Spherical particle size measurement method and device based on deep learning numerical prediction

Abstract

The invention discloses a spherical particle size measurement method and device based on deep learning numerical prediction, and belongs to the field of particle measurement and image processing. The method comprises the following steps: acquiring interference fringe image data of spherical particles with different sizes, marking a corresponding particle size label for each image, and dividing the images into a training set, a verification set and a test set; the training set is used to train a pre-constructed neural network model, the verification set monitors the training process and stores the optimal model weight, and the test set tests the prediction precision of the model; the pre-constructed neural network model is based on an original UNet + + network and comprises a plurality of residual blocks and a space attention module, and a full-connection output head is added to realize end-to-end mapping from an image to a size value; and finally, inputting an interference fringe image of a to-be-measured particle into the trained network, and directly outputting a particle size prediction value. According to the method, high-precision, real-time and end-to-end prediction of the spherical particle size is realized, and the method is suitable for scenes such as cloud particle field on-line monitoring.

Description

Technical Field

[0001] This invention relates to the field of particle measurement and image processing technology in optical imaging, and more specifically to a method and apparatus for measuring the size of spherical particles based on deep learning numerical prediction. Background Technology

[0002] Warm clouds are composed of spherical water droplet particles. Measuring cloud particle size not only helps improve the accuracy of weather forecasts and reduce the risk of meteorological disasters, but also plays a crucial role in revealing the mechanisms of cloud formation and evolution and optimizing processes such as artificial rain enhancement. Interferometric particle imaging (IPI) is a measurement method based on the distribution of particle scattered light fields. This method uses Mie scattering theory to study the relationship between particle size and the distribution of scattered light intensity, and acquires particle scattered light information through an optical imaging system to achieve particle size measurement. In research on particle size measurement using interferometric particle imaging technology, existing patent document CN116559033A extracts the fringe frequencies in the interference fringe pattern using a two-dimensional Fourier transform algorithm to obtain particle size, but it cannot acquire cloud particle field information in real time.

[0003] Currently, the UNet++ network has been applied to the field of spherical particle size measurement. The existing patent CN119935850A has achieved spherical particle size measurement, but the image reconstruction method requires converting the reconstructed image into size values, which is a complex process. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a method and apparatus for measuring the size of spherical particles based on deep learning numerical prediction that overcomes or at least partially solves the above problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, embodiments of the present invention provide a method for measuring the size of spherical particles based on deep learning numerical prediction, comprising the following steps: Dataset preparation steps: Obtain interference fringe image data of spherical particles of different sizes, and label each image with the corresponding particle size to construct a labeled dataset; Dataset partitioning steps: Divide the dataset into a training set, a validation set, and a test set; Network training steps: The pre-built neural network model is trained using the training set, the training process is monitored and the optimal model weights are saved using the validation set, and the prediction accuracy of the model is tested using the test set; The pre-built neural network model is based on the original UNet++ network, including multiple residual blocks and spatial attention modules, and a fully connected output head is added to achieve end-to-end mapping from image to size value; Size prediction step: Input the interference fringe image of the particle to be measured into the trained network, and directly output the predicted particle size value.

[0007] In one embodiment, the pre-built neural network model specifically includes: Seven residual blocks are concatenated after the two-layer convolutional structure at the encoding and decoding ends; Each residual block embeds a spatial attention module, which is then converted into spatial attention weights by 1×1 convolution and sigmoid activation. Furthermore, a fully connected regression head is added to the network output, which maps high-dimensional features into a single scalar through the fully connected layer, achieving end-to-end size prediction from pixel-level representation to image-level regression index.

[0008] In one embodiment, the spherical particle interference fringe image data in the dataset preparation step includes: experimentally acquired data and simulation-generated data; The experimental data was acquired using an interferometric particle imaging system, which includes a laser, a spatial filter, a collimating lens, a cylindrical lens group, a sample cell, an imaging lens, and a CCD camera. Diluted polystyrene spherical particles of different sizes were placed in the sample cell as experimental particles and irradiated by a sheet-like laser beam. The CCD camera then captured the interference fringe pattern of the spherical particles.

[0009] In one embodiment, the CCD camera has a resolution of 2448×2048 and a pixel size of 3.45μm.

