A method and system for microparticle measurement based on bounding box interaction

Abstract

This application relates to the field of micron-particle image detection technology, and discloses a method and system for measuring micron-particles based on interactive detection boxes. This method introduces a Dynamic Multi-Dilatation Fusion (DMDF-Block) module and an Adaptive Cross-Scale Feedback Fusion (ACF-Fusion) module into a CNN object detection model, enhancing the model's feature extraction and cross-scale fusion capabilities for small particle targets in microscopic images. Soft cue prior maps are generated from the predicted bounding boxes and confidence scores output during the detection phase, guiding UNet++ to perform interactive segmentation and thus improving segmentation accuracy. Simultaneously, a sparse sliding window inference strategy is employed to reduce redundant inference and improve segmentation efficiency. This method can significantly improve the accuracy and stability of small target particle detection, segmentation, and measurement under complex backgrounds and multiple interference conditions.

Description

Technical Field

[0001] This invention relates to the field of micron particle image detection technology, and in particular to a method and system for measuring micron particles based on detection frame interaction. Background Technology

[0002] Micron-sized particles are widely present in nature and various industrial production processes. Their particle size and morphology significantly influence particle motion, reactivity, and the quality of final products. In materials science, environmental monitoring, and biomedicine, particle size and shape are key indicators for evaluating performance and application potential. For example, in the field of power batteries, ternary precursor particles, as the core raw material for cathode materials, directly affect the specific surface area, charge-discharge performance, and cycle stability of the material due to their particle size distribution and morphology. In biomedicine, precise control over the size and shape of micron-sized particles can lead to the development of drug carriers with targeted release capabilities. In environmental monitoring, the detection of micron-sized particles is a crucial basis for air quality assessment and pollution source tracing.

[0003] With the increasing demands of applications, various industries are placing higher requirements on the accuracy and efficiency of particle size and shape detection for micron-sized particles. Currently, commonly used detection methods mainly include laser particle size analyzers and image-based detection. Laser particle size analyzers can achieve rapid measurement and are suitable for analyzing the particle size distribution of large batches of samples, but they are difficult to provide information on particle morphology. Image-based detection, on the other hand, can simultaneously acquire particle size and shape characteristics, providing a more intuitive reflection of the particle's microstructure. However, traditional image detection typically relies on microscopy or scanning electron microscopy imaging, and analysis is performed manually or semi-automatically. This process is cumbersome, inefficient, and highly subjective, making it difficult to meet the demands of automated and high-throughput detection.

[0004] In recent years, image-based automated detection methods have gradually become an important research direction in the field of micron-sized particle detection. Early image detection methods relied heavily on classic algorithms such as edge detection and threshold segmentation. While simple to implement, these methods lacked robustness and accuracy in situations involving complex backgrounds, particle aggregation, or drastic changes in particle morphology. The rapid development of computer vision and deep learning technologies has provided new solutions for the efficient online analysis of micron-sized particles through intelligent detection algorithms. Among these, deep learning models based on target detection and image segmentation can automatically learn and extract particle boundary, morphology, and location information through large-scale data training, thereby achieving accurate identification and quantitative analysis of particle size and morphology.

[0005] However, these methods still face many challenges in practical applications. On the one hand, existing image segmentation models often ignore the differences in structural features between clustered and non-clustered particles, which can easily lead to problems such as blurred boundaries, particle adhesion, or missegmentation when processing high-density images or images with clustering. On the other hand, direct full-image segmentation is not only computationally complex but also susceptible to background interference, especially when processing small target particles, which can easily lead to missed detections and false detections, thus affecting the accuracy of particle count and morphological parameter extraction.

[0006] It is evident that existing methods for detecting micron-sized particles have shortcomings in terms of particle aggregation, small target recognition, and segmentation accuracy. Summary of the Invention

[0007] This invention provides a method and system for measuring micron-sized particles based on detection frame interaction, in order to solve the problems that are difficult to solve in the prior art.

[0008] In a first aspect, this application provides a method for measuring micron-sized particles based on detection frame interaction, including:

[0009] S1: Collect images of micron-sized particles under multiple backgrounds and interference types, construct an image dataset of particles in multiple scenarios, and train an improved CNN object detection network model. S2: Input the set of images to be inferred into the improved CNN object detection network model to obtain the detection box information of small target particles; S3: Generate a soft cue prior map based on the detection box and concatenate it with the image to be inferred to form a four-channel input feature. Select the intersecting region from the four-channel input feature through a sparse sliding window to perform image segmentation and obtain the sliding window result. Output a binarized segmentation mask image based on the sliding window result. S4: Calculate the target parameters of the binarized mask image to obtain the measurement results.

