Flocculation feature parameter extraction method based on YOLO model deep learning

Abstract

The application discloses a flocculation feature parameter extraction method based on a YOLO model deep learning, which comprises: a collected flocculation movement sequence image; based on the flocculation movement sequence image, a training data set is constructed and a YOLO target detection model is trained; according to the trained YOLO target detection model, the flocculation movement sequence image is automatically detected to obtain a detection result; a Kalman filtering algorithm is used to track and estimate the motion state of the detection result to obtain a motion state estimation result of the flocculation; and based on the motion state estimation result, multi-dimensional feature parameter extraction is carried out to obtain flocculation feature parameters. The application can realize efficient, stable and automatic processing of large-scale image data without interfering with the natural motion state of the flocculation, and significantly improves the accuracy and processing efficiency of the flocculation feature parameter extraction.

Description

Technical Field

[0001] This invention belongs to the field of image processing and hydrological sediment measurement technology, and particularly relates to a method for extracting floc feature parameters based on YOLO model deep learning. Background Technology

[0002] The particle size, settling velocity, and structural characteristics of sediment flocs are important parameters for studying the transport, sedimentation, and flocculation mechanisms of sediment in water bodies. Existing floc measurement methods mostly rely on field or laboratory instruments such as particle size analyzers and flow meters. These methods usually require direct contact with the sediment or water environment, which can easily damage the original structure of the flocs during sampling or measurement. Furthermore, the instruments themselves can disturb the water flow, leading to deviations between the measurement results and actual operating conditions.

[0003] To overcome the aforementioned problems, non-invasive high-speed camera systems have been increasingly applied to the study of floc characteristics. These systems record the movement of flocs through images and then extract relevant parameters using image processing techniques. However, most existing image processing methods based on ImageJ or Matlab require manual target selection, threshold setting, or frame-by-frame matching, resulting in cumbersome processes, low efficiency, and insufficient stability under large-scale image data conditions.

[0004] With the development of deep learning technology, the target detection model's ability to identify targets in complex backgrounds has been significantly improved. However, mature technical solutions for its systematic application in automatic floc identification, motion tracking, and physical parameter inversion are still relatively lacking. Therefore, it is necessary to propose a deep learning-based method for automatic extraction of floc feature parameters to achieve efficient, accurate, and repeatable floc characteristic analysis. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for extracting floc feature parameters based on YOLO model deep learning. This method can achieve efficient, stable, and automated processing of large-scale image data without interfering with the natural motion state of the flocs, significantly improving the accuracy and processing efficiency of floc feature parameter extraction.

[0006] To achieve the above objectives, this invention provides a method for extracting clumping feature parameters based on YOLO model deep learning, comprising: Images of the collected floc movement sequence; Based on the floc movement sequence images, a training dataset is constructed and a YOLO object detection model is trained; The YOLO target detection model, after training, is used to automatically detect the floc movement sequence images and obtain the detection results. The Kalman filter algorithm is used to perform motion state tracking and estimation on the detection results to obtain the motion state estimation results of the flocs. Based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters.

[0007] Optionally, based on the floc movement sequence images, constructing a training dataset and training a YOLO object detection model includes: A training dataset is constructed by manually annotating the floc targets in the floc motion sequence images to acquire floc motion sequence images without direct contact with flocs or water using a non-invasive high-speed camera system. Based on the training dataset, the YOLO object detection algorithm is used to train the model. By iteratively learning the visual features and spatial distribution of flocs in the image, the trained YOLO object detection model is obtained.

[0008] Optionally, training the model using the YOLO object detection algorithm based on the training dataset includes: Based on the training dataset, configure the network parameters of the YOLO object detection model, which adopts a multi-scale anchor box structure. The detection results are screened using non-maximum suppression based on the multi-scale anchor frame structure to improve the identification accuracy of flocs of different sizes.

[0009] Optionally, the floc motion sequence image is automatically detected based on the trained YOLO target detection model, and the detection results include: The YOLO target detection model, after training, is used to automatically detect the floc motion sequence images and obtain the position information and bounding box information of the floc in each frame image. Extract the segmentation mask of each floc based on the bounding box information to obtain the pixel-level contour information of the floc; The detection result is obtained based on the pixel-level contour information of the flocs.

[0010] Optionally, the motion state estimation of the detection results is performed using the Kalman filter algorithm to obtain the motion state estimation results of the floc, including: Based on the location information of the same floc in adjacent frames, the location information is used as the observation value and input into the Kalman filter algorithm; Based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the motion state estimation result of the floc.

