A particle morphology quantification method based on an improved YOLOv8 model and generated data enhancement

Abstract

This invention provides a particle morphology quantization method based on an improved YOLOv8 model and generative data augmentation, belonging to the fields of image processing and geotechnical engineering. The method includes: labeling particle images to construct a training dataset containing morphological features; building a particle image synthesis module based on a generative adversarial network (GAN) to generate diverse particle images using semantic tags to expand the training data; improving the YOLOv8 model by adding a morphological parameter regression branch to construct a YOLOv8-PMQ model capable of end-to-end prediction of particle position and morphological parameters, achieving simultaneous prediction of particle position and morphological parameters; performing mixed training using synthetic and real data; and realizing rapid particle image detection and parameter output through a mobile application. This method achieves efficient morphological recognition of densely packed particles without segmentation computation, has low computational resource consumption, and flexible deployment, providing an efficient, lightweight, and practical particle morphology quantization solution.

Description

Technical Field

[0001] This invention relates to the fields of image processing and geotechnical engineering, and in particular to a particle morphology quantization method based on an improved YOLOv8 model and generated data augmentation. Background Technology

[0002] Particle morphology, including particle size and shape, represents a set of fundamental physical parameters characterizing granular soils. These parameters largely control interparticle interactions, thereby influencing basic macroscopic properties such as porosity, compaction, and compressive or shear strength. Traditionally, particle size distribution is determined by mechanical sieving, while particle shape is assessed by manual measurement or visual comparison with standard reference images. Although these methods are widely adopted due to their simplicity, their inherent inefficiency and high subjectivity make them unsuitable for large-scale quantitative analysis. While deep learning methods based on image segmentation utilize semantic segmentation networks such as Mask R-CNN and UNet to extract masks from particle images and perform geometric calculations, enabling automated analysis, they are computationally complex, have slow inference speeds, and cannot be deployed on mobile devices. To address these shortcomings, this invention proposes a particle morphology quantification method combining Generative Adversarial Networks (GANs) and the lightweight detection model YOLOv8-PMQ, achieving end-to-end prediction of morphology parameters and mobile deployment. Summary of the Invention

[0003] The purpose of this invention is to provide a particle morphology quantization method based on an improved YOLOv8 model and generative data augmentation. This addresses the technical problems of existing deep learning methods based on image segmentation, which rely on semantic segmentation networks such as Mask R-CNN and UNet to extract masks from particle images before performing geometric calculations. These methods are computationally complex, have slow inference speeds, and are unsuitable for mobile deployment. This new method achieves efficient morphology recognition of densely packed particles without requiring segmentation calculations. It consumes minimal computational resources, is flexible in deployment, and provides an efficient, lightweight, and practical particle morphology quantization solution.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A particle morphology quantization method based on an improved YOLOv8 model and augmented generative data, the method comprising the following steps: Step S1: Obtain and label particle image samples to form a dataset containing particle boundaries and morphological parameter labels; Step S2: Construct a particle image synthesis module based on generative adversarial network to generate synthesized particle images from the input semantic labels, which are used to expand the training dataset; Step S3: Construct the YOLOv8-PMQ particle detection model. Add a morphological parameter regression branch to the YOLOv8 detection head to simultaneously output the particle detection box and morphological parameters. Step S4: Use the synthetic image generated in step S2 to perform mixed training with the real image to obtain a target detection model with higher morphological parameter regression accuracy; S5. Input the particle image to be analyzed into the trained YOLOv8-PMQ particle detection model, and output the particle bounding box and morphological parameters. S6. Embed the target detection model into a mobile application to achieve rapid on-site particle morphology detection and statistical analysis.

[0005] Furthermore, in step S2, the generative adversarial network adopts a semantic image generation model based on the OASIS structure, which includes a generator and a discriminator. The generator converts the label information into particle images through several convolution operations, batch normalization and activation functions. The discriminator adopts a U-Net structure to distinguish between generated images and real images.