[0010] Secondly, this application also provides a micron particle measurement system based on detection frame interaction, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0011] The present invention has the following beneficial effects: The micron-sized particle measurement method based on detection box interaction proposed in this application effectively enhances the model's feature extraction and cross-scale fusion capabilities for small particle targets in microscopic images by introducing a dynamic multi-expansion fusion module (DMDF-Block) and an adaptive cross-scale feedback fusion module (ACF-Fusion) into the CNN target detection model. By adaptively adjusting the receptive field and optimizing feature interaction, it can significantly improve the detection accuracy and stability of small target particles under complex backgrounds and multiple interference conditions.

[0012] In a further technical solution, the detection boxes and their confidence scores output during the detection phase are used to generate a soft prior map, guiding the segmentation network to perform interactive constraints. By combining the soft boundary prior with the center heatmap and adaptive weighting of confidence scores, high-confidence target constraints are enhanced, low-confidence interference is suppressed, missegmentation and missed segmentation are reduced, and boundary consistency is improved. Simultaneously, sparse sliding window inference is employed, performing forward computation only on sliding window blocks intersecting with the detection box regions, reducing redundant background inference and significantly improving inference efficiency while maintaining overall image output consistency.

[0013] In addition to the objectives, features and advantages described above, the present invention has other objectives, features and advantages.

[0014] The present invention will now be described in further detail with reference to the figures. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a preferred embodiment of the present invention for measuring micron-sized particles based on detection frame interaction; Figure 2 The following is a schematic diagram of a microscopic image of a multi-background particle dataset according to a preferred embodiment of the present invention, wherein (a) is a schematic diagram of aggregation, (b) is a schematic diagram of impurities, (c) is a schematic diagram of defocus blur, (d) is a schematic diagram of bubbles, (e) is a schematic diagram of backgrounds with different color gains, and (f) is a schematic diagram of motion blur. Figure 3 This is a schematic diagram of the overall architecture of the improved CNN object detection network according to a preferred embodiment of the present invention; Figure 4 This is a flowchart illustrating the feature processing of the dynamic multi-expansion fusion module in the improved Backbone structure of a preferred embodiment of the present invention. Figure 5 This is a flowchart of the feature processing of the adaptive cross-scale feedback fusion module in the Neck structure of a preferred embodiment of the present invention; Figure 6 The diagram shows a comparison of the confusion matrices of the CNN object detection network before and after the improvement according to a preferred embodiment of the present invention. (a) is the confusion matrix of the original YOLOv8 test set, and (b) is the confusion matrix of the improved model test set. Figure 7 This is a schematic diagram of the SoftPrompt+ sparse sliding window interaction process based on Unet++ in a preferred embodiment of the present invention. Figure 8This is a schematic diagram of the non-agglomerated particle segmentation effect of a preferred embodiment of the present invention, wherein (a) is a comparison diagram of the original image and the segmentation mask superimposed, and (b) is a binarized segmentation mask image; Figure 9 This is a particle size distribution diagram of the non-agglomerated particles of the ternary precursor in a preferred embodiment of the present invention; Figure 10 The diagram shows the particle shape parameter distribution of the non-agglomerated ternary precursor particles according to a preferred embodiment of the present invention, wherein (a) is the frequency distribution curve of roundness R, (b) is the frequency distribution curve of aspect ratio AR, (c) is the frequency distribution curve of ellipticity E, and (d) is the frequency distribution curve of irregularity IR. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms, do not indicate a quantity limitation, but rather indicate the presence of at least one.

[0018] It should be understood that existing micron-sized particle image detection methods have shortcomings in terms of particle aggregation, small target recognition, and segmentation accuracy. Based on this, this application provides a micron-sized particle measurement method based on detection box interaction. This method can achieve high-precision and automated acquisition of micron-sized particle size and shape information in complex scenarios, thereby balancing detection accuracy, robustness, and efficiency, and providing a reliable technical means for quantitative analysis of particles in complex scenarios.

[0019] Please see Figure 1 This application provides a method for measuring micron-sized particles based on detection frame interaction, including: S1: Collect images of micron-sized particles under multiple backgrounds and interference types, construct an image dataset of particles in multiple scenarios, and train an improved CNN object detection network model. S2: Input the set of images to be inferred into the improved CNN object detection network model to obtain the detection box information of small target particles; S3: Generate a soft cue prior map based on the detection box and concatenate it with the image to be inferred to form a four-channel input feature. Select the intersecting region from the four-channel input feature through a sparse sliding window to perform image segmentation and obtain the sliding window result. Output a binarized segmentation mask image based on the sliding window result. S4: Calculate the target parameters of the binarized mask image to obtain the measurement results.

