Automatic plastic particle mixing equipment adopting image recognition and evaluation method

Abstract

The invention relates to the field of plastic processing, in particular to automatic plastic particle mixing equipment adopting image recognition and an evaluation method. The method comprises the steps that an industrial camera is used for shooting multi-angle images of plastic particles at fixed time intervals, contrast adjustment, brightness adjustment, denoising and normalization processing are conducted on the collected images, and the image quality is improved; converting the image from an RGB color space to an HSV color space; identifying the plastic particles by adopting a multi-target identification model to obtain position information of the plastic particles with multiple colors; the automation of the material mixing process is realized through automatic equipment; by improving the YOLO model, recognizing images of various plastic particles and extracting features, the plastic particles with corresponding colors can be recognized more accurately, position information can be provided, meanwhile, the dispersity of the plastic particles is judged through an algorithm, mixing uniformity information of the plastic particles is obtained, and the problem that manual naked eye recognition of the mixing degree depends on experience is solved.

Description

Technical Field

[0001] The present invention relates to the field of plastic processing, and in particular to an automatic mixing device for plastic particles using image recognition and an evaluation method. Background Art

[0002] Plastic mixing is a crucial step in plastics processing. Properly mixing new or masterbatch with old stock can effectively reduce production costs for plastics companies. However, traditional plastic pellet mixing often relies on manual operation or simple mechanical control, which is not only inefficient but also requires manual inspection before proceeding to the next step. Furthermore, it relies on visual inspection, which can lead to subjective and empirical differences in determining whether the mixing is uniform. Uneven mixing of new or masterbatch with old stock can lead to inconsistent product quality and, in more serious cases, even frequent breakage of plastic wire during drawing. In light of this, the present application is filed. Summary of the Invention

[0003] The technical problem to be solved by the present invention is that the determination of whether the mixing is uniform is subjective and based on experience. If the new material or masterbatch is not mixed evenly with the old material, the quality of the mixed product will be uneven. If the quality is to be guaranteed, it may even lead to frequent breakage of the plastic wire drawing. In order to solve the above technical problem, the present invention is implemented by the following technical solutions:

[0004] A method for automatically evaluating plastic particle mixing using image recognition, comprising:

[0005] An industrial camera is used to capture multi-angle images of plastic particles at fixed time intervals. The captured images are then subjected to contrast adjustment, brightness adjustment, denoising, and normalization to improve image quality. The images are then converted from the RGB color space to the HSV color space.

[0006] The plastic particles are identified using a multi-target recognition model to obtain the location information of plastic particles of various colors.

[0007] The coordinate set of each color particle is extracted from the position information of the plastic particles. The cluster analysis method is used to calculate the dispersion index of each color plastic, evaluate the uniformity of plastic particles of different colors in the image, and output the dispersion value of each color plastic.

[0008] According to the requirements of the mixing process, the dispersion threshold of each color of plastic particles is set; the dispersion value is compared with the dispersion threshold to determine whether it meets the mixing requirements.

[0009] Furthermore, the method of identifying the plastic particles using a multi-target recognition model to obtain position information of plastic particles of various colors includes:

[0010] The multi-target recognition model uses an improved YOLO model to identify plastic particles. The improved YOLO model uses YOLOv10 as its basic architecture, including a feature extraction network, a feature fusion network and a detection head. A C2f_CAA module is constructed in the feature extraction network, the ordinary convolution module in the feature extraction network is replaced by a DBB module, and the GSConv module and VovGSCSP module are introduced in the feature fusion network.

[0011] Furthermore, the C2f_CAA module introduces a CAA attention mechanism; the CAA attention mechanism divides the channel dimension into multiple sub-feature groups, optimizes the feature distribution of plastic particles through the CAA attention mechanism, and enhances the feature extraction capability of the improved YOLO model for plastic particles and plastic particle occlusions.

[0012] Furthermore, the ordinary convolution module in the feature extraction network is replaced with a DBB module; the DBB module introduces a nonlinear enhancement mechanism in the training stage through multi-branch design and structural reparameterization technology, enriches the feature expression space, improves the adaptability of the feature extraction network to plastic particles of different colors and shapes, and enhances the feature extraction capability of plastic particles of different colors and shapes; the DBB module adopts the mathematical formula:

[0013]

[0014] Output multi-branch structure, where P represents the output feature map of the DBB module and L represents the input feature map; R q and r q are the convolution kernel and bias term of the qth branch respectively; * indicates the convolution operation.