[0010] In one embodiment, the experimental particles are polystyrene spherical particles of 30μm, 45μm, 60μm, and 90μm, which are diluted and placed in a sample cell containing deionized water.

[0011] In one embodiment, the training process in the network training step includes: selecting Adam as the optimizer and setting the initial learning rate to 1×10. -4 The weight decay factor is set to 1×10. -8 A learning rate adjustment strategy is adopted, in which the learning rate is decayed by 0.1 every 10 training epochs; the number of training epochs is 150 epochs. During the training process, the model parameters with the minimum current loss are dynamically saved based on the validation set loss, thereby obtaining the optimal network weights.

[0012] In a second aspect, embodiments of the present invention provide a spherical particle size measurement device based on deep learning numerical prediction, for implementing the spherical particle size measurement method based on deep learning numerical prediction as described in any of the first aspects, the device comprising: The dataset preparation module acquires interference fringe image data of spherical particles of different sizes, and labels each image with the corresponding particle size to construct a labeled dataset. The dataset partitioning module divides the dataset into a training set, a validation set, and a test set. The network construction and training module uses the training set to train the pre-built neural network model, monitors the training process and saves the optimal model weights through the validation set, and tests the prediction accuracy of the model through the test set. The pre-built neural network model is based on the original UNet++ network, includes multiple residual blocks and spatial attention modules, and adds a fully connected output head to realize end-to-end mapping from image to size value. The size prediction module inputs the interference fringe image of the particle to be tested into the trained network and directly outputs the predicted particle size value.

[0013] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of the first aspects.

[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of the first aspects.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: First, real-time processing. The method directly outputs predicted particle sizes from the input particle interference fringe pattern. Compared to traditional image processing methods, this method offers higher real-time performance.

[0016] Second, it has high accuracy. Experimental noise affects the accuracy of measurements, and this method achieves higher accuracy compared to traditional methods.

[0017] Third, size prediction. By improving the size measurement network, the function of predicting the size of spherical particles was realized. Compared with traditional image processing methods, this avoids formula derivation and reduces complexity.

[0018] This invention combines interferometric imaging technology with UNet++ neural network to predict the size of spherical particles, providing technical support for the acquisition of cloud particle field information. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for measuring the size of spherical particles based on deep learning numerical prediction provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the interferometric particle imaging system provided in an embodiment of the present invention; Among them: 1. Laser, 2. Microscope objective, 3. Pinhole, 4. Collimating lens, 5. Plano-convex cylindrical lens, 6. Plano-concave cylindrical lens, 7. Sample cell, 8. Imaging lens, 9. CCD camera. Figure 3 This is a schematic diagram of the interference fringe pattern and processing results provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the improved UNet++ prediction network structure provided in the embodiments of the present invention. Figure 5 The above are the predicted particle size results for spherical particles provided in the embodiments of the present invention, wherein: (a)-(c) are predicted particle size results for spherical particles with diameters of 30μm, 45μm and 90μm respectively; Figure 6 This is a block diagram of a spherical particle size measurement device based on deep learning numerical prediction provided in an embodiment of the present invention. Figure 7 This is a structural diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0022] Example 1: Reference Figure 1As shown, this embodiment of the invention provides a method for measuring the size of spherical particles based on deep learning numerical prediction, including the following steps: S1. Dataset preparation steps: Obtain interference fringe image data of spherical particles of different sizes, and label each image with the corresponding particle size label to construct a labeled dataset; For example, an interferometric particle imaging system can be built to obtain interference fringe patterns of spherical particles of different sizes; such as Figure 2 As shown, the interferometric particle imaging system, arranged horizontally from left to right, consists of a 532nm laser 1, a spatial filter composed of an objective lens 2 and a pinhole 3, a collimating lens 4, a lens group composed of a plano-convex cylindrical lens 5 and a plano-concave cylindrical lens 6, a sample cell 7, followed by an imaging lens 8 and a CCD camera 9. This system is used to acquire spherical particle interference fringe patterns and construct a dataset: diluted polystyrene spherical particles are placed in the sample cell as experimental particles, irradiated by a sheet-like laser beam, and the spherical particle interference fringe patterns are acquired at the CCD end. Simulated interference fringe patterns can also be generated using an optical transfer matrix simulation program. Each image is labeled with a size.