[0020] The aforementioned micron-sized particle measurement method based on detection box interaction effectively enhances the model's feature extraction and cross-scale fusion capabilities for small particle targets in microscopic images by introducing a dynamic multi-expansion fusion module (DMDF-Block) and an adaptive cross-scale feedback fusion module (ACF-Fusion) into the CNN target detection model. By adaptively adjusting the receptive field and optimizing feature interaction, it can significantly improve the detection accuracy and stability of small target particles under complex backgrounds and multiple interference conditions.

[0021] First, it is worth noting that this application is based on the following technical system and materials, using the Motic B1 223 microscope as the core hardware foundation. Three stepper motors are configured for precise drive, enabling accurate movement of the stage in the X and Y axes and controlling the microscope's focal length adjustment within the micrometer range. An integrated camera and image acquisition card complete image capture and data transmission, along with optical components such as a light source, objective lens, eyepiece, and condenser. The experimental sample is a ternary precursor particle prepared using a specific process, dispersed with 0.05% sodium hexametaphosphate reagent.

[0022] The steps of the above-described method for measuring micron-sized particles based on detection box interaction are described below with a complete example: Step 1: Acquisition of microscopic images of micro and nanoparticles.

[0023] To address practical application scenarios such as agglomeration, bubbles, impurities, defocusing, motion blur, and different color gains, we collect images of micron-sized particles with multiple backgrounds and interference types, and construct a multi-scenario particle dataset.

[0024] Step 2: Small target particle detection.

[0025] Based on a CNN object detection network, this paper optimizes the model structure to address the challenges of small particle size and complex backgrounds in microscopic images. A Dynamic Multi-Dilation Fusion Module (DMDF-Block) is introduced into the backbone network. This module dynamically adjusts the receptive field through multi-dilation convolutional branches and an adaptive weight selection mechanism, enhancing the ability of shallow features to represent the edges and details of small particles. The computation process of the DMDF-Block is as follows: Input features The parallel inputs flow into multiple dilated convolutional branches and dynamic dilation selectors, and after weighted fusion, channel enhancement and residual connection, the output features are finally fed into the Concat structure to participate in the feature fusion process of the Neck part.

[0026] The set of multi-dilated convolution branches is B is the number of branches. For the first The convolution dilation rate of the nth branch, the nth The output of each branch for the input features is:

[0027] In the above formula Indicates the expansion rate 3×3 convolution, For the first Batch normalization operation for each branch The SiLU activation function is used. This represents the nonlinear mapping of the i-th dilated convolution branch, which is formed by the concatenation of the above three components.

[0028] The dynamic expansion selector generates branch weights based on global context information, enabling adaptive adjustment of the contribution of each expansion branch. First, it processes the input... Perform global average pooling:

[0029] In the formula, s represents the input features The channel description vector is obtained by performing global average pooling in the spatial dimension. , The number of channels for X; Then, the intermediate feature vector z is obtained by first performing a 1×1 convolution and SiLU activation, and then a 1×1 convolution is performed followed by Softmax normalization in the branch dimension to generate the branch weights:

[0030] in , These are the weight matrices corresponding to a 1×1 convolution. For each expansion branch, the adaptive weights are... This is the intermediate feature vector after SiLU activation.

[0031] Next, we will utilize weights. The local feature maps output by each branch are dynamically weighted and summed to obtain multi-scale fused features. :

[0032] Fusion features Perform global average pooling, and then obtain channel attention by sequentially applying two 1×1 convolutions followed by a SiLU + 1×1 convolution and a Sigmoid function: , ,

[0033] in, This is the channel description vector obtained by global average pooling of the fused features in the spatial dimension. The intermediate channel features are those activated by SiLU. It is the Sigmoid activation function. The attention map is obtained by using 1×1 convolution parameters. The characteristics after channel enhancement are:

[0034] In the formula, This represents element-wise multiplication. right Perform 1×1 convolution and BatchNorm fusion compression:

[0035] In the formula, This indicates that the convolution output is normalized using BatchNorm. The weights are the parameters for a 1×1 convolution. Finally, residual connections are used to obtain the module output:

[0036] Simultaneously, an Adaptive Cross-Scale Feedback Fusion (ACF-Fusion) module is introduced in the feature fusion section. This module employs multi-scale convolution and channel attention feedback mechanisms to achieve adaptive fusion of high- and low-level features, improving cross-scale information transmission and detection accuracy. The computation process of the ACF-Fusion adaptive cross-scale feature fusion module is as follows: Input features are first processed through 1×1 and 3×3 convolutional branches to extract local details and contextual information. Then, feature reweighting and enhancement are completed through channel attention and residual connections to obtain a fused feature map used for the external detection head structure.