[0006] Furthermore, the generator minimizes the adversarial loss by feeding the synthesized image into the discriminator network. L adv To update the generator's parameters, the generator loss function. L G The definition is as follows: Where y represents the input label, z is the noise tensor introduced to enhance the diversity of the synthesized image, N is the total number of categories (set to N=3, including complete particles, incomplete particles, and background), G(z, y) represents the image synthesized by the generator given the noise z and the label y, and D(G(z, y)) represents the pixel prediction on the synthesized image, outputting the probability that each pixel belongs to a specific category. y, j, c refers to the pixel at position (i, j) in label y belonging to category c. H × W α represents the resolution of the image. c The weights assigned to address pixel frequency imbalances are calculated using the following formula: .

[0007] Furthermore, discriminator loss L D Depend on L real and L fake Composed of two components, L D The definition is as follows: in x It refers to a true image of particles.

[0008] Furthermore, in step S1, the morphological parameters include roundness R and sphericity S. A Length, width, and sphericity S E Equivalent projected circle diameter d EQPC Maximum Freette diameter d Fmax and minimum Freette diameter d Fmin .

[0009] Furthermore, in step S3, the probe network of the YOLOv8-PMQ particle detection model extends the existing YOLOv8 design and includes three branches: a classification branch, a bounding box regression branch, and a new particle morphology parameter regression branch. These three branches are responsible for particle classification, localization, and morphology estimation, respectively.

[0010] Furthermore, in step S3, the total loss function of the YOLOv8-PMQ particle detection model... L Detect Classification branch loss L Cls Bounding box regression branch loss L Bbox and morphological parameter regression branch loss L Param The summation is constructed, and its expression is: Where N is the number of samples; .

[0011] Furthermore, in step S4, the hybrid training adopts a two-stage training strategy. In the first stage, real particle image data is used to train the YOLOv8-PMQ particle detection model. In the second stage, synthetic particle images generated by adversarial networks are introduced for joint training. A reduction factor λ is introduced, where 0 < λ < 1, to adjust the weight of the synthetic image in the loss function. The optimal value of λ is selected by searching in increments of 0.1.

[0012] Furthermore, step S6 also includes deploying the trained model on an Android smartphone for offline operation, requiring no cloud computing resources and allowing for on-site use.

[0013] The present invention, by adopting the above-described technical solution, has the following beneficial effects: The YOLOv8-PMQ model proposed in this invention simplifies the traditional two-step process of segmentation + geometric calculation into a one-step process of detection + regression, eliminating complex instance segmentation and post-processing calculations. The model has a small number of parameters and fast inference speed, making it ideal for deployment on mobile devices or embedded platforms with limited computing resources. By introducing an OASIS-based GAN image synthesis framework, a large number of diverse synthetic particle images can be generated from limited labeled data, effectively expanding the training set and improving the model's generalization ability and the accuracy of morphological parameter regression. Although the detection accuracy is slightly lower than that of large segmentation models, this invention exhibits better stability in morphological parameter (especially roundness) prediction and has significant advantages in efficiency, model size, and deployment flexibility, achieving a good balance between accuracy and practicality. It provides a complete technology chain from data preparation and model training to mobile deployment, allowing users to quickly obtain multiple morphological parameters of particles simply by capturing particle images, greatly improving the efficiency and convenience of on-site detection. Attached Figure Description

[0014] Figure 1 This is a flowchart of the overall method of an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of the GAN image synthesis framework in an embodiment of the present invention.

[0016] Figure 3 This is a network architecture diagram of the YOLOv8-PMQ model in an embodiment of the present invention.

[0017] Figure 4 This is a schematic diagram of the hybrid training strategy in an embodiment of the present invention.

[0018] Figure 5 These are the training and verification loss curves of the generator and discriminator in this embodiment of the invention.

[0019] Figure 6 This is a schematic diagram of the particle morphology descriptor used in the embodiments of the present invention.

[0020] Figure 7 This is a schematic diagram of the field test results of the mobile application in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.