[0023] The CCD camera used a resolution of 2448×2048 and a pixel size of 3.45μm. The experimental particles were polystyrene spherical particles of 30μm, 45μm, 60μm and 90μm, which were diluted and placed in a sample cell containing deionized water.

[0024] S2. Dataset partitioning steps: Divide the dataset into a training set, a validation set, and a test set; for example, it can be partitioned in a 7:2:1 ratio.

[0025] S3. Network training steps: The pre-built neural network model is trained using the training set. The training process is monitored and the optimal model weights are saved using the validation set. The prediction accuracy of the model is tested using the test set. The pre-built neural network model is based on the original UNet++ network, including multiple residual blocks and spatial attention modules, and a fully connected output head is added to achieve end-to-end mapping from image to size value.

[0026] Specific improvements include: First, based on the original UNet++ network, seven residual modules are introduced into the encoding and decoding paths. These residual connections superimpose shallow details with deep semantic features, mitigating gradient decay and preserving edge structure information. After the input image undergoes double convolution to extract local texture, the residual modules enhance deep representations and stabilize feature transfer, ensuring key structures are preserved during multi-scale fusion. Specifically, these residual blocks are concatenated after the double convolutional structures at the encoding and decoding ends. Second, seven spatial attention modules are introduced into the encoding and decoding paths to generate spatial weights and weight the feature maps pixel-by-pixel. This enhances the response of the target region and suppresses background and redundant interference. The attention branch recalibrates the output based on the feature spatial distribution, enabling the network to prioritize the target region and improve boundary detail recovery during reconstruction and fusion. Specifically, spatial attention modules are embedded within each residual block, and then spatial attention weights are generated through 1×1 convolution and Sigmoid activation. Third, a fully connected output head is added. The fully connected layer maps high-dimensional features into a single scalar. The fused features are aggregated into a global vector and the size parameters are obtained by regression. This reduces the dependence on post-processing such as threshold segmentation and connected components, and improves real-time performance and consistency. Specifically, a fully connected regression head is added at the network output. The fully connected layer maps high-dimensional features into a single scalar, realizing end-to-end size prediction from pixel-level representation to image-level regression index.

[0027] During training, the hyperparameters are first set: the optimizer is Adam, and the initial learning rate is set to 1×10. -4 The weight decay factor is set to 1×10. -8 A learning rate adjustment strategy is adopted, in which the learning rate is decayed by 0.1 every 10 training epochs; the number of training epochs is 150 epochs. During the training process, the model parameters with the minimum current loss are dynamically saved based on the validation set loss, thereby obtaining the optimal network weights.

[0028] By learning to construct a mapping relationship between interference fringe patterns and particle size using the training set, and utilizing the optimal weights obtained from training on the validation set, predicted values ​​for particles of different sizes can be obtained based on the test set (without size labels). For example, when the relative error between the actual size and the measurement result in the test set data is less than 0.093%, and the average detection rate is 72.99 fps, the corresponding pre-built neural network model has the most objective and reliable real-world capabilities and generalization performance, and can be used for inference applications to achieve high-precision real-time prediction of spherical particle size.

[0029] S4. Size Prediction Step: Input the interference fringe image of the particle to be measured into the trained network, and directly output the predicted particle size value. For example, by using the interference fringe image of the particle whose size needs to be measured as input, the predicted particle size can be output.

[0030] This invention achieves high-precision, real-time, end-to-end prediction of spherical particle size, and is applicable to scenarios such as online monitoring of cloud particle fields.