[0037] Let the input features be , The number of input channels is denoted as and the number of output channels of the module is denoted as . H and W represent the spatial dimensions of the fused feature map in the vertical and horizontal directions, respectively. The module extracts features through two parallel 1×1 and 3×3 convolutional branches, and the output is:

[0038] The two branches are added element-wise along the channel dimension to obtain the initial multi-scale fused features. :

[0039] In fusion features A lightweight channel attention mechanism is introduced, which performs global average pooling on each channel to obtain a channel-level global response vector. ,in:

[0040] In the formula, For fusion features No. Coordinates in each channel The feature values ​​at the location, i=1,…,H, j=1,…,W, represent the pixel indices of the feature map in the vertical and horizontal directions, respectively; Subsequently, two 1×1 convolutional layers with SiLU and Sigmoid activations are used to generate channel attention weight vectors. :

[0041] in The SiLU activation function is used. For the Sigmoid function, , These are the weight matrices corresponding to the two 1×1 convolutional layers. This represents the channel compression ratio.

[0042] Using channel attention weight vectors Fusion features Perform channel-by-channel reweighting to obtain the weighted feature map. ; For fusion features The Each channel, its weighted eigenvalue The calculation method is as follows:

[0043] in For the first Attention coefficients for each channel For fusion features No. Coordinates in each channel The eigenvalue at that location.

[0044] The weighted features are then subjected to 1×1 convolution and BatchNorm for channel compression and fusion to obtain intermediate features:

[0045] A residual connection structure is added to the output. When the number of input and output channels is the same, the input features are added to the intermediate features; when they are not the same, a 1×1 convolution is used to align the number of input channels first, and the residual term is then processed. Represented as:

[0046] Final output features The output features are obtained by residual superposition followed by SiLU nonlinear activation. :

[0047] The output feature map The input feature map is used as the target detection head. Based on the detection head parameters, the category of micron-sized particles and the coordinates of the detection box are regressed and classified, and this information is provided to the segmentation model.

[0048] The dataset is labeled with the locations and labels of clustered and non-clustered particles, and the improved detection network described above is trained. For the detected image, the output is the coordinates (x1, y1, x2, y2) of the detection box for small target particles in YOLO format. In this application, the improved detection network is as follows: Figure 2 As shown, (a) is a schematic diagram of agglomeration, (b) is a schematic diagram of impurities, (c) is a schematic diagram of defocusing and blurring, (d) is a schematic diagram of bubbles, (e) is a schematic diagram of backgrounds with different color gains, and (f) is a schematic diagram of motion blur.

[0049] Step 3: Interactive segmentation of detection boxes.

[0050] Detection boxes based on particle images output by the YOLO detector and the corresponding confidence level A soft prompt prior graph, SoftPrompt, is constructed for the segmentation stage, and a sparse sliding window strategy is combined to complete efficient interactive segmentation inference. The specific process includes the following steps: First, by analyzing the detection frame area... Gaussian blurring is performed to generate boundary responses, and soft boundary priors are constructed for each detection box. Furthermore, the soft boundary priors corresponding to each detection box are fused at the pixel level. The formula for generating the soft boundary prior is:

[0051] The central heat map is only in the detection frame Construct within the region. Based on the center coordinates of the detection box ( , The spatial distance between the pixel location (x, y) and the detection box region is used to construct a two-dimensional Gaussian distribution, thus obtaining the prior central heatmap. :

[0052] To enhance the constraint effect of high-confidence detection boxes on segmentation results, an adaptive confidence intensity coefficient is introduced:

[0053] Soft boundary prior for each detection box Prior to the central heat map Weighted fusion is performed, and the maximum value is taken at the pixel level to obtain a soft cue prior map of the same size as the input image. :

[0054] To reduce redundant computation in the background region, a sparse sliding window strategy is adopted, selecting only sliding window blocks that overlap with the detection box for segmentation inference. The set of sliding windows to be inferred is defined as follows:

[0055] The input image to be inferred (RGB image) is compared with the soft cue prior image. Four-channel features are formed by concatenating the channel dimensions. Based on a sparse sliding window strategy, sliding window blocks intersecting with the detection box region are selected from these features and fed into UNet++ for segmentation inference to obtain the corresponding sliding window probability prediction results. The probability prediction results of each sliding window block are fused back into the overall probability map according to their spatial location, and then thresholded to finally obtain the binary segmentation mask for the micron-sized particles.

[0056] Step 4: Post-processing and statistical analysis.