[0015] Furthermore, the GSConv module includes standard convolution of the main branch, depthwise separable convolution of the auxiliary branch, and feature concatenation and shuffling;

[0016] The standard convolution uses the formula:

[0017] P CS =Conv K×K (R in );

[0018] Calculate, where K×K is the convolution kernel size; R in is the input feature map; P CS Features output by the main branch;

[0019] The depth-wise separable convolution uses the formula:

[0020] P DSC =DConv K×K (PConv(R in));

[0021] Calculate, where P DSC Features output by the auxiliary branch; PConv is used for 1×1 convolution of channel mapping; DConv is used to reduce computational cost and focus on spatial features;

[0022] The VovGSCSP module includes input feature segmentation, main branch multi-scale feature extraction and feature fusion; the input feature segmentation adopts the formula:

[0023] R1, R2=Split(R in );

[0024] Perform feature segmentation, where R1 is the main branch feature and R2 is the bypass retention feature; Split is a segmentation operation, and the segmentation method is based on the number of channels, height or width;

[0025] The main branch multi-scale feature extraction adopts the formula:

[0026] R1 (i) =GSConv i (R1),i∈[1,n];

[0027] Extraction is performed, where i represents the i-th GSConv module in the VovGSCSP module, which is used to extract multi-scale features from the main branch features to enhance the model's detection ability for targets of different scales; n is the number of branches;

[0028] The feature fusion adopts the formula:

[0029]

[0030] Perform feature fusion, where R mg The final fused feature map is used for subsequent detection tasks to improve the detection performance of the model; GSConv f The final VovGSCSP module is used to further extract and compress the spliced ​​feature maps, enhance the feature expression ability through the feature shuffling mechanism, and reduce the amount of calculation; Concat represents the channel splicing operation, which is used to splice the feature maps of multiple branches in the channel dimension to form a richer feature representation; R1 (s) Represents the feature map extracted by the s-th branch.

[0031] Furthermore, the method of extracting the coordinate set of each color of plastic particles from the position information of the plastic particles, calculating the dispersion of each color of plastic using a cluster analysis method, evaluating the uniformity of plastic particles of different colors in the image, and outputting the dispersion value of each color of plastic includes:

[0032] Use the improved YOLO model to identify plastic particles in the image and extract the location coordinates of each plastic particle.

[0033] K-Means cluster analysis was used to calculate the dispersion of each color of plastic.

[0034] Randomly select K colors of plastic particles as the initial cluster centers, set as C1, C2, C3, ..., C K .

[0035] Calculate the distance d((x u ,y u ),C j ), where (x u ,y u ) is the position coordinate of each scattered plastic particle, C j is the location of the cluster center of the j-th color plastic particle.

[0036] The plastic particles are divided into Q clusters, each cluster represents the collection of plastic particle positions in a local area, using the formula:

[0037]

[0038] Assign the location of the scattered plastic particles to the cluster with the nearest cluster center, where (x Cj ,y Cj ) is the location coordinate of the j-th cluster center.

[0039] Using the formula:

[0040]

[0041] Calculate the dispersion index of each color plastic; where D is the dispersion index, C Q is the Qth cluster, O Q are the cluster center coordinates.

[0042] The uniformity of plastic particles of different colors in the image was evaluated based on the calculated dispersion index.

[0043] Furthermore, the improved YOLO model also obtains manufacturer information of the plastic particles, and learns the characteristics of plastic particles from different manufacturers by the improved YOLO model, thereby improving the generalization ability of the improved YOLO model.

[0044] A classification model is used to classify the features extracted by the improved YOLO model to distinguish plastic particles from different manufacturers.

[0045] Furthermore, the ratio information of the plastic particles is obtained, and the ratio information is used to adjust the dispersion thresholds of the plastic particles of different colors.

[0046] Furthermore, the step of identifying the plastic particles using a multi-target recognition model to obtain position information of plastic particles of various colors further includes:

[0047] The multi-target recognition model uses Wide ResNet (WRN) as the backbone network of the improved YOLO model.

[0048] Construct a synthetic dataset model of particles of various colors. In the synthetic dataset model, particles of different colors and sizes can be randomly added, and interference elements can be introduced. This allows the synthetic dataset model to learn how to distinguish and identify target particles in complex scenes during training, thereby improving the accuracy and robustness of recognition of particles of the same color. Interference elements include background noise, other objects, and human influence.

[0049] The improved YOLO model was trained on synthetic and real particle datasets. The performance of the improved YOLO model on different datasets was verified by comparing the training results. The trained improved YOLO model can more accurately identify particles of different colors in complex backgrounds and provide the coordinate distribution of particles of the same color.