[0022] like Figure 1 As shown, a particle morphology quantization method based on an improved YOLOv8 model and generated data augmentation is proposed. The method includes the following steps: S1. Obtain and label particle image samples to form a dataset containing particle boundaries and morphological parameter labels; S2. Construct a particle image synthesis module based on Generative Adversarial Network (GAN) to generate diverse synthetic particle images from the input semantic labels to expand the training dataset. S3. Construct a lightweight particle detection model using YOLOv8-PMQ. Add a morphological parameter regression branch to the YOLOv8 detection head to simultaneously output the detection box and morphological parameters of the particles. S4. Use the synthetic image generated in step (2) and the real image to train in order to obtain a target detection model with higher morphological parameter regression accuracy; S5. Input the particle image to be analyzed into the trained YOLOv8-PMQ model and output the particle bounding box and morphological parameters. S6. Embed the detection model into a mobile application to achieve rapid on-site particle morphology detection and statistical analysis.

[0023] The following is a detailed explanation of each of the above steps: In step S1 above: 820 images containing gravel and sand are obtained from a public dataset (such as the Clear Data Bay platform). 420 of these are gravel images with a resolution of 6720 × 4480 pixels, and 400 are sand images with a resolution of 896 × 896 pixels. To ensure consistent input size, all images and their corresponding labels are resized to 1024 × 1024 pixels using bicubic interpolation. The label for each image consists of three pixel categories: intact grains (white), incomplete grains (gray), and background area (black, representing blurry or unrecognizable areas).

[0024] A total of 3,744 complete particles and 31,112 incomplete particles were annotated across the entire dataset, with each image containing an average of approximately 4 complete particles and 35 incomplete particles.

[0025] Furthermore, to improve recognition accuracy and generalization ability, especially when data is limited, basic data augmentation techniques, such as image rotation and flipping, are applied to expand the dataset. The dataset is then divided into three different training sets in a ratio of 7:1:2 to optimize model parameters, a validation set to tune hyperparameters and monitor potential overfitting during training, and a test set to evaluate the model's performance after training is complete.

[0026] In step S2 above: refer to Figure 2 Generative adversarial networks employ a semantic image generation model based on the OASIS structure, which includes a generator and a discriminator.

[0027] The generator employs a typical decoder structure, converting label information into particle images (e.g., ReLU, LeakyReLU, and Tanh) through multiple convolution operations, batch normalization, and activation functions; the specific steps are as follows: The generator will generate a size of H × W × N The labels are used as input, where H and W The height and width are represented by , and N represents the number of channels, with each channel corresponding to a semantic category (e.g., complete particle, incomplete particle, and background). To introduce variability in image synthesis, the size is . H × W A randomly generated noise tensor of size ×64 is concatenated along the channel dimension with the labels. Each pixel in the noise tensor is sampled from a normal distribution, serving as a "random seed" for the generator. Combined with the labels, it forms a tensor of size ×64. H × W ×( N The input of +64) enables the generator to generate different particle images under the same label conditions.

[0028] The input is first downsampled to H / 32× W / 32×( N To reduce computational cost, a feature map with 1024 channels is generated through 3×3 convolution operations. This is followed by processing the feature map through six SPADE residual blocks (each consisting of SPADE, ReLU, and convolution operations) and five upsampling operations to progressively restore spatial resolution and enhance detail representation. The final feature map size is [size missing]. H × W ×64. In the output stage, a 1×1 convolution is applied to reduce the number of channels to 3. Then, the output is passed through LeakyReLU and Tanh activation functions to generate a new particle image aligned with the semantic structure of the input labels.

[0029] The generator passes the synthesized image to the discriminator network, minimizing the adversarial loss. L adv This loss function updates the generator's parameters. It evaluates the probability that the discriminator will identify the synthesized image as a real image. Generator Loss Function L G The definition is as follows: Where y represents the input label, and z is a noise tensor introduced to enhance the diversity of the synthesized image. N It is the total number of categories, set to [value] in this study. N = 3, including complete particles, incomplete particles, and background. G (z, y) represents the image synthesized by the generator given noise z and label y, while D (G(z, y)) represents the pixel prediction on the synthetic image, outputting the probability that each pixel belongs to a specific class. y, j, c refers to the pixel at position (i, j) in label y belonging to class c. H × W Indicates the image resolution. α c The weights assigned to address pixel frequency imbalances are calculated using the following formula: The discriminator employs a symmetric U-Net structure consisting of two stages: encoding and decoding. It will have a size of... H × W A 3x3 image is used as input; this image can be a real image or a synthesized image by the generator. The purpose is to determine whether the image is real or fake. The specific steps are as follows: During the encoding phase, the image is processed sequentially through five downsampled residual blocks (ResBlocks-Down). Spatial resolution from H × W Gradually decrease to H / 32× W / 32, while the number of channels is increased to 512 to extract rich discriminative features. Each ResBlock-Down consists of convolutional operations, spectral normalization, LeakyReLU activation, and average pooling. Unlike batch normalization used in the generator to accelerate convergence, spectral normalization is used to regulate the convergence speed of the feature map and prevent it from dominating the adversarial training process. Average pooling reduces the resolution of the feature map by sliding a 2×2 window across it and calculating the average value within each window, which helps preserve global semantic information.