[0057] The relevant grain size and morphology parameters are calculated for the segmentation mask results, and statistical results are output. Grain size parameters include: equivalent circle diameter, maximum Ferrette diameter, minimum Ferrette diameter, perimeter, and area. Grain shape parameters include roundness, ellipticity, aspect ratio (length-to-width ratio), and irregularity.

[0058] The equivalent circle diameter refers to the diameter of a circle with the same projected area A as the particle. It is defined as follows:

[0059] Let the projected area of ​​the particle be A, and the projected perimeter be P. Then, the roundness index R based on the area and perimeter is defined as:

[0060] Ellipticity describes the degree to which a particle approximates an elliptical shape, and is expressed as the ratio of the minor axis to the major axis of the Legendre-fitted ellipse. Let the minor axis of the ellipse be... Major axis is The ellipticity E is defined as:

[0061] The aspect ratio is used to characterize the elongation of a particle in both its length and width directions, and is generally expressed as the ratio of the Freret minimum diameter to the Freret maximum diameter. Let the Freret minimum diameter be... The maximum diameter of the Ferret is The formula for aspect ratio AR is:

[0062] Irregularity measures the degree to which particles deviate from an ideal circular shape. Let the diameter of the largest inscribed circle inside the particle be... The smallest circumscribed circle diameter is The formula for irregularity IR is:

[0063] In summary, this application achieves synergy between fine segmentation and efficient inference through a detection-segmentation linkage architecture: the target detection model outputs detection boxes of non-clustered particles and their confidence scores, and the SoftPrompt module generates a soft prior map of the same size as the input image based on this, and concatenates the soft prior map with the original RGB image in the channel dimension to form a four-channel input; at the same time, a sparse sliding window inference strategy is adopted, performing forward computation of UNet++ only on sliding window blocks that intersect with the detection box region, and fusing the outputs of each sliding window back into the whole image probability map and thresholding to obtain a binary segmentation mask. In this way, the sparse sliding window inference strategy reduces background redundant computation and improves segmentation stability and efficiency.

[0064] Below, in an example, the online monitoring of ternary precursor particles is used to verify the practical application effect of the method of the present invention.

[0065] Small target particle detection performance: The preprocessed images were input into the improved target detection model. The dataset consisted of 2336 images, with a training, validation, and test set ratio of 7:2:1, and was trained for 300 epochs. As shown in Table 1, the improved model, using a joint improvement strategy of DMDF and ACF, achieved the best overall performance on the independent test set, building upon the baseline model. The model outputs detection boxes for non-agglomerated ternary precursor particles. The improved model achieved a precision of 0.9887, a recall of 0.9852, and an mAP50 of 0.9940 on the independent test set. Furthermore, a comparative analysis was conducted with a control model that did not incorporate the relevant improved features; the performance comparison is shown in Table 1. From the confusion matrix, compared with the previous model, the direct misclassification between non-clustered particles and clustered particles was reduced to 0, and the misclassification of background particles was also significantly reduced. The false negative rate and false positive rate of non-clustered particles in the improved model were 1.66% and 0.22%, respectively, which were reduced by 5.41 and 6.91 percentage points compared with the previous model. This fully demonstrates that the improved model has better classification performance in the particle clustering state recognition task. The total processing speed of a single image can reach about 207.47 FPS, which meets the real-time requirements.

[0066] Table 1 Ablation Experiment Results

[0067] To verify the segmentation effect of small target particles, the experiment used a dataset containing 300 images and 1548 particle mask pairs to train the segmentation model for 100 rounds. In the experiment, the non-clustered particle detection boxes and confidence scores output by the detection model were used as interactive prompts and input into the UNet++ segmentation model, which finally output the corresponding binary segmentation mask of the particle. The segmentation performance and efficiency comparison of the two schemes are shown in Table 2. The SoftPrompt + sparse sliding window scheme outperformed the UNet++ baseline model: the AP50 index reached 0.989968, an improvement of 1.12 percentage points over the baseline model; in terms of segmentation accuracy, the IoU and Dice coefficients increased to 0.975676 and 0.985525, respectively, with increases of 2.87% and 1.49%; the BF1 coefficient increased to 0.988219, and the instance-level F1 score reached 0.984945, an improvement of 3.79 percentage points over the control model. This indicates that the interactive scheme achieves significant optimization in both instance recognition accuracy and segmentation completeness. Furthermore, by employing a sparse sliding window with a size of 384×384, the model inference efficiency is significantly improved. The average time for pure segmentation is reduced from 197.32 milliseconds in the baseline model to 113.89 milliseconds, and the total time for the entire detection and segmentation process is only 156.71 milliseconds. Overall, it meets the dual requirements of segmentation accuracy and real-time performance in small target particle analysis scenarios.