[0011] Optionally, based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the motion state estimation results of the floc, including: According to the prediction steps of the Kalman filter algorithm, the prior state estimate and prior covariance matrix at the current time are obtained based on the posterior state estimate and state transition matrix of the previous time step. According to the observation update steps of the Kalman filter algorithm, the Kalman gain is calculated based on the current observation value, observation matrix, and measurement noise covariance matrix. The posterior state estimate and posterior covariance matrix are then updated based on the Kalman gain to obtain the optimal spatial position and velocity estimate of the floc at the current time.

[0012] Optionally, the floc characteristic parameters include: two-dimensional morphological parameters, kinematic parameters, and physical characteristic parameters of the floc.

[0013] Optionally, based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters, including: Based on the floc segmentation mask corresponding to the bounding box information, the two-dimensional morphological parameters of the floc are calculated, wherein the two-dimensional morphological parameters include one or more of the floc's area, perimeter, equivalent diameter, and roundness. The settling velocity of the floc is calculated based on the motion state estimation results and the image acquisition time interval, and the physical characteristic parameters of the floc are obtained by inversion using a modified Stokes formula. The physical characteristic parameters include one or more of the effective density and fractal dimension of the floc. The motion parameters are calculated by measuring the change in the vertical position of the same particle in two images taken at a preset time interval.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: 1. Target Recognition Level: Deep learning improves the accuracy of complex floc recognition. Based on YOLOv8 multi-scale anchor boxes + non-maximum suppression, it is suitable for flocs with significant size differences and has strong robustness in recognizing overlapping and blurred flocs. The model is trained on labeled datasets and automatically outputs bounding boxes and pixel-level segmentation masks, eliminating the need for manual thresholding and avoiding contour extraction errors caused by manual binarization. The false negative and false positive rates in complex multi-floc scenarios are much lower than those of traditional image algorithms.

[0015] 2. Motion Tracking Level: Kalman filtering smooths the trajectory, resulting in more accurate motion parameters. A state vector containing two-dimensional position and velocity is constructed. Prediction-update recursive operations suppress inter-frame jitter and image noise detected by YOLO. The system automatically matches the same floc across frames, outputting a continuous and smooth motion trajectory, eliminating the need for manual frame-by-frame matching. Vertical sink velocity is automatically calculated based on the smooth trajectory and frame interval, significantly reducing velocity calculation errors.

[0016] 3. Fully automated process, integrated output of multi-dimensional feature parameters. A single process synchronously outputs morphological parameters (area, perimeter, equivalent diameter, roundness, etc.), motion parameters (trajectory coordinates, settling velocity), and physical feature parameters (effective density, fractal dimension); relying on Stokes' theorem for automatic inversion, eliminating manual step-by-step conversion, and achieving full automation of "raw image input → structured parameter CSV file + visualization result output".

[0017] 4. Objective standardization eliminates subjective bias. The entire process involves no human intervention in threshold adjustment, target selection, or profile correction. Parameter calculation rules are fixed and uniform, ensuring high comparability of data from different batches and experimental groups, significantly improving experimental reproducibility. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the floc feature parameter extraction method based on YOLO model deep learning according to an embodiment of the present invention; Figure 2 This is a flowchart of the manual binarization method based on ImageJ according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the effect of the embodiment of the present invention compared with the traditional ImageJ manual processing method. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] This embodiment proposes a method for extracting floc feature parameters based on deep learning of the YOLO model, such as... Figure 1 As shown, the specific steps include: Images of the collected floc movement sequence; Based on the floc movement sequence images, a training dataset is constructed and a YOLO object detection model is trained; The YOLO target detection model, after training, is used to automatically detect the floc movement sequence images and obtain the detection results. The Kalman filter algorithm is used to perform motion state tracking and estimation on the detection results to obtain the motion state estimation results of the flocs. Based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters.

[0022] Specifically, this invention provides a method for extracting floc feature parameters based on YOLO model deep learning, aiming to solve the problems of high invasiveness, low efficiency, and poor accuracy in traditional sediment measurement methods. This method integrates non-invasive image acquisition, deep learning target detection, motion state estimation, and multi-parameter analysis to achieve fully automated processing from raw images to feature parameters. Specifically, it includes the following five core steps: Step 1: Image Acquisition and Annotation Dataset Construction: A non-invasive high-speed camera system was used to continuously capture images of the natural motion sequence of sediment flocs, ensuring the original data was authentic and lossless. Subsequently, professional annotation tools (such as Roboflow) were used to manually and meticulously annotate the floc targets in the images, selecting targets and assigning them category labels to form a structured training dataset. This dataset, after being partitioned (e.g., training and validation sets) and preprocessed (e.g., size normalization, data augmentation), was used for subsequent model training.