[0050] Based on the training and verification results, the structure and parameters of the YOLO model are optimized and improved; the width factor of the WRN is adjusted and the module configuration of the Neck layer is optimized to improve the detection performance of the improved YOLO model.

[0051] Evaluate the performance of the improved YOLO model on an independent test set, including detection accuracy, recall rate, and mAP metrics. Based on the evaluation results, adjust the improved YOLO model or optimize the training strategy to ensure that the improved YOLO model can stably and accurately identify particles of different colors in the image and output the coordinate distribution of particles of the same color.

[0052] Based on the same inventive concept, on the other hand, the present invention also provides an automatic mixing equipment for plastic particles using image recognition, including a central control device, a first mixer, a second mixer, a feed hopper, a first silo, a second silo and a third silo; the first mixer includes a first mixing silo, a first feed hopper, a first mixing port, a first discharge port and a first elevator, and the second mixer includes a second mixing silo, a second feed hopper, a second mixing port, a second discharge port and a second elevator.

[0053] The discharge ports of the first silo and the second silo are connected to the first feed hopper; the first discharge port and the discharge port of the third silo are connected to the second feed hopper; the second discharge port is connected to the feed hopper; and the conveying hopper is connected to the plastic melting machine.

[0054] The central control device is used to control the mixing process of the first mixer and the second mixer and the discharge of the first silo, the second silo and the third silo.

[0055] Weighing sensors are provided in the first silo, the second silo and the third silo; the weighing sensors transmit weight data to the central control device.

[0056] Industrial cameras are installed at the first feed hopper and the second feed hopper respectively; the industrial cameras are used to collect mixed images of plastic particles.

[0057] Compared with the existing technology, the present invention has the following advantages and beneficial effects: the automation of the mixing process is achieved through automated equipment; by improving the YOLO model, images of multiple plastic particles are recognized and features are extracted, so that plastic particles of corresponding colors can be identified more accurately and position information can be provided. At the same time, the dispersion of plastic particles is judged by the algorithm to obtain the mixing uniformity information of plastic particles, solving the problem that the human eye relies on experience to identify the mixing degree. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 This is a flow chart of an automatic plastic particle mixing evaluation method using image recognition according to the present invention.

[0060] Figure 2 This is the architecture diagram of the improved YOLO model of the present invention.

[0061] Figure 3 This is the structural diagram of the C2f_CAA module of the present invention.

[0062] Figure 4 This is the structural diagram of the DBB module of the present invention.

[0063] Figure 5 This is the structural diagram of the GSConv module and VoVGSCSP module of the present invention.

[0064] Figure 6This is a schematic diagram of the plastic particle image annotation and recognition method according to the present invention.

[0065] Figure 7 This is a schematic diagram of labeling and identification under the condition of obstruction by plastic particles of the present invention. DETAILED DESCRIPTION

[0066] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0067] In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, it will be apparent to one skilled in the art that these specific details are not necessarily required to practice the present invention. In other embodiments, well-known structures, circuits, materials, or methods are not described in detail to avoid obscuring the present invention.

[0068] Throughout this specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present invention. Therefore, appearances of the phrases "one embodiment," "an embodiment," "an example," or "an example" in various places throughout this specification are not necessarily all referring to the same embodiment or example. Furthermore, the particular features, structures, or characteristics may be combined in one or more embodiments or examples in any suitable combinations and / or subcombinations. Furthermore, it will be understood by those of ordinary skill in the art that the figures provided herein are for illustrative purposes only and are not necessarily drawn to scale. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0069] In the description of the present invention, the terms "front", "back", "left", "right", "up", "down", "vertical", "horizontal", "high", "low", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the scope of protection of the present invention.

[0070] Plastic particle detection is a typical object detection task. Object detection methods can generally be divided into two categories: traditional machine vision methods and deep artificial intelligence learning methods. However, traditional machine vision methods are not only sensitive to changes in lighting and viewpoint, but also exhibit significant limitations when dealing with small objects in complex backgrounds.

[0071] Example 1: Figure 1As shown, an automatic plastic particle mixing evaluation method using image recognition is provided in an embodiment of the present invention, comprising:

[0072] An industrial camera is used to capture multi-angle images of plastic particles at fixed time intervals. The captured images are then subjected to contrast adjustment, brightness adjustment, denoising, and normalization to improve image quality. The images are then converted from the RGB color space to the HSV color space.

[0073] The plastic particles are identified using a multi-target recognition model to obtain the location information of plastic particles of various colors.

[0074] The coordinate set of each color particle is extracted from the position information of the plastic particles. The cluster analysis method is used to calculate the dispersion index of each color plastic, evaluate the uniformity of plastic particles of different colors in the image, and output the dispersion value of each color plastic.