[0031] For example, in specific implementation, the spherical particle size measurement method based on deep learning numerical prediction provided by this invention includes the following: Step 1: Set up as follows Figure 2 The interferometric particle imaging system shown acquires interference fringe patterns of spherical particles. The system includes: a laser, a spatial filter consisting of an objective lens and a pinhole, a collimating lens, a lens group consisting of a plano-convex cylindrical lens and a plano-concave cylindrical lens, an imaging lens, and a CCD camera. From left to right along the horizontal direction, the system consists of: a 532nm laser 1, a spatial filter consisting of an objective lens 2 and a pinhole 3, a collimating lens 4, a lens group consisting of a plano-convex cylindrical lens 5 and a plano-concave cylindrical lens 6, and a sample cell 7 containing polystyrene spherical particles. Above the sample cell 7, the imaging lens 8 and the CCD camera 9 are positioned sequentially. The system is used to acquire a dataset of spherical particle interference fringe patterns. The system parameters are set as follows: object distance 140.90mm, image distance 77.50mm, scattering angle 60.73°, and magnification 0.55. The particles to be tested were 30μm, 45μm, 60μm and 90μm polystyrene spherical particles, which were placed in a sample cell filled with deionized water; after being irradiated by a sheet laser beam, the interference fringe pattern of the spherical particles was finally acquired at the CCD end; Step 2: Generate simulation images using the optical transfer matrix and construct a size prediction network dataset; obtain simulated interference fringe patterns using an optical transfer matrix simulation program. Combine the simulation images with the experimental interference fringe patterns from Step 1 to build the size prediction network dataset. The resulting dataset contains 9984 images, including 3726 simulation images and 6258 experimental images, with a resolution of 256×256. Label the spherical particles of different sizes, with the label value being the particle size; divide the spherical particle interference fringe dataset into training, validation, and test sets in a 7:2:1 ratio. like Figure 3 As shown, (a)-(d) are interference fringe patterns of 30 μm spherical particle experiments; (e)-(h) are interference fringe patterns of 45 μm spherical particle experiments; and (i)-(l) are interference fringe patterns of 90 μm spherical particle experiments.

[0032] Step 3: Network structure improvement. For example... Figure 4As shown, the improvements include: First, based on the original UNet++ network, seven residual modules are introduced into the encoding and decoding paths, specifically, residual blocks are concatenated after the double-layer convolutional structures at the encoding and decoding ends. Second, seven spatial attention modules are introduced into the encoding and decoding paths, specifically, spatial attention modules are embedded within each residual block, and then spatial attention weights are generated through 1×1 convolution and Sigmoid activation, and multiplied point-by-point with the original feature map, thereby adaptively highlighting key information regions and suppressing redundant background features. Third, a fully connected output head is added, specifically, a fully connected regression head is added at the network output end, which maps high-dimensional features to a single scalar through a fully connected layer, and uses Sigmoid activation to achieve normalized output, realizing end-to-end size prediction from pixel-level representation to image-level regression indicators; Specifically: This network uses UNet++ as its backbone and employs an encoder-decoder fusion-regression output head structure: the input is a 512×512 single-channel interference fringe image. First, a reconstructed representation at the same scale as the input is formed, then global pooling and fully connected regression are applied, finally outputting a single particle size prediction. At the encoder, the channel configuration is nb_filter=[16,32,64,128,256]: the first encoding layer maps the input to 16 channels while maintaining the original scale; subsequently, max pooling and encoding units with strides are used to progressively downsample and increase the number of channels, forming multi-scale encoded features for the decoder. Residual blocks are introduced at the end of the encoder and decoder. Their main branches consist of multiple 3×3 convolutions followed by BN layers and ReLU activation, and are combined with shortcut branches to mitigate gradient decay. Furthermore, each residual block embeds a spatial attention mechanism module to enhance the particle fringe-related regions and suppress the background. At the decoder, multi-node dense fusion is performed using the nested approach of UNet++: each node concatenates the features of historical nodes at the same layer with the upsampled features from deeper layers before fusing them. Finally, the decoded top-level features are convolved with 1×1 to obtain the output feature map, and then a single size value is output through global average pooling, fully connected layers and sigmoid activation function, realizing end-to-end fast prediction of IPI image input to size result, reducing the complex steps of traditional image reconstruction and frequency extraction.