[0023] Furthermore, based on the floc movement sequence images, constructing a training dataset and training a YOLO object detection model includes: A training dataset is constructed by manually annotating the floc targets in the floc motion sequence images to acquire floc motion sequence images without direct contact with flocs or water using a non-invasive high-speed camera system. Based on the training dataset, the YOLO object detection algorithm is used to train the model. By iteratively learning the visual features and spatial distribution of flocs in the image, the trained YOLO object detection model is obtained.

[0024] Furthermore, based on the aforementioned training dataset, model training using the YOLO object detection algorithm includes: Based on the training dataset, configure the network parameters of the YOLO object detection model, which adopts a multi-scale anchor box structure. The detection results are screened using non-maximum suppression based on the multi-scale anchor frame structure to improve the identification accuracy of flocs of different sizes.

[0025] Specifically, step 2, YOLO deep learning model training and optimization: Based on the dataset constructed in step 1, the YOLO object detection algorithm was used to train the model. Deep learning frameworks (such as Ultralytics) were used to configure the network parameters, and the visual features and spatial distribution of flocs in images were learned through multiple iterations. During training, the loss function and evaluation metrics (such as precision and recall) were monitored, and the model performance was optimized using a validation set, ultimately resulting in a highly robust detection model capable of accurately identifying and locating flocs.

[0026] Furthermore, the floc motion sequence image is automatically detected based on the trained YOLO target detection model, and the detection results include: The YOLO target detection model, after training, is used to automatically detect the floc motion sequence images and obtain the position information and bounding box information of the floc in each frame image. Extract the segmentation mask of each floc based on the bounding box information to obtain the pixel-level contour information of the floc; The detection result is obtained based on the pixel-level contour information of the flocs.

[0027] Specifically, step 3 involves automatic identification, location, and segmentation mask extraction of flocs. The trained YOLO model was integrated into an image processing program to batch process a large number of experimental image sequences. The model automatically performs floc detection, bounding box localization, and category determination for each frame. Furthermore, the program extracts a segmentation mask for each floc based on the detection results, providing accurate pixel-level contour information for subsequent quantitative calculation of morphological parameters.

[0028] Furthermore, the motion state estimation of the detection results is performed using the Kalman filter algorithm, and the motion state estimation results of the floc include: Based on the location information of the same floc in adjacent frames, the location information is used as the observation value and input into the Kalman filter algorithm; Based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the motion state estimation result of the floc.

[0029] Furthermore, based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the floc's motion state estimation results, including: According to the prediction steps of the Kalman filter algorithm, the prior state estimate and prior covariance matrix at the current time are obtained based on the posterior state estimate and state transition matrix of the previous time step. According to the observation update steps of the Kalman filter algorithm, the Kalman gain is calculated based on the current observation value, observation matrix, and measurement noise covariance matrix. The posterior state estimate and posterior covariance matrix are then updated based on the Kalman gain to obtain the optimal spatial position and velocity estimate of the floc at the current time.

[0030] Specifically, step 4, tracking and estimating the motion state of the flocs based on Kalman filtering: To obtain accurate motion parameters, a Kalman filter algorithm is introduced to perform cross-frame tracking and state smoothing on the flocs identified in step 3. Using the position information of the flocs between previous and subsequent frames as observations, the optimal motion state (including position and velocity) of each floc is estimated through the prediction and update recursive steps of the Kalman filter, effectively suppressing detection jitter and noise, and forming a continuous and stable motion trajectory.

[0031] Furthermore, the floc characteristic parameters include: two-dimensional morphological parameters, kinematic parameters, and physical characteristic parameters of the floc.

[0032] Furthermore, based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters, including: Based on the floc segmentation mask corresponding to the bounding box information, the two-dimensional morphological parameters of the floc are calculated, wherein the two-dimensional morphological parameters include one or more of the floc's area, perimeter, equivalent diameter, and roundness. The settling velocity of the floc is calculated based on the motion state estimation results and the image acquisition time interval, and the physical characteristic parameters of the floc are obtained by inversion using a modified Stokes formula. The physical characteristic parameters include one or more of the effective density and fractal dimension of the floc. The motion parameters are calculated by measuring the change in the vertical position of the same particle in two images taken at a preset time interval. The motion parameters are settling velocity parameters.