[0075] According to the requirements of the mixing process, the dispersion threshold of each color of plastic particles is set; the dispersion value is compared with the dispersion threshold to determine whether it meets the mixing requirements.

[0076] Furthermore, the method of identifying the plastic particles using a multi-target recognition model to obtain position information of plastic particles of various colors includes:

[0077] The multi-target recognition model uses an improved YOLO model to identify plastic particles. The improved YOLO model uses YOLOv10 as its basic architecture, including a feature extraction network, a feature fusion network and a detection head. A C2f_CAA module is constructed in the feature extraction network, the ordinary convolution module in the feature extraction network is replaced by a DBB module, and the GSConv module and VovGSCSP module are introduced in the feature fusion network.

[0078] Specifically, as the latest version of the YOLO series, YOLOv10 further optimizes its network structure while maintaining high efficiency and accuracy. It boasts advantages such as a small number of parameters, fast computation, and strong compatibility, making it ideal for scenarios with limited hardware resources. YOLOv10 primarily consists of a backbone feature extraction network (Backbone), a neck network (Neck), and a detection head (Head). The input image first passes through the Backbone network, where features are extracted using designs such as large convolution kernels, a local self-attention module (PSA), and a spatial-channel separation downsampling module. The Backbone outputs feature maps, including multi-scale feature maps downsampled to 8×, 16×, and 32×. These feature maps are then fed into the Neck network. The Neck portion of YOLOv10 utilizes enhanced feature fusion modules (such as BiC and RepGFPN) to efficiently fuse multi-scale features, significantly improving small object detection capabilities. Finally, the fused feature maps are passed to the detection head, which outputs the classification result and the precise location of the object. The head portion of YOLOv10 adopts a lightweight design, further reducing computational cost and inference latency while maintaining detection accuracy. Improved YOLO model architecture diagram Figure 2 shown.

[0079] Furthermore, the C2f_CAA module introduces a CAA attention mechanism; the CAA attention mechanism divides the channel dimension into multiple sub-feature groups, optimizes the feature distribution of plastic particles through the CAA attention mechanism, and enhances the feature extraction capability of the improved YOLO model for plastic particles and plastic particle occlusions.

[0080] Specifically, the CAA attention mechanism is introduced into the Backbone network of YOLOv10, and the CAA attention module is used to improve the C2f module. CAA is inserted into the output stage of the C2f module, the C2f_CAA module is constructed and the C2f module in the Backbone network is replaced to enhance Backbone's ability to express key areas of the input feature map. Through the anchor mechanism, CAA can more accurately capture global semantics and key area features, thereby effectively focusing on the significant features of plastic particle targets in complex backgrounds while suppressing background noise interference. This improvement significantly improves Backbone's feature extraction capabilities for small targets and occluded targets, providing more efficient and accurate feature representation for subsequent Neck and Head parts, thereby improving the overall detection performance of the model. The C2f_CAA module structure diagram is shown below. Figure 3 shown.

[0081] Furthermore, the ordinary convolution module in the feature extraction network is replaced with a DBB module; the DBB module introduces a nonlinear enhancement mechanism in the training stage through multi-branch design and structural reparameterization technology, enriches the feature expression space, improves the adaptability of the feature extraction network to plastic particles of different colors and shapes, and enhances the feature extraction capability of plastic particles of different colors and shapes; the DBB module adopts the mathematical formula:

[0082]

[0083] Output multi-branch structure, where P represents the output feature map of the DBB module and L represents the input feature map; R q and r q are the convolution kernel and bias term of the qth branch respectively; * indicates the convolution operation.

[0084] Specifically, as a general convolution module, the DBB module enriches the feature expression space through a multi-branch design, and introduces a nonlinear enhancement mechanism during the training phase, effectively improving the feature diversity and network representation capabilities. In addition, the DBB module uses structural reparameterization technology to simplify the complex multi-branch structure into only a single convolution during the inference phase, thereby balancing model performance and inference efficiency. This design makes the DBB module not only suitable for single-stage target detection models, but also demonstrates excellent feature learning capabilities in a variety of tasks. The introduction of the DBB module into YOLOv10 enhances the adaptability of the backbone network to complex spatial and semantic features. By adopting the multi-branch design of the DBB module, the network can flexibly capture diverse and complex feature patterns, which is crucial for detecting plastic particle targets of different colors and shapes. Structural reparameterization technology ensures that this enhanced learning ability does not impair the efficiency of the model during inference, and the YOLOv10 model after replacing the original convolution with the DBB module convolution greatly improves the expression capability of a single convolution. The DBB convolution structure diagram is shown below. Figure 4 shown.