[0033] Step 4: Train the network and obtain the optimal weights. Using the dataset from Step 2 and the improved network from Step 3, train the network. First, set the hyperparameters: select Adam as the optimizer and set the initial learning rate to 1×10⁻⁶. -4 The weight decay factor is set to 1×10. -8 A learning rate adjustment strategy is adopted, in which the learning rate is decayed by 0.1 every 10 training epochs; the number of training epochs is 150 epochs; during the training process, the model parameters with the minimum current loss are dynamically saved based on the validation set loss, thereby obtaining the optimal network weights. Step 5: Test set testing: The size prediction value is obtained by detecting the interference fringe pattern. For example... Figure 3 The right side of the middle section shows a visualization of the prediction results of the improved UNet++ network for the (a)-(l) stripe pattern. The numerical results are simultaneously exported as a .csv data file.

[0034] By constructing a mapping relationship between interference fringe patterns and particle size, and utilizing the optimal weights obtained during training, predicted values ​​for particles of different sizes in the test set are obtained. For example... Figure 5 As shown in (a)-(c), the prediction results were obtained by randomly selecting 20 interference fringe patterns of 30μm, 45μm, and 90μm particles. The relative errors were 0.15%, 0.04%, and 0.08%, respectively. The average relative error of the measurement results was less than 0.093%, and the average detection rate was 72.99fps. This trained model can achieve high-precision real-time prediction of the size of spherical particles.

[0035] Step 6: The deployed IPI system acquires interference fringe images of cloud particles of different sizes to be predicted. These images are then input into the trained network for prediction and inference. The network can output the corresponding cloud particle size prediction value in real time in a single forward computation. Compared to the multi-step processing of traditional processes, which requires image reconstruction, frequency extraction, and particle size conversion, this method integrates imaging to size regression into a one-step end-to-end prediction. Therefore, it has the advantages of fast processing speed, high measurement accuracy, simplified implementation link, and easy online real-time deployment, which can meet the requirements for real-time and accurate acquisition of cloud particle size.

[0036] This invention combines interferometric particle imaging (IPI) technology with convolutional neural networks to propose a method for measuring the size of spherical particles based on deep learning numerical prediction. This method enables direct acquisition of the size of spherical particles, reducing complexity and saving measurement time. By collecting a large amount of experimental data to train the network, improving the network to enable prediction, and adjusting the network hyperparameters to obtain optimal weights, this invention achieves high-precision real-time prediction of the size of spherical particles.

[0037] Example 2: Based on the same inventive concept, this embodiment of the invention also provides a spherical particle size measurement device based on deep learning numerical prediction, used to implement the spherical particle size measurement method based on deep learning numerical prediction as described in Embodiment 1, referring to... Figure 6 As shown, the device includes: The dataset preparation module acquires interference fringe image data of spherical particles of different sizes, and labels each image with the corresponding particle size to construct a labeled dataset. The dataset partitioning module divides the dataset into a training set, a validation set, and a test set. The network training module uses the training set to train the pre-built neural network model, monitors the training process and saves the optimal model weights through the validation set, and tests the prediction accuracy of the model through the test set. The pre-built neural network model is based on the original UNet++ network, includes multiple residual blocks and spatial attention modules, and adds a fully connected output head to realize end-to-end mapping from image to size value. The size prediction module inputs the interference fringe image of the particle to be tested into the trained network and directly outputs the predicted particle size value.

[0038] Since the principles by which these devices solve problems are similar to those of the methods described above, the implementation of these devices can be found in the implementation of the methods described above, and the repetitions will not be repeated.

[0039] Example 3: Based on the same inventive concept, the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it is able to implement the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of Embodiments 1.

[0040] like Figure 7 As shown, the electronic device may include: a processor 10, a communication interface 20, a memory 30, and a communication bus 40, wherein the processor 10, the communication interface 20, and the memory 30 communicate with each other via the communication bus 40. The processor 10 can call logical instructions in the memory 30 to execute a method for measuring the size of spherical particles based on deep learning numerical prediction. This method includes: Dataset preparation steps: Obtain interference fringe image data of spherical particles of different sizes, and label each image with the corresponding particle size to construct a labeled dataset; Dataset partitioning steps: Divide the dataset into a training set, a validation set, and a test set; Network training steps: The pre-built neural network model is trained using the training set, the training process is monitored and the optimal model weights are saved using the validation set, and the prediction accuracy of the model is tested using the test set; The pre-built neural network model is based on the original UNet++ network, including multiple residual blocks and spatial attention modules, and a fully connected output head is added to achieve end-to-end mapping from image to size value; Size prediction step: Input the interference fringe image of the particle to be measured into the trained network, and directly output the predicted particle size value.