[0033] Specifically, step 5: Calculation of multi-dimensional feature parameters and visualization of results. Combining the segmentation mask from step 3 with the motion state estimation results from step 4, multi-dimensional feature parameters are extracted. Two-dimensional morphological parameters of the floc (such as area, perimeter, equivalent diameter, and roundness) are calculated using image processing libraries (e.g., OpenCV). Sinking velocity is calculated based on the motion trajectory and time interval, and physical parameters such as effective density and fractal dimension are derived using Stokes' theorem. Finally, the program automatically generates a visualized image with bounding boxes, trajectory lines, and parameter labels, and outputs a structured parameter data file, completing the entire extraction process from image to quantitative features.

[0034] This method enables simultaneous, non-destructive, and batch measurement of key characteristics such as particle size and settling velocity of sediment flocs, significantly improving measurement efficiency and accuracy, and providing an advanced technical tool for sediment kinematics research.

[0035] The following is a detailed description of this embodiment, such as... Figure 1 As shown, it specifically includes: Step 1: Acquisition of floc motion sequence images and construction of training dataset: First, a non-invasive high-speed camera observation system was built and aimed at the area of ​​sediment floc movement. Camera parameters were set (20fps frame rate, 1280×860 resolution), and under stable lighting conditions, continuous image sequences of floc movement were acquired for different experimental groups, ensuring image clarity and target visibility. Then, using the professional image annotation tool Roboflow, the acquired images were opened frame by frame, and rectangular bounding boxes were manually drawn for each floc target, uniformly labeled as "floc". All annotation results were exported in YOLO format (each image corresponds to a .txt file, containing the target category number and normalized bounding box coordinates). All labeled data were randomly divided according to a preset ratio (80% training set, 20% validation set), and the image data underwent uniform size normalization (adjusted to 640×640 pixels) and data augmentation processing (such as random rotation and brightness adjustment) to construct a standard dataset for model training.

[0036] Step 2: Training and validation of the YOLO-based floc detection model: The YOLO model (YOLOv8) under the Ultralytics framework was selected as the detection network. The training and validation sets prepared in step 1 were loaded, and key model training parameters were configured: initial weights were selected using a pre-trained model (e.g., yolov8n.pt), and the learning rate (e.g., 0.01), batch size (e.g., 16), and total training epochs (e.g., 100) were set. The training process was initiated, and the model continuously learned the visual features of the flocs through forward and backward propagation. During training, the changes in the loss function on the training and validation sets were monitored in real time, and metrics such as precision, recall, and mean average precision (mAP) were evaluated. When the model's performance on the validation set stabilized and reached a preset standard (e.g., mAP@0.5>0.9), training was terminated, and the optimal model weight file (e.g., best.pt) was saved for subsequent recognition.

[0037] Step 3: Automatic identification, location, and segmentation mask extraction of flocs: Deploy the YOLO model trained in step 2 in a Python environment. Write a batch processing program to sequentially read all sequence images for one experimental group. For each input image, call the model to perform inference. The model outputs the confidence score, bounding box coordinates, and corresponding instance segmentation mask for each detected floc target. The program filters out false positives with low confidence based on a confidence threshold (set to 0.5), retaining reliable detection results. This step ultimately outputs the unique ID, bounding box information, and binarized segmentation mask image for all flocs in each frame, providing basic data for subsequent morphological analysis and cross-frame tracking.

[0038] Step 4: Tracking the motion state of flocs and smoothing their trajectory based on Kalman filtering: The application of Kalman filtering is based on the following three assumptions: the current state depends only on the previous state; the system state evolution and the observation process both satisfy a linear relationship; and the introduced noise follows a Gaussian distribution. Based on this, two formulas characterizing the system state and the observation are introduced as follows: (1); (2); Equation (1) is the system state equation, and equation (2) is the observation equation, where, Let be the state vector (such as position and velocity) at time (k). This is the state transition matrix, which describes the relationship between the system state and time. To control the degree of influence of input on the state (control matrix). To represent the control input at time k-1, The noise is a process noise that follows a Gaussian distribution. Let be the observation vector at time (k) (e.g., the detected position). The observation matrix maps the state space to the observation space. The noise is observed to follow a Gaussian distribution.