[0030] In the decoding phase, this method uses five upsampled residual blocks (ResBlocks-Up) – each including a convolution operation, spectral normalization, LeakyReLU activation, and upsampling – to progressively restore the spatial resolution of the feature map from... H / 32× W / Skip connections are used to connect corresponding feature maps from the encoder and decoder levels along the channel dimension. Finally, a 1×1 convolution maps the features to a shape of size 1. H × W ×( N The output tensor is (N+1) channels. The first N channels represent the predicted probability that each pixel belongs to one of N semantic categories, i.e., complete particles, incomplete particles, and the (N+1)th channel indicates the probability that the pixel originates from a synthetic background. If the input is a real image, the classifier is encouraged to assign each pixel to its real semantic category; if the input is a synthetic image, the pixel is encouraged to be assigned to the “false” class represented by the (N+1)th channel.

[0031] Discriminator loss L D Depend on L real and L fake Composed of two components, the discriminator will... L D Minimize and continuously improve the ability to distinguish between real and fake images, thereby providing the generator with more accurate adversarial feedback. L D The definition is as follows: in x Refers to real particle images In step S3 above: a model called YOLOv8-PMQ is proposed based on the YOLOv8 object detection framework. YOLOv8-PMQ introduces a regression branch specifically for predicting particle morphology parameters. By transforming traditional segmentation-based morphological measurements into a regression task, this design eliminates the need for complex segmentation and geometric calculations. Its overall architecture is as follows: Figure 3 As shown, it includes a backbone network, a neck network, and a detection head network.

[0032] The backbone network retains the basic architecture of YOLOv8, including Convolution-Batch Normalization-SiLU (CBS), a cross-stage part (C2f) with two convolutions and feature concatenation, and a Spatial Pyramid Pooling-Fast (SPPF) module. CBS, as the basic backbone module, consists of convolution operations, batch normalization, and the SiLU activation function. It performs basic feature extraction. The C2f module divides the input feature map evenly along the channel dimension. One part is passed directly, while the other part is convolved before concatenation, enabling efficient feature extraction with reduced redundancy. The SPPF module applies max pooling to the feature map using a 5×5 sliding window, extracting the maximum value within each window to emphasize important features without significant computational cost. Input images of size H×W×3 are processed by these modules to extract multi-scale feature maps: C3, C4, and C5, with sizes of H / 8×W / 8×64, H / 16×W / 16×128, and H / 32×W / 32×256, respectively.

[0033] YOLOv8-PMQ adopts the architecture of the PAFPN (Path Aggregation Feature Pyramid Network) neck network. This network aggregates high-level semantic information into lower-level feature maps via a bottom-up path. Simultaneously, it propagates detailed information from lower to higher levels via a top-down path, achieving bidirectional fusion of semantic and spatial information. After PAFPN processing, the neck outputs fused multi-scale feature maps: P3, P4, and P5. These correspond to fused versions of the original C3, C4, and C5 feature maps. Their size is [size missing]. H / 8× W / 8×64, H / 16× W / 16×128, H / 32× W / 32×256.

[0034] The YOLOv8-PMQ probe network extends the original YOLOv8 design, comprising three branches: a classification branch, a bounding box regression branch, and a new particle morphology parameter regression branch. These branches are responsible for particle classification, localization, and morphology estimation, respectively.

[0035] In step S4 above: To further verify the effectiveness of the synthesized image in the detection task and the feasibility of the morphology quantization method based on object detection, three training strategies were designed (see...). Figure 4 ): True training: The model is trained using only real images, which serve as a baseline for evaluating model performance.