[0068] Table 2 compares the performance of the segmentation method based on detection box interaction with the UNet++ baseline.

[0069] Post-processing calculations and statistical results: Characterization parameters of the ternary precursor particles were calculated based on particle segmentation masks and statistical analysis was performed. According to statistics, the calculated results of the equivalent area diameter and the minimum Ferete diameter showed a consistent trend. Specifically, the equivalent area diameters D10, D50, and D90 were 9.52 μm, 11.28 μm, and 12.27 μm, with an average particle size of 11.06 μm; the minimum Ferete diameters D10, D50, and D90 were 9.00 μm, 11.08 μm, and 12.00 μm, with an average particle size of 10.8 μm. Both diameter calculations indicate that the overall particle size range is small and uniformly distributed, with no obvious differentiation between coarse and fine particles. Furthermore, the consistency of the parameters further reflects the regularity of the particle morphology, providing a foundation for subsequent particle shape analysis. Particle shape characteristics: The aspect ratio was 0.94, the ellipticity was 0.93, and the irregularity was 0.99. All three parameters were close to 1, indicating high particle morphological symmetry. The roundness reached 0.82, which is at a high level, indicating that the sample particles have smooth edges and no obvious sharp corners or irregular protrusions.

[0070] In yet another example, the invention will be further described in detail below with reference to the accompanying drawings: In this example, the technology system and materials are based on the following: the materials include ternary precursor particles and anhydrous ethanol; the system includes: a camera, an image acquisition card, a light source, an objective lens, an eyepiece, a condenser, etc.

[0071] Step 1: Obtain a multi-background dataset of particles using an optical microscope, including aggregates, bubbles, impurities, and blurred scenes, such as... Figure 2 As shown.

[0072] Step 2: Train the improved object detection model based on the labeled dataset, and output non-clustered particle detection box information. The architecture of the improved object detection model is as follows: Figure 3-5 As shown in Table 1, the detection performance comparison is shown in Table 1, and the confusion matrix comparison is shown in Table 2. Figure 6 As shown, (a) is the confusion matrix of the original YOLOv8 test set, and (b) is the confusion matrix of the improved model test set.

[0073] Step 3: Train a Softprompt + sparse sliding window interaction model based on Unet++ to further segment non-clustered particles and output a segmentation mask for non-clustered particles. A schematic diagram of the interaction process is shown below. Figure 7 As shown, the output reasoning effect is as follows: Figure 8As shown, (a) is a comparison image of the original image and the segmentation mask superimposed, (b) is a binary segmentation mask image, where (a) is the frequency distribution curve of roundness R, (b) is the frequency distribution curve of aspect ratio AR, (c) is the frequency distribution curve of ellipticity E, and (d) is the frequency distribution curve of irregularity IR.

[0074] Step 4: Perform collaborative inference based on the trained detection model and interactive segmentation model. Calculate particle size and shape parameters based on the mask obtained from the inference, and output the particle size distribution map as shown below. Figure 9 The output particle shape distribution map is as follows: Figure 10 As shown, (a) is the frequency distribution curve of roundness R, (b) is the frequency distribution curve of aspect ratio AR, (c) is the frequency distribution curve of ellipticity E, and (d) is the frequency distribution curve of irregularity IR.

[0075] In summary, this invention integrates deep learning object detection and semantic segmentation networks. First, it utilizes a CNN object detection network to locate small target particle regions and accurately distinguish between clustered and non-clustered particles. Then, a Softprompt + sparse sliding window interactive segmentation model based on UNet++ performs high-precision segmentation only on non-clustered particle regions. Finally, it combines morphological image processing algorithms to accurately extract particle size and shape information. This addresses the shortcomings of existing micron-sized particle image detection methods in terms of particle clustering, small target recognition, and segmentation accuracy.

[0076] This application also provides a micron-particle measurement system based on detection frame interaction, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method. This micron-particle measurement system based on detection frame interaction can implement various embodiments of the above-described micron-particle measurement method based on detection frame interaction and achieve the same beneficial effects; further details are omitted here.

[0077] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure 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 disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for measuring micron-sized particles based on detection frame interaction, characterized in that, include: S1: Collect images of micron-sized particles under multiple backgrounds and interference types, construct an image dataset of particles in multiple scenarios, and train an improved CNN object detection network model. S2: Input the set of images to be inferred into the improved CNN object detection network model to obtain the detection box information of small target particles; S3: Generate a soft cue prior map based on the detection box and concatenate it with the image to be inferred to form a four-channel input feature. Select the intersecting region from the four-channel input feature through a sparse sliding window to perform image segmentation and obtain the sliding window result. Output a binarized segmentation mask image based on the sliding window result. S4: Calculate the target parameters of the binarized mask image to obtain the measurement results.