[0039] The filtering process includes: ① Prediction steps: Equations (3) and (4) are the prior state estimation and prior covariance update, respectively. The former predicts the current state through the state transition model (without observation update), while the latter quantifies the prediction uncertainty through covariance propagation and superimposes the influence of process noise Q.

[0040] (3); (4); in, Let k be the prior state estimate at time k. For the posterior state estimate at time k-1, Let be the prior covariance matrix at time k. Let be the posterior covariance matrix at time k-1. Let be the process noise covariance matrix.

[0041] ②The observation update steps include: For the observation equation, by using the assumptions of unbiased estimation and the definition of error, we obtain the optimal estimate (5), the Kalman gain K (6), and the posterior error covariance matrix update equation (7): (5); (6); (7); in, To measure the noise covariance matrix, It is an identity matrix with the same dimension as the system state.

[0042] ③ Recursive execution: The above prediction and update steps are performed recursively at each time step to achieve continuous optimization estimation of the target state.

[0043] In this embodiment, Kalman filtering is used to perform temporal smoothing estimation of the spatial position and velocity of each target based on YOLO detection results, thereby improving the accuracy and robustness of motion parameters. The target state vector is set as a two-dimensional position and velocity (e.g., ([x,y,v_x,v_y])), and the state transition matrix and observation matrix reflect the target's physical motion model and the information obtainable from the detection results, respectively. Using the target position detected in each frame as the observation input, Kalman filtering performs dynamic smoothing estimation of position and velocity, effectively suppressing jitter caused by detection errors and measurement noise, and improving the accuracy of motion analysis.

[0044] Step 5, calculation of floc morphology and physical parameters, and visualization of results, includes: Based on the segmentation mask obtained in step 3 and the smoothing trajectory obtained in step 4, the feature parameters of each floc in each frame are calculated.

[0045] Morphological parameter calculation: Call the OpenCV library function to calculate the projected area A and perimeter P of the floc based on the contour of the segmentation mask, and then calculate the equivalent diameter and roundness (8) and (9).

[0046] Physical parameter calculation: Based on the time-series data of the floc vertical coordinate y obtained in step 4, combined with the known inter-frame time interval Δt (0.05s), the settling velocity is calculated using equation (10). Furthermore, based on a modified version of Stokes' formula, combined with the known dynamic viscosity μ and gravitational acceleration g, the effective density Δρ of the floc is calculated using equation (11). Finally, the fractal dimension N is estimated using the relationship between effective density and particle size. f Equation (12).

[0047] (8); (9); (10); (11); (12); In the formula, μ is the dynamic viscosity of water (1.005 × 10⁻⁶). -3 Pa·s), g is the acceleration due to gravity (9.8 m / s²). 2 ), ρ s , ρ w ρ represents the density of the dispersed particles and water, respectively, and d represents the particle size of the dispersed particles.

[0048] During program execution, a visualization window is displayed in real time, drawing ID-bearing bounding boxes and displaying historical motion trajectory lines on the original image. All calculation results (morphological parameters, sinking velocity, trajectory coordinates) are stored in structured files (CSV format) according to floc ID and frame number. After processing, the program automatically saves all visualization images and corresponding data files, completing the entire extraction process from the original image to quantitative feature parameters.

[0049] Verification of the improvement effect in calculation: To objectively evaluate the accuracy and superiority of this embodiment (an automatic extraction method based on YOLO), it is compared with traditional classical methods (manual binarization methods based on ImageJ). Figure 2 The data used for verification were derived from the 24 experimental groups involved in this study. 500 typical original images containing different floc concentrations, particle size ranges and motion states were randomly selected as the test set.

[0050] To further evaluate the reliability of this embodiment from a statistical perspective, the error distribution of the results obtained by the two methods was calculated for all floc particle size data (a total of 513 valid flocs) extracted from 500 images in the test set, as shown below. Figure 3As shown in the figure, the relative error of most data points is within ±5% on average, with a mean absolute error (MAE) of 3.1 μm and a root mean square error (RMSE) of 3.8 μm. The particle size data extracted by both methods show a high linear correlation, with a coefficient of determination (R²) of 0.97. This fully demonstrates that the automatic extraction method in this embodiment has reached a level of measurement accuracy that can replace manual analysis. Furthermore, a quantitative comparison of processing efficiency was made: processing the same test set (500 images, totaling 513 flocs), skilled researchers using ImageJ for manual threshold adjustment, segmentation checks, and data export took approximately 4 hours. However, using the fully automated processing of this embodiment, including model inference, tracking, and parameter calculation, took only about 3 minutes on the same computing platform, representing an efficiency improvement of over 80 times. In addition, this embodiment avoids subjective fluctuations caused by manual operation, ensuring a high degree of reproducibility of the results. Compared to traditional manual ImageJ methods, the YOLO-based automatic extraction method for floc feature parameters proposed in this invention achieves an order-of-magnitude leap in processing efficiency while maintaining measurement accuracy (average error <5%, R²>0.96). It also completely eliminates human subjectivity, achieving standardization and objectification of results. This verifies the comprehensive improvements in accuracy, efficiency, and reliability of this embodiment, providing a powerful high-throughput analysis tool for the study of sediment floc dynamics.