[0036] Fake training: The model is trained using only synthetic images. Specifically, labels are input into a trained generator, whose parameters are frozen to prevent further training. Synthetic images (fake images) are then used to train the detection network. This tests the applicability of synthetic images in a standalone training scenario.

[0037] Hybrid Training: The combine harvester undergoes a two-step training process that combines real and synthetic images. The first step uses real images, and the second step adds synthetic images. Considering the potential differences in detail between synthetic and real images, a reduction factor λ (0 < λ < 1) is introduced to adjust the weight of synthetic images in the loss function. The optimal value of λ is selected by searching in increments of 0.1.

[0038] To comprehensively evaluate the performance of the proposed YOLOv8-PMQ model in particle detection and morphological parameter quantification tasks, this study employed three metrics. The first evaluation metric is mean precision (mAP), which is widely used in detection tasks to evaluate overall detection performance. mAP is calculated by first calculating the precision (P) and recall (R) for each class, and then integrating the precision-recall curve to obtain the mean precision (AP). The average of all classes yields the final mAP value. It is defined as follows: Here, TP represents the number of correctly predicted bounding boxes, and FP represents incorrectly predicted bounding boxes, specifically ground truth objects missed by the model. Precision (P) reflects the proportion of correctly predicted objects among all predictions, while recall (R) reflects the proportion of correctly predicted objects among all ground truth objects. p (R) is the precision (P) as a function of recall (R). N Cls This indicates the number of categories. In this model, N Cls = 1.

[0039] The second evaluation metric is frames per second (FPS), which measures the model's processing speed and reflects its real-time performance. FPS is calculated as follows: The third evaluation metric is the Mean Absolute Error (MAE), used to assess the quantitative accuracy of morphological parameters. MAE reflects the average absolute deviation between the predicted values ​​and the true ground values, and is defined as: in, n It is the total number of particles detected by the model. This represents the predicted value of the i-th particle. s i This represents the corresponding ground truth value.

[0040] The generator and validator were trained simultaneously for a total of 1000 epochs. The loss curves for the training and validation phases are as follows: Figure 5 As shown, the generator loss decreases slowly, while the validator loss decreases rapidly. The training and validation loss curves show similar trends, without obvious divergence or abnormal fluctuations, indicating that the training process is relatively stable.

[0041] Table 1 summarizes the final performance of each strategy on the test set. The hybrid training strategy achieved the highest detection accuracy, with a b_mAP of 54.13%, outperforming both the pure real number (53.11%) and pure synthetic (26.72%) training strategies, and also achieving the highest accuracy across all six morphological descriptors (R, S A , S E , d EQPC , d Fmin and d Fmax See Figure 6 The mean absolute error (MAE) of morphological parameter regression is the lowest. total The value is 0.1097. Among them, the size descriptor ( d EQPC , d Fmin , d Fmax The prediction error of ) is significantly smaller than that of the shape descriptor ( R , S A , S E ).

[0042] Table 1 shows the final performance of each strategy on the test set. This invention was compared with two state-of-the-art instance segmentation methods (Mask R-CNN and SOLOv2), and the results are as follows: Table 2 summarizes the model's performance on the test set, including mean accuracy (mAP), inference speed (FPS), and the number of model parameters. For the segmentation model, segmentation mAP (s_mAP) is reported, while for the detection-based YOLOv8-PMQ, bounding box mAP (B_mAP) is used. As the results show, Mask R-CNN and SOLOv2 achieve high segmentation accuracy, with s_mAP values ​​of 79.30% and 82.41%, respectively. However, both models are computationally expensive, with model sizes exceeding 100 M parameters and inference speeds below 2 FPS. Notably, the bottleneck in inference speed mainly comes from post-processing geometric analysis, which takes nearly 0.7 seconds per image.

[0043] In comparison, YOLOv8-PMQ has a lower B_mAP of 54.13%, but requires no additional geometric analysis. It contains far fewer parameters (3.38 M) and achieves significantly higher inference speed (74.63 FPS), making it more efficient and lightweight in real-world deployments.