[0085] Furthermore, the GSConv module includes standard convolution of the main branch, depthwise separable convolution of the auxiliary branch, and feature concatenation and shuffling;

[0086] The standard convolution uses the formula:

[0087] P CS =Conv K×K (R in );

[0088] Calculate, where K×K is the convolution kernel size; R in is the input feature map; P CS Features output by the main branch;

[0089] The depth-wise separable convolution uses the formula:

[0090] P DSC =DConv K×K (PConv(R in ));

[0091] Calculate, where P DSC Features output by the auxiliary branch; PConv is used for 1×1 convolution of channel mapping; DConv is used to reduce computational cost and focus on spatial features;

[0092] The VovGSCSP module includes input feature segmentation, main branch multi-scale feature extraction and feature fusion; the input feature segmentation adopts the formula:

[0093] R1, R2=Split(R in );

[0094] Perform feature segmentation, where R1 is the main branch feature and R2 is the bypass retention feature; Split is a segmentation operation, and the segmentation method is based on the number of channels, height or width;

[0095] The main branch multi-scale feature extraction adopts the formula:

[0096] R1 (i) =GSConv i (R1),i∈[1,n];

[0097] Extraction is performed, where i represents the i-th GSConv module in the VovGSCSP module, which is used to extract multi-scale features from the main branch features to enhance the model's detection ability for targets of different scales; n is the number of branches;

[0098] The feature fusion adopts the formula:

[0099]

[0100] Perform feature fusion, where R mg The final fused feature map is used for subsequent detection tasks to improve the detection performance of the model; GSConv f The final VovGSCSP module is used to further extract and compress the spliced ​​feature maps, enhance the feature expression ability through the feature shuffling mechanism, and reduce the amount of calculation; Concat represents the channel splicing operation, which is used to splice the feature maps of multiple branches in the channel dimension to form a richer feature representation; R1 (s) Represents the feature map extracted by the s-th branch.

[0101] Specifically, the feature fusion network is an efficient neck architecture designed for lightweight target detection networks. Its core is to significantly reduce the computational cost while ensuring the model detection performance by introducing an efficient GSConv module and an optimized feature fusion strategy. Compared with the traditional neck network, the feature fusion network significantly reduces the number of parameters and computational complexity while maintaining detection accuracy, which makes it perform particularly well in real-time detection tasks and resource-constrained scenarios. On this basis, the feature fusion network further introduces an optimized VoVGSCSP (VoVNet grouping and CSP) module to enhance the multi-scale feature fusion capability. The VoVGSCSP module combines the multi-branch feature extraction of VoVNet with the cross-stage feature fusion mechanism of CSP (cross-stage part). The structure diagram of the GSConv module and the VoVGSCSP module is shown below. Figure 5 shown.

[0102] Furthermore, the method of extracting the coordinate set of each color of plastic particles from the position information of the plastic particles, calculating the dispersion of each color of plastic using a cluster analysis method, evaluating the uniformity of plastic particles of different colors in the image, and outputting the dispersion value of each color of plastic includes:

[0103] Use the improved YOLO model to identify plastic particles in the image and extract the location coordinates of each plastic particle.

[0104] K-Means cluster analysis was used to calculate the dispersion of each color of plastic.

[0105] Randomly select K colors of plastic particles as the initial cluster centers, set as C1, C2, C3, ..., C K .

[0106] Calculate the distance d((x u ,y u ),C j ), where (x u ,y u ) is the position coordinate of each scattered plastic particle, C j is the location of the cluster center of the j-th color plastic particle.

[0107] The plastic particles are divided into Q clusters, each cluster represents the collection of plastic particle positions in a local area, using the formula:

[0108]

[0109] Assign the location of the scattered plastic particles to the cluster with the nearest cluster center, where (x Cj ,y Cj) is the location coordinate of the j-th cluster center.

[0110] Using the formula:

[0111]

[0112] Calculate the dispersion index of each color plastic; where D is the dispersion index, C Q is the Qth cluster, O Q are the cluster center coordinates.

[0113] The uniformity of plastic particles of different colors in the image was evaluated based on the calculated dispersion index.

[0114] Specifically, the uniformity of plastic particles of different colors within the image is assessed based on the calculated dispersion index. A lower dispersion index indicates a more even distribution of particles, while a higher dispersion index indicates a more concentrated or uneven distribution. The dispersion value for each color of plastic particles is output, and a preset dispersion threshold is used to determine whether the particle size meets the mixing requirements. For example, if the dispersion value is less than or equal to the dispersion threshold, the plastic particles of that color are considered evenly distributed; otherwise, they are considered unevenly distributed.