[0051] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for extracting clumping feature parameters based on YOLO model deep learning, characterized in that, include: Images of the collected floc movement sequence; Based on the floc movement sequence images, a training dataset is constructed and a YOLO object detection model is trained; The YOLO target detection model, after training, is used to automatically detect the floc movement sequence images and obtain the detection results. The Kalman filter algorithm is used to perform motion state tracking and estimation on the detection results to obtain the motion state estimation results of the flocs. Based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters.

2. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 1, characterized in that, Based on the aforementioned floc movement sequence images, constructing a training dataset and training a YOLO object detection model includes: A training dataset is constructed by manually annotating the floc targets in the floc motion sequence images to acquire floc motion sequence images without direct contact with flocs or water using a non-invasive high-speed camera system. Based on the training dataset, the YOLO object detection algorithm is used to train the model. By iteratively learning the visual features and spatial distribution of flocs in the image, the trained YOLO object detection model is obtained.

3. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 2, characterized in that, Based on the training dataset, model training using the YOLO object detection algorithm includes: Based on the training dataset, configure the network parameters of the YOLO object detection model, which adopts a multi-scale anchor box structure. The detection results are screened using non-maximum suppression based on the multi-scale anchor frame structure to improve the identification accuracy of flocs of different sizes.

4. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 1, characterized in that, The YOLO target detection model, after training, is used to automatically detect the floc motion sequence images, and the detection results include: The YOLO target detection model, after training, is used to automatically detect the floc motion sequence images and obtain the position information and bounding box information of the floc in each frame image. Extract the segmentation mask of each floc based on the bounding box information to obtain the pixel-level contour information of the floc; The detection result is obtained based on the pixel-level contour information of the flocs.

5. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 4, characterized in that, The motion state estimation of the flocs is obtained by using the Kalman filter algorithm to perform motion state tracking estimation on the detection results, including: Based on the location information of the same floc in adjacent frames, the location information is used as the observation value and input into the Kalman filter algorithm; Based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the motion state estimation result of the floc.

6. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 5, characterized in that, Based on the prediction and observation update steps of the Kalman filter algorithm, the two-dimensional spatial position and corresponding two-dimensional velocity components of the floc are recursively estimated to obtain the motion state estimation results of the floc, including: According to the prediction steps of the Kalman filter algorithm, the prior state estimate and prior covariance matrix at the current time are obtained based on the posterior state estimate and state transition matrix of the previous time step. According to the observation update steps of the Kalman filter algorithm, the Kalman gain is calculated based on the current observation value, observation matrix, and measurement noise covariance matrix. The posterior state estimate and posterior covariance matrix are then updated based on the Kalman gain to obtain the optimal spatial position and velocity estimate of the floc at the current time.

7. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 4, characterized in that, The floc characteristic parameters include: two-dimensional morphological parameters, kinematic parameters, and physical characteristic parameters of the floc.

8. The method for extracting floc feature parameters based on YOLO model deep learning according to claim 7, characterized in that, Based on the motion state estimation results, multi-dimensional feature parameters are extracted to obtain the floc feature parameters, including: Based on the floc segmentation mask corresponding to the bounding box information, the two-dimensional morphological parameters of the floc are calculated, wherein the two-dimensional morphological parameters include one or more of the floc's area, perimeter, equivalent diameter, and roundness. The settling velocity of the floc is calculated based on the motion state estimation results and the image acquisition time interval, and the physical characteristic parameters of the floc are obtained by inversion using a modified Stokes formula. The physical characteristic parameters include one or more of the effective density and fractal dimension of the floc. The motion parameters are calculated by measuring the change in the vertical position of the same particle in two images taken at a preset time interval.

Images