[0044] Table 2 shows the model's performance data on the test set. Finally, the trained YOLOv8-PMQ model was integrated into the developed Android application. This application supports offline operation. For processing high-resolution images, the application utilizes a built-in sliding window mechanism. Figure 7 Users can customize the window size and sliding step size based on particle size and density. The application also provides a scale calibration function, allowing users to convert pixel dimensions to actual physical dimensions by annotating a reference object with known physical dimensions. Figure 7 The results of field tests conducted at the Xinguang Baitong quarry in Lingshan County, Qinzhou, Guangxi, demonstrate the effectiveness and practicality of this method in actual engineering projects.

[0045] Matters not covered in this invention are common knowledge.

[0046] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A particle morphology quantization method based on an improved YOLOv8 model and augmented generative data, characterized in that, The method includes the following steps: Step S1: Obtain and label particle image samples to form a dataset containing particle boundaries and morphological parameter labels; Step S2: Construct a particle image synthesis module based on generative adversarial network to generate synthesized particle images from the input semantic labels, which are used to expand the training dataset; Step S3: Construct the YOLOv8-PMQ particle detection model. Add a morphological parameter regression branch to the YOLOv8 detection head to simultaneously output the particle detection box and morphological parameters. Step S4: Use the synthetic image generated in step S2 to perform mixed training with the real image to obtain a target detection model with higher morphological parameter regression accuracy; S5. Input the particle image to be analyzed into the trained YOLOv8-PMQ particle detection model, and output the particle bounding box and morphological parameters. S6. Embed the target detection model into a mobile application to achieve rapid on-site particle morphology detection and statistical analysis.

2. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: In step S2, the generative adversarial network adopts a semantic image generation model based on the OASIS structure, which includes a generator and a discriminator. The generator converts the label information into particle images through several convolution operations, batch normalization and activation functions. The discriminator adopts the U-Net structure to distinguish between generated images and real images.

3. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 3, characterized in that: The generator passes the synthesized image to the discriminator network, minimizing the adversarial loss. L adv To update the generator's parameters, the generator loss function. L G The definition is as follows: Where y represents the input label, z is the noise tensor introduced to enhance the diversity of the synthesized image, N is the total number of categories (set to N=3, including complete particles, incomplete particles, and background), G(z, y) represents the image synthesized by the generator given the noise z and the label y, and D(G(z, y)) represents the pixel prediction on the synthesized image, outputting the probability that each pixel belongs to a specific category. y, j, c refers to the pixel at position (i, j) in label y belonging to category c. H × W α represents the resolution of the image. c The weights assigned to address pixel frequency imbalances are calculated using the following formula: 。 4. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: Discriminator loss L D Depend on L real and L fake Composed of two components, L D The definition is as follows: in x It refers to a true image of particles.

5. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: In step S1, the morphological parameters include roundness R and sphericity S. A Length, width, and sphericity S E Equivalent projected circle diameter d EQPC Maximum Freette diameter d Fmax and minimum Freette diameter d Fmin .

6. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: In step S3, the probe network of the YOLOv8-PMQ particle detection model extends the existing YOLOv8 design and includes three branches: a classification branch, a bounding box regression branch, and a new particle morphology parameter regression branch. The three branches are responsible for particle classification, localization, and morphology estimation, respectively.

7. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: In step S3, the total loss function of the YOLOv8-PMQ particle detection model L Detect Classification branch loss L Cls Bounding box regression branch loss L Bbox and morphological parameter regression branch loss L Param The summation is constructed, and its expression is: Where N is the number of samples; 。 8. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: In step S4, the hybrid training adopts a two-stage training strategy. In the first stage, real particle image data is used to train the YOLOv8-PMQ particle detection model. In the second stage, synthetic particle images generated by adversarial networks are introduced for joint training. A reduction factor λ is introduced, where 0 < λ < 1, to adjust the weight of the synthetic images in the loss function. The optimal value of λ is selected by searching in increments of 0.

1.

9. The particle morphology quantization method based on an improved YOLOv8 model and augmented generation data as described in claim 1, characterized in that: Step S6 also includes deploying the trained model on an Android smartphone for offline operation, requiring no cloud computing resources and allowing for on-site use.

Images