[0115] Furthermore, the improved YOLO model also obtains manufacturer information of the plastic particles, and learns the characteristics of plastic particles from different manufacturers by the improved YOLO model, thereby improving the generalization ability of the improved YOLO model.

[0116] A classification model is used to classify the features extracted by the improved YOLO model to distinguish plastic particles from different manufacturers.

[0117] Furthermore, the ratio information of the plastic particles is obtained, and the ratio information is used to adjust the dispersion thresholds of the plastic particles of different colors.

[0118] Specifically, the experimental environment was Windows 11, with an Intel Core i7-12700K CPU, an NVIDIA GeForce RTX 3080 GPU, and 64GB of RAM. Model training and testing were performed using the PyTorch framework, with the following training parameters: learning rate: 0.01; momentum: 0.937; weight decay: 0.00005; image size: 640×640; number of training epochs: 180; and batch size: 16.

[0119] Precision, recall, and mean average precision (mAP@0.5) are used as evaluation indicators. The experimental results are shown in the following table:

[0120] Model Name Accuracy (%) Recall rate (%) Average precision (%) YOLOv10 model 91.7 93.7 97.4 Improving the YOLO model 96.4 96.1 98.9

[0121] Results analysis: 1. The improved YOLO model has an accuracy improvement of 4.7% compared to the YOLOv10 model. This is because the introduced CAA attention mechanism can effectively enhance the feature extraction capability of the central area of ​​plastic particles, allowing the model to pay more attention to the key features of the target, thereby improving the accuracy of detection. At the same time, the introduction of the DBB module enriches the feature expression space and further improves the model's ability to identify plastic particles in complex backgrounds. 2. The recall rate of the improved YOLO model is 2.4% higher than that of the YOLOv10 model. This shows that the improved model can better detect plastic particles in the image and reduce missed detections. At the same time, the GSConv and VoVGSCSP modules optimize the feature fusion process, enabling the model to more effectively extract multi-scale features, thereby improving the detection capability of plastic particles of different sizes. 3. The average precision of the improved YOLO model reached 98.9%. This result shows that the improved model has been significantly improved in both detection accuracy and robustness.

[0122] Furthermore, the step of identifying the plastic particles using a multi-target recognition model to obtain position information of plastic particles of various colors further includes:

[0123] The multi-target recognition model uses Wide ResNet (WRN) as the backbone network of the improved YOLO model.

[0124] Construct a synthetic dataset model of particles of various colors. In the synthetic dataset model, particles of different colors and sizes can be randomly added, and interference elements can be introduced. This allows the synthetic dataset model to learn how to distinguish and identify target particles in complex scenes during training, thereby improving the accuracy and robustness of recognition of particles of the same color. Interference elements include background noise, other objects, and human influence.

[0125] The improved YOLO model was trained on synthetic and real particle datasets. The performance of the improved YOLO model on different datasets was verified by comparing the training results. The trained improved YOLO model can more accurately identify particles of different colors in complex backgrounds and provide the coordinate distribution of particles of the same color.

[0126] Based on the training and verification results, the structure and parameters of the YOLO model are optimized and improved; the width factor of the WRN is adjusted and the module configuration of the Neck layer is optimized to improve the detection performance of the improved YOLO model.

[0127] Evaluate the performance of the improved YOLO model on an independent test set, including detection accuracy, recall rate, and mAP metrics. Based on the evaluation results, adjust the improved YOLO model or optimize the training strategy to ensure that the improved YOLO model can stably and accurately identify particles of different colors in the image and output the coordinate distribution of particles of the same color.

[0128] Based on the same inventive concept, on the other hand, the present invention also provides an automatic mixing equipment for plastic particles using image recognition, including a central control device, a first mixer, a second mixer, a feed hopper, a first silo, a second silo and a third silo; the first mixer includes a first mixing silo, a first feed hopper, a first mixing port, a first discharge port and a first elevator, and the second mixer includes a second mixing silo, a second feed hopper, a second mixing port, a second discharge port and a second elevator.

[0129] The discharge ports of the first silo and the second silo are connected to the first feed hopper; the first discharge port and the discharge port of the third silo are connected to the second feed hopper; the second discharge port is connected to the feed hopper; and the conveying hopper is connected to the plastic melting machine.

[0130] The central control device is used to control the mixing process of the first mixer and the second mixer and the discharge of the first silo, the second silo and the third silo.

[0131] Weighing sensors are provided in the first silo, the second silo and the third silo; the weighing sensors transmit weight data to the central control device.

