boxes based on the tobacco dataset.

[0084] Both models were trained with the same hyperparameters: Adam optimizer, learning rate 0.001, batch size 16, and 100 epochs in total. To enhance the robustness of the models under different lighting conditions and object orientations, data augmentation techniques were applied during training, including random scaling, horizontal flipping, and colorization. Performance evaluation was performed on a held-out test set of 1,000 images annotated with non-encountered foreign objects embedded in tobacco chains.

[0085] (2) Performance comparison between the optimized YOLOv10 model in this embodiment and the baseline YOLOv10 model:

[0086] To evaluate the impact of anchor box optimization, we first degraded the performance of the baseline YOLOv10 model and the optimized YOLOv10 model in this example by several metrics. Table 1 provides a comparison of the precision, recall, mAP@50, and mAP@50-95 on the test set for the two models. The optimized YOLOv10 model shows significant improvements in small object detection, which is reflected in higher recall and map scores.

[0087] Table 1 Performance comparison

[0088] index Baseline YOLOv10 model Optimized YOLOv10 model Accuracy (%) 86.45 89.32 recall(%) 83.52 87.14 mAP@50(%) 90.57 93.28 mAP@50-95(%) 62.47 67.81

[0089] The results show:

[0090] The optimized YOLOv10 model achieves 93.28% at mAP@50, reflecting a 2.71% improvement over the baseline YOLOv10 model. In addition, the optimized YOLOv10 model achieves 67.81% at mAP@50-95, indicating a significant improvement over the baseline YOLOv10 model at 62.47% at mAP@50-95. These results underestimate the effectiveness of anchor box optimization in improving detection accuracy, especially for small objects, and thus emphasize the importance of customizing the anchor box size in this embodiment to improve small object detection performance.

[0091] (3) Precision recall analysis of the optimized YOLOv10 model in this embodiment and the baseline YOLOv10 model:

[0092] like Figure 1 As shown, the precision-recall curve of the algorithm is displayed, highlighting the trade-off between precision and recall for different confidence thresholds. The YOLOv10 model optimized by this embodiment consistently outperforms the baseline YOLOv10 model, achieving higher precision at multiple recalculation levels. This improvement is particularly evident in high recall areas, where the optimized YOLOv10 model maintains a good balance between precision and recall, indicating that it is able to detect the smallest foreign objects while identifying positive negative samples. The area under the precision-recall curve of the optimized YOLOv10 model is also high, further verifying the superior performance of this embodiment in detecting small non-metallic objects embedded in tobacco chains.

[0093] (4) The object detection performance of the optimized YOLOv10 model in this embodiment is smaller than that of the baseline YOLOv10 model:

[0094] Table 2 summarizes the performance of the baseline YOLOv10 model and the optimized YOLOv10 model in small object detection. The optimized YOLOv10 model shows a significant improvement in the recall rate of small samples, reaching 85.73%, while the small sample recall rate of the baseline YOLOv10 model is 78.91%. The precision of the optimized YOLOv10 model for small objects is also improved, reaching 87.65% after the optimization, an increase of 3.24% over the baseline YOLOv10 model.

[0095] Table 2 Performance comparison

[0096]

[0097]

[0098] The significant improvement in small object recall is attributed to the optimized anchor boxes, which better align the size and shape ratios of tiny foreign objects in the tobacco chain, resulting in more accurate bounding box predictions and fewer missed detections.

[0099] (5) The optimized YOLOv10 model in this embodiment detects two different types of non-metallic foreign matter in the tobacco chain:

[0100] The two different types of non-metallic foreign bodies are: Type 0 (brighter objects) and Type 1 (objects similar in color to tobacco chains but larger in size).

[0101] like Figure 2 Figure 2 shows examples of two test objects detected by the optimized YOLOv10 model. The confidence scores are shown on top of the bounding boxes, indicating how confident the optimized YOLOv10 model is in detecting each object.

[0102] (6) Inference time and efficiency test of the optimized YOLOv10 model in this embodiment and the baseline YOLOv10 model:

[0103] First, we ensure that the optimized YOLOv10 model maintains comparable inference time to the baseline YOLOv10 model. Table 3 shows the average inference time per image for the two models.

[0104] Model Average inference time (ms) Baseline YOLOv10 model 9.43 Optimized YOLOv10 model 16.0

[0105] The results show:

[0106] The optimized YOLOv10 model has only a slight increase in inference time (16.7ms for the optimized YOLOv10 model, compared to 16.0ms for the baseline YOLOv10 model). The slight increase in inference time for the optimized YOLOv10 model is outweighed by its significant detection accuracy, especially for small objects. This balance between accuracy and efficiency demonstrates the practicality of the method of this embodiment, which is suitable for real-time industrial applications.