[0041] Example 4: This invention also provides a computer-readable storage medium containing a program for executing a method for measuring the size of spherical particles based on deep learning numerical prediction as described in Embodiment 1 above. This program can be executed on a processor.

[0042] Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0043] The program stored on this medium is loaded into the processor's memory and executed to perform various functions. This storage medium, connected to hardware devices, enables the computer to perform the steps of Embodiment 1 described above.

[0044] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0045] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for measuring the size of spherical particles based on deep learning numerical prediction, characterized in that, Includes the following steps: Dataset preparation steps: Obtain interference fringe image data of spherical particles of different sizes, and label each image with the corresponding particle size to construct a labeled dataset; Dataset partitioning steps: Divide the dataset into a training set, a validation set, and a test set; Network training steps: The pre-built neural network model is trained using the training set, the training process is monitored and the optimal model weights are saved using the validation set, and the prediction accuracy of the model is tested using the test set; The pre-built neural network model is based on the original UNet++ network, including multiple residual blocks and spatial attention modules, and a fully connected output head is added to achieve end-to-end mapping from image to size value; Size prediction step: Input the interference fringe image of the particle to be measured into the trained network, and directly output the predicted particle size value.

2. The method according to claim 1, characterized in that, The pre-built neural network model specifically includes: Seven residual blocks are concatenated after the two-layer convolutional structure at the encoding and decoding ends; Each residual block embeds a spatial attention module, which is then converted into spatial attention weights by 1×1 convolution and sigmoid activation. Furthermore, a fully connected regression head is added to the network output, which maps high-dimensional features into a single scalar through the fully connected layer, achieving end-to-end size prediction from pixel-level representation to image-level regression index.

3. The method according to claim 1, characterized in that, In the dataset preparation step, the spherical particle interference fringe image data includes: experimentally acquired data and simulation-generated data; The experimental data was acquired using an interferometric particle imaging system, which includes a laser, a spatial filter, a collimating lens, a cylindrical lens group, a sample cell, an imaging lens, and a CCD camera. Diluted polystyrene spherical particles of different sizes were placed in the sample cell as experimental particles and irradiated by a sheet-like laser beam. The CCD camera then captured the interference fringe pattern of the spherical particles.

4. The method according to claim 3, characterized in that, The CCD camera has a resolution of 2448×2048 and a pixel size of 3.45μm.

5. The method according to claim 3, characterized in that, The experimental particles were polystyrene spherical particles of 30μm, 45μm, 60μm, and 90μm, which were diluted and placed in a sample cell containing deionized water.

6. The method according to claim 1, characterized in that, The training process in the network training steps includes: selecting Adam as the optimizer and setting the initial learning rate to 1×10. -4 The weight decay factor is set to 1×10. -8 A learning rate adjustment strategy is adopted, in which the learning rate is decayed by 0.1 every 10 training epochs; the number of training epochs is 150 epochs. During the training process, the model parameters with the minimum current loss are dynamically saved based on the validation set loss, thereby obtaining the optimal network weights.

7. A device for measuring the size of spherical particles based on deep learning numerical prediction, characterized in that, For implementing the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of claims 1 to 6, the apparatus comprises: The dataset preparation module acquires interference fringe image data of spherical particles of different sizes, and labels each image with the corresponding particle size to construct a labeled dataset. The dataset partitioning module divides the dataset into a training set, a validation set, and a test set. The network construction and training module uses the training set to train the pre-built neural network model, monitors the training process and saves the optimal model weights through the validation set, and tests the prediction accuracy of the model through the test set. The pre-built neural network model is based on the original UNet++ network, includes multiple residual blocks and spatial attention modules, and adds a fully connected output head to realize end-to-end mapping from image to size value. The size prediction module inputs the interference fringe image of the particle to be tested into the trained network and directly outputs the predicted particle size value.

8. A computer device, 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 program, it implements the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for measuring the size of spherical particles based on deep learning numerical prediction as described in any one of claims 1 to 6.

Images