[0132] Industrial cameras are installed at the first feed hopper and the second feed hopper respectively; the industrial cameras are used to collect mixed images of plastic particles.

[0133] Specifically, through deep learning of large amounts of plastic pellet images and mixing data, the system can automatically identify subtle differences between batches of plastic pellets and dynamically adjust the mixing control strategy based on actual production conditions. For example, if the physical properties of plastic pellets change due to changes in the raw material supplier or production environment, the system can automatically optimize the image recognition model and mixing control parameters without manual intervention, ensuring that the mixing process remains highly accurate.

[0134] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatically evaluating plastic particle mixing using image recognition, characterized in that: The method comprises: Use an industrial camera to capture multi-angle images of plastic particles at fixed time intervals, perform contrast adjustment, brightness adjustment, denoising, and normalization on the captured images to improve image quality; and convert the images from RGB color space to HSV color space. Plastic particles are identified using a multi-target recognition model to obtain the location information of plastic particles of various colors; The coordinate set of each color particle is extracted from the position information of the plastic particles. The dispersion index of each color plastic is calculated using cluster analysis method to evaluate the uniformity of different color plastic particles in the image and output the dispersion value of each color plastic. According to the requirements of the mixing process, the dispersion threshold of each color of plastic particles is set; the dispersion value is compared with the dispersion threshold to determine whether it meets the mixing requirements.

2. The method for automatically evaluating plastic particle mixing using image recognition according to claim 1, characterized in that: The method of using a multi-target recognition model to identify plastic particles and obtaining position information of plastic particles of various colors includes: The multi-target recognition model uses an improved YOLO model to identify plastic particles. The improved YOLO model uses YOLOv10 as its basic architecture, including a feature extraction network, a feature fusion network and a detection head. A C2f_CAA module is constructed in the feature extraction network, the ordinary convolution module in the feature extraction network is replaced by a DBB module, and the GSConv module and VovGSCSP module are introduced in the feature fusion network.

3. The method for automatically evaluating plastic particle mixing using image recognition according to claim 2, characterized in that: The C2f_CAA module introduces a CAA attention mechanism; the CAA attention mechanism divides the channel dimension into multiple sub-feature groups, optimizes the feature distribution of plastic particles through the CAA attention mechanism, and enhances the feature extraction capability of the improved YOLO model for plastic particles and plastic particle occlusions.

4. The method for automatically evaluating plastic particle mixing using image recognition according to claim 2, wherein: The ordinary convolution module in the feature extraction network is replaced with a DBB module; the DBB module introduces a nonlinear enhancement mechanism in the training stage through multi-branch design and structural reparameterization technology to enrich the feature expression space, improve the adaptability of the feature extraction network to plastic particles of different colors and shapes, and enhance the feature extraction capability of plastic particles of different colors and shapes; the DBB module adopts the mathematical formula: Output multi-branch structure, where P represents the output feature map of the DBB module and L represents the input feature map; R q and r q are the convolution kernel and bias term of the qth branch respectively; * indicates the convolution operation.

5. The method for automatically evaluating plastic particle mixing using image recognition according to claim 2, characterized in that: The GSConv module includes standard convolution of the main branch, depth-wise separable convolution of the auxiliary branch, and feature concatenation and shuffling; The standard convolution uses the formula: P CS =Conv K×K (R in ); Calculate, where K×K is the convolution kernel size; R in is the input feature map; P CS Features output by the main branch; The depth-wise separable convolution uses the formula: P DSC =DConv K×K (PConv(R in )); Calculate, where P DSC Features output by the auxiliary branch; PConv is used for 1×1 convolution of channel mapping; DConv is used to reduce computational cost and focus on spatial features; The VovGSCSP module includes input feature segmentation, main branch multi-scale feature extraction and feature fusion; the input feature segmentation adopts the formula: R1,R2=Split(R in ); Perform feature segmentation, where R1 is the main branch feature and R2 is the bypass retention feature; Split is a segmentation operation, and the segmentation method is based on the number of channels, height or width; The main branch multi-scale feature extraction adopts the formula: R1 (i) =GSConv i (R1),i∈[1,n]; Extraction is performed, where i represents the i-th GSConv module in the VovGSCSP module, which is used to extract multi-scale features from the main branch features to enhance the model's detection ability for targets of different scales; n is the number of branches; The feature fusion adopts the formula: Perform feature fusion, where R mg The final fused feature map is used for subsequent detection tasks to improve the detection performance of the model; GSConv f The final VovGSCSP module is used to further extract and compress the spliced ​​feature maps, enhance the feature expression ability through the feature shuffling mechanism, and reduce the amount of calculation; Concat represents the channel splicing operation, which is used to splice the feature maps of multiple branches in the channel dimension to form a richer feature representation; R1 (s) Represents the feature map extracted by the s-th branch.