2. The method for measuring micron-sized particles based on detection frame interaction according to claim 1, characterized in that, The improved CNN object detection network model includes a backbone network, a neck network, and a head network connected in sequence; the backbone network includes multiple convolutional modules and at least one dynamic multi-dilation fusion module; the neck network adopts a feature pyramid structure and sets an adaptive cross-scale feature fusion module at the feature fusion node.

3. The method for measuring micron-sized particles based on detection frame interaction according to claim 2, characterized in that, The dynamic multi-expansion fusion module includes a first dynamic multi-expansion fusion module and a second dynamic multi-expansion fusion module, and the processing steps of the two dynamic multi-expansion fusion modules are the same, as follows: The feature maps output from multiple convolutional modules in the backbone network are used as input features. The input features flow in parallel into the multi-dilation convolutional branch and dynamic dilation selector of the dynamic multi-dilation fusion module. After weighted fusion, channel enhancement and residual connection, the output features are obtained. The set of multi-dilated convolution branches is Where B is the number of branches. For the first The convolution dilation rate of the nth branch, the nth The output of each branch for the input features is: ； In the formula, As input features, Indicates the expansion rate 3×3 convolution, For the first Batch normalization operation for each branch The SiLU activation function is used. This represents the nonlinear mapping of the i-th dilated convolution branch, which is formed by concatenating the above three components. The dynamic dilation selector generates branch weights based on global context information, applying them to the input features. Global average pooling satisfies the following relationship: ； In the formula, s represents the input features The channel description vector is obtained by performing global average pooling in the spatial dimension. , The number of channels for X; Then, the intermediate feature vector z is obtained by first performing a 1×1 convolution and SiLU activation, and then a 1×1 convolution is performed followed by Softmax normalization in the branch dimension to generate the branch weights: ； in, , These are the weight matrices corresponding to a 1×1 convolution. For each expansion branch, the adaptive weights are... The intermediate feature vector after SiLU activation; Using weights The local feature maps output by each branch are dynamically weighted and summed to obtain multi-scale fused features. It satisfies the following relationship: Fusion features Perform global average pooling, and then sequentially obtain channel attention through two 1×1 convolutions and a SiLU+1×1 convolution followed by a Sigmoid function, satisfying the following relationship: ， ， in, This is the channel description vector obtained by global average pooling of the fused features in the spatial dimension. The intermediate channel features are those activated by SiLU. It is the Sigmoid activation function. The attention map is obtained by using 1×1 convolution parameters. The characteristics after channel enhancement are: In the formula, This represents element-wise multiplication. right Performing 1×1 convolution and BatchNorm fusion compression satisfies the following relationship: In the formula, This indicates that the convolution output is normalized using BatchNorm. The weights are the parameters for a 1×1 convolution. Finally, residual connections are used to obtain the module output, which satisfies the following relationship: ； The features output by the dynamic multi-expansion fusion module are input to the neck network through the Concat structure and participate in the subsequent multi-scale feature fusion processing as one of the input features of the adaptive cross-scale feature fusion module.

4. The method for measuring micron-sized particles based on detection frame interaction according to claim 2, characterized in that, The processing steps of the adaptive cross-scale feature fusion module are as follows: The system receives multi-scale feature maps from different layers of the backbone network, aligned via bidirectional propagation paths (top-down and bottom-up) in the neck network. These multi-scale feature maps are first processed through 1×1 and 3×3 convolutional branches to extract local details and contextual information. Then, feature reweighting and enhancement are performed using channel attention and residual connections to obtain a fused feature map for external head structure detection. Let the input features be , The number of input channels is denoted as and the number of output channels of the module is denoted as . H and W represent the spatial dimensions of the fused feature map in the vertical and horizontal directions, respectively. The module extracts features through two parallel 1×1 and 3×3 convolutional branches, and the output is: The two branches are added element-wise along the channel dimension to obtain the initial multi-scale fused features. : In fusion features A lightweight channel attention mechanism is introduced, which performs global average pooling on each channel to obtain a channel-level global response vector. ,in: ； In the formula, For fusion features No. Coordinates in each channel The feature values ​​at the location, i=1,…,H, j=1,…,W, represent the pixel indices of the feature map in the vertical and horizontal directions, respectively; Subsequently, two 1×1 convolutional layers with SiLU and Sigmoid activations are used to generate channel attention weight vectors. It satisfies the following relationship: ； in, The SiLU activation function is used. For the Sigmoid function, , These are the weight matrices corresponding to the two 1×1 convolutional layers. This refers to the channel compression ratio. Using channel attention weight vectors Fusion features Perform channel-by-channel reweighting to obtain the weighted feature map. For fusion features The Each channel, its weighted eigenvalue The calculation method is as follows: in For the first Attention coefficients for each channel For fusion features No. Coordinates in each channel Eigenvalues ​​at; The weighted features are then subjected to 1×1 convolution and BatchNorm for channel compression and fusion to obtain intermediate features that satisfy the following relationship: A residual connection structure is added to the output. When the number of input and output channels is the same, the input features are added to the intermediate features; when they are not the same, a 1×1 convolution is used to align the number of input channels first, and the residual term is then processed. Represented as: The output features are finally obtained by superimposing residuals and performing SiLU nonlinear activation. : ； The output feature map The input feature map is used as the target detection head. Based on the detection head parameters, the category of micron-sized particles and the coordinates of the detection box are regressed and classified, and this information is provided to the segmentation model.