[0107] Therefore, the present invention adopts the above-mentioned anchor frame tobacco chain small object detection method based on optimized YOLOv10, and detects small targets in the tobacco chain by refining the anchor frame in the YOLOv10 model, which can better adapt to small impurities, thereby significantly improving the detection performance, and the optimization and improvement have minimal impact on the overall performance, making the detection method feasible in real-time industrial applications where accuracy is prioritized.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for detecting small objects in tobacco chains based on anchor frame optimization YOLOv10, characterized in that: The following steps are involved: Step S1, data set construction, customizing a custom data set; The custom dataset consists of high-resolution images of cut tobacco chains, each containing various non-metallic foreign objects, with each object having a bounding box and corresponding class label; Step S2, anchor frame optimization; Step S21: Given a set of real bounding boxes (w i ,h i ), where w i and h i are the width and height of different groups of anchor boxes respectively; Step S22: normalize the width W and height H by the corresponding image dimension; the objective function of the K-means clustering algorithm is expressed as: In the formula, μ k , μ j are all center points, corresponding to the optimized anchor box size; where μ j =(μ w j , μ h j ), μ w j and μ h j are the width and height of the image corresponding to the center point of the j group of anchor boxes; N is the total number of bounding boxes; w i and h i are the width and height of different groups of anchor boxes respectively; Step S3, integrating into the YOLOv10 model architecture; Utilize multiple multi-technique detector heads across different scales to capture objects of different sizes, each working on feature maps at different resolutions; Step S4, training procedure; Step S41, the Adam optimizer is used, the learning rate is 0.001, and the optimization problem of the training model is expressed as follows: In the formula, θ * is the optimal value of parameter θ; is the loss function; f(x i ; θ) is the prediction function of the model; y i is the actual observed value; N is the total number of data points; Step S42, calculate the mean average precision (AP) of each class, and average mAP over all classes and IoU thresholds; where mAP is the mean average precision, providing a comprehensive method for model accuracy in different object scales and detection scenarios; The average precision mAP is used to evaluate the performance of the model. The calculation method is as follows: Where, IoU thresholds is the preset threshold; Classes and class are both classes; Step S5, evaluating the performance of the optimization model; Calculate key indicators, including precision P, recall R and F1 score, which are defined as follows: In the formula, TP is a positive sample, FP is a positive negative sample, and FN is a negative positive sample.

2. The method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 according to claim 1, characterized in that: Specifically, step S1 is to collect a number of test samples to ensure different performances of lighting conditions, tobacco textures and object types, and accurately reflect the diversity of real-world tobacco processing environments.

3. The method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 according to claim 1, characterized in that: In step S1, non-metallic foreign bodies include tiny fragments of plastic, paper, and wood; wherein the bounding box removes the spatial dimension of the prefix experimental object, and the class label specifies the type of material, which includes plastic and paper.

4. The method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 according to claim 1, characterized in that: Specifically, step S3 is that the YOLOv10 model processes feature maps of three scales, using p j represents the predicted bounding box; p j =(x j ,y j ,w j ,h j ); In the formula, (x j ,y j ) is the center coordinate of the predicted bounding box, (w j ,h j ) are the width and height of the predicted bounding box, scaled relative to the input image size.

5. The method for detecting small objects in tobacco chains based on anchor frame optimization of YOLOv10 according to claim 4, characterized in that: At each detection scale, the YOLOv10 model predicts a bounding box and the corresponding confidence score for the presence of an object, as well as the class probability of each detected object; the architecture uses a semi-partial loss function to optimize the following predictions, specifically, L = Lloc + Lconf + Lcls; Where L is the total loss function of the target recognition system, which penalizes the design between the predicted ground truth bounding box coordinates; Lconf is the centroid prediction probability loss, which penalizes the wrong object and ensures that the model accurately identifies whether an object exists in the predicted bounding box; Lcls is the target category loss, which penalizes incorrect classification using cross entropy loss; Lloc is the positioning loss.

6. The method for detecting small objects in an anchor frame tobacco chain based on optimized YOLOv10 according to claim 5, characterized in that: In step S3, a combination of smoothing loss and joint loss is used to ensure improved overlap between the predicted bounding box regression and the actual box; the loss is calculated using double cross entropy, in which the model learns to distinguish between foreground objects and background areas; the Lcls component shows that the YOLOv10 model can detect the subjects and correctly assign them to appropriate categories, including different types of impurities and backgrounds.

Images