6. A method for automatically evaluating plastic particle mixing using image recognition according to any one of claims 2 to 5, characterized in that: The method of extracting the coordinate set of each color particle from the position information of the plastic particles, calculating the dispersion index of each color plastic using a cluster analysis method, evaluating the uniformity of the different color plastic particles in the image, and outputting the dispersion value of each color plastic includes: Use the improved YOLO model to identify plastic particles in the image and extract the location coordinates of each plastic particle; The K-Means cluster analysis method is used to calculate the dispersion of each color of plastic including: Randomly select K colors of plastic particles as the initial cluster centers, set as C1, C2, C3, ..., C K ; Calculate the distance d((x u ,y u ),C j ), where (x u ,y u ) is the position coordinate of each scattered plastic particle, C j is the location of the cluster center of the j-th color plastic particle; The plastic particles are divided into Q clusters, each cluster represents the collection of plastic particle positions in a local area, using the formula: Assign the location of the scattered plastic particles to the cluster with the nearest cluster center, where (x Cj ,y Cj ) is the location coordinate of the j-th cluster center; Using the formula: Calculate the dispersion index of each color plastic; where D is the dispersion index, C Q is the Qth cluster, O Q is the cluster center position coordinate; The uniformity of plastic particles of different colors in the image was evaluated based on the calculated dispersion index.

7. The method for automatically evaluating plastic particle mixing using image recognition according to claim 6, characterized in that: The improved YOLO model also obtains manufacturer information of plastic particles, and learns the characteristics of plastic particles from different manufacturers by improving the generalization ability of the improved YOLO model; A classification model is used to classify the features extracted by the improved YOLO model to distinguish plastic particles from different manufacturers.

8. The method for automatically evaluating plastic particle mixing using image recognition according to claim 7, characterized in that: The ratio information of the plastic particles is also obtained, and the ratio information is used to adjust the dispersion threshold value of the plastic particles of different colors.

9. The method for automatically evaluating plastic particle mixing using image recognition according to claim 1, characterized in that: The method of identifying the plastic particles using a multi-target recognition model to obtain position information of plastic particles of various colors further includes: The multi-target recognition model uses Wide ResNet (WRN) as the backbone network of the improved YOLO model; Construct a synthetic dataset model for particles of various colors. This model can randomly add particles of different colors and sizes, and introduce interference elements. This allows the synthetic dataset model to learn how to distinguish and identify target particles in complex scenes during training, improving the accuracy and robustness of identifying particles of the same color. Interference elements include background noise, other objects, and human influence. The improved YOLO model was trained on a synthetic dataset and a real particle dataset. The performance of the improved YOLO model on different datasets was verified by comparing the training results. The trained improved YOLO model was able to more accurately identify particles of different colors in complex backgrounds and provide the coordinate distribution of particles of the same color. Based on the training and verification results, the structure and parameters of the YOLO model were optimized and improved. The width factor of the WRN and the module configuration of the Neck layer were adjusted to improve the detection performance of the improved YOLO model. Evaluate the performance of the improved YOLO model on an independent test set, including detection accuracy, recall rate, and mAP metrics. Based on the evaluation results, adjust the improved YOLO model or optimize the training strategy to ensure that the improved YOLO model can stably and accurately identify particles of different colors in the image and output the coordinate distribution of particles of the same color.

10. An automatic mixing device for plastic particles using image recognition, characterized in that: The device implements the steps of the method according to any one of claims 1 to 9, comprising a central control device, a first mixer, a second mixer, a feed hopper, a first silo, a second silo, and a third silo; the first mixer comprises a first mixing silo, a first feed hopper, a first mixing port, a first discharge port, and a first elevator; the second mixer comprises a second mixing silo, a second feed hopper, a second mixing port, a second discharge port, and a second elevator; The discharge ports of the first silo and the second silo are connected to the first feed hopper; the first discharge port and the discharge port of the third silo are connected to the second feed hopper; the second discharge port is connected to the feed hopper; and the feed hopper is connected to the plastic melter; The central control device is used to control the mixing process of the first mixer and the second mixer and the discharge of the first silo, the second silo and the third silo; The first silo, the second silo, and the third silo are provided with weighing sensors; the weighing sensors transmit weight data to the central control device; Industrial cameras are installed at the first feed hopper and the second feed hopper respectively; the industrial cameras are used to collect mixed images of plastic particles.

Images