5. The method for measuring micron-sized particles based on detection frame interaction according to claim 1, characterized in that, S3 includes: Detection boxes based on particle images output by the YOLO detector and the corresponding confidence level A soft prompt prior graph, SoftPrompt, is constructed for the segmentation stage, and a sparse sliding window strategy is combined to complete efficient interactive segmentation inference. The specific process includes the following steps: First, by analyzing the detection frame area... Gaussian blurring is performed to generate boundary responses, and soft boundary priors are constructed for each detection box. Furthermore, the soft boundary priors corresponding to each detection box are fused at the pixel level. The formula for generating the soft boundary priors is as follows: In the formula, i is the index of the detection box; Represents the Gaussian kernel. Related to the detection frame size, it is used to control the soft boundary diffusion range. For detection box The indicator function is set to 1 if the pixel (x,y) is inside the bounding box, and 0 otherwise; The central heat map is only in the detection box Constructed within the region, based on the center coordinates of the detection box ( , The spatial distance between the pixel location (x, y) and the detection box region is used to construct a two-dimensional Gaussian distribution, thus obtaining the prior central heatmap. : In the formula, , The standard deviation of the Gaussian distribution is used to control the diffusion range of the heatmap; To enhance the constraint effect of high-confidence detection boxes on segmentation results, an adaptive confidence intensity coefficient is introduced: In the formula, Let be the confidence level of the i-th detection box; This is a hyperparameter used to adjust the degree of influence of confidence level on prior strength; This means truncating the result to the interval [0, 1]. Soft boundary prior for each detection box Prior to the central heat map Weighted fusion is performed, and the maximum value is taken at the pixel level to obtain a soft cue prior map of the same size as the input image. : In the formula, This indicates that the prior responses of all detection boxes are fused by taking the maximum value pixel by pixel; , These represent the weighting coefficients for the soft boundary branch and the heatmap branch, respectively. To reduce redundant computation in the background region, a sparse sliding window strategy is adopted, selecting only sliding window blocks that overlap with the detection box for segmentation inference. The set of sliding windows to be inferred is defined as follows: In the formula, For the j-th sliding window region, This indicates the existence of a detection box that intersects with the sliding window. Input RGB image and soft cue prior image Four-channel input features are formed by splicing the channel dimensions, and sliding window blocks that intersect with the detection box region are selected from the features according to the sparse sliding window strategy and sent to UNet++ for segmentation inference to obtain the corresponding sliding window probability prediction results. The probability prediction results of each sliding window block are fused back into the whole image probability map according to their spatial position, and the binarized segmentation mask of micron particles is finally obtained through thresholding.

6. The method for measuring micron-sized particles based on detection frame interaction according to claim 1, characterized in that, S4 includes: The target parameters of the binary mask image are calculated and statistical results are output. The target parameters include granularity parameters and grain shape parameters. The granularity parameters include: equivalent circle diameter, maximum Ferrette diameter, minimum Ferrette diameter, perimeter, and area. The grain shape parameters include: roundness, ellipticity, aspect ratio, and irregularity.

7. The method for measuring micron-sized particles based on detection frame interaction according to claim 6, characterized in that, The calculation of the equivalent circle diameter satisfies the following relationship: In the formula, A is the projected area of ​​the particle, and P is the projected perimeter; The roundness R index, based on area and perimeter, is defined as follows: In the formula, P is the projected perimeter; Let the minor axis of the ellipse be... The major axis is The ellipticity E is defined as: Let the minimum diameter of the Ferrette be... The maximum diameter is The formula for the aspect ratio AR is: Let the diameter of the largest inscribed circle inside the particle be... The smallest circumscribed circle diameter is The formula for regularity IR is: 。 8. A micron-particle measurement system based on detection frame interaction, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any of the methods described in claims 1-7.

Images