Cloud particle diffraction image acquisition device and classification method

Abstract

The invention discloses a cloud particle diffraction image acquisition device and a classification method, and belongs to the field of image processing and meteorology. The cloud particle diffraction image acquisition device comprises a detection light path and an imaging plane, wherein the detection light path is mainly used for generating a proper sampling light beam and transmitting diffraction information through the light beam; a light-transmitting calibration turntable is placed on an imaging plane, a diffraction effect is generated after light beams are sampled to generate propagation light beams containing diffraction information, and the propagation light beams receive light beam information through a linear array cloud particle detector and are used for identifying whether shadows in the propagation light beams change or not and identifying diffraction fringe structures. The classification method comprises the steps of obtaining a cloud particle diffraction image data set, building a convolutional neural network classification model of the cloud particle diffraction images and carrying out classification and identification on the cloud particle diffraction images. According to the method, the problem that the particle diffraction images are not clear under different conditions can be solved, the efficiency and accuracy of obtaining the cloud particle diffraction images in any shape are improved, and classification research of particle graphs obtained by an optical array imaging device can be promoted.

Description

Technical Field

[0001] The invention belongs to the fields of image processing and meteorology, and relates to a cloud particle diffraction image acquisition device and a classification method. Background Art

[0002] Clouds are visible aggregates composed of a large number of cloud particles (water droplets, ice crystals) with diverse shapes and a wide range of sizes. Meteorological agencies typically use cloud particle images acquired by airborne probes penetrating clouds to reveal microphysical processes within clouds, which are then used for weather forecasting, environmental protection, and weather modification. To meet the requirements of real-time and continuous detection and further improve the reliability of practical applications, imaging systems based on linear array cloud particle detectors are generally used for detection, acquiring high-precision in-situ cloud particle images for analysis. Linear array cloud particle detectors, based on the principle of light intensity attenuation caused by particle obstruction, offer advantages such as a large size measurement range and simple shape acquisition. They can achieve real-time measurement of micron-sized cloud particles, and have led to the development of highly mature optical array imaging devices such as cloud image probes (CIPs) and two-dimensional stereo probes (2D-S). The scale of cloud particles usually ranges from tens of microns to one thousand microns. When the particle size, light wavelength, and the distance from the detector to the particle plane meet the diffraction conditions, the detector records not only the obstruction shadow, but also the diffraction fringe structure. Due to the influence of the phase distribution of the diffracted light, the diffraction fringe structure has many categories and complex transformations.

[0003] Chinese invention patent CN106092835A proposes an atmospheric precipitation particle measurement device based on the diffraction principle. This device records the shadow transformation of particle occlusions through a linear array detector to achieve large-scale, high-precision cloud particle measurement. However, it does not further discuss the imaging of particles with a diameter of less than 100 μm. Chinese invention patent CN116559033A proposes a hybrid particle field particle size measurement method based on IPI technology. The Resnet50 network model is used to classify and distinguish the interference images of particles of different morphologies, completing the shape classification of particle images as spherical or non-spherical, but it cannot further distinguish the types of non-spherical particles. Chinese invention patent CN111860570A proposes a cloud particle image extraction and classification method. By filtering and completing the image data of cloud particle images obtained by CIP, a cloud particle image dataset is established, and a deep neural network classification model based on transfer learning is used to classify the cloud particle images, improving the classification accuracy and reliability. However, its dataset only comes from images collected in the early stage of CIP, and the data source is relatively single.

[0004] In fact, due to the differences in the structure and methods of cloud particle detection equipment, the cloud particle images obtained are also different. The amount of data obtained by simply building an experimental device to simulate cloud particle images is extremely small, and classification is relatively complicated. Although the images collected by existing equipment such as CIP are easy to obtain, there are problems with the accuracy and update of the database. In addition, it is difficult to meet the high-precision measurement requirements of particles by only distinguishing between circular and non-circular shapes through neural network models, and there are still great difficulties for computers to accurately identify and classify the various shapes of cloud particles at the micron level. These problems have affected the accuracy of cloud particle size parameters and the reliability of using cloud particle classification results for meteorological research. Summary of the Invention

[0005] In response to the above-mentioned problems existing in the prior art, the present invention provides a cloud particle diffraction image acquisition device and classification method. Specifically, a cloud particle diffraction image acquisition device is combined with a simulation algorithm to generate a large number of cloud particle diffraction images, which are then identified and classified through neural network learning. This method can solve the problems of unclear cloud particle diffraction images and low efficiency and accuracy of manual or general algorithm classification. The present invention utilizes the characteristics of micron-level cloud particles, whose contour features are severely distorted while their internal diffraction features are different, and combines convolutional neural network training to obtain different cloud particle diffraction image classification models, ultimately achieving rapid and accurate classification of cloud particle images. The process flow chart is shown below. Figure 1 shown.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A cloud particle diffraction image acquisition device is an adjustable cloud particle diffraction image generation device comprising a detection optical path and an imaging plane. The detection optical path primarily generates a suitable sampling beam and propagates diffraction information through the beam. A translucent fixed turntable is placed on the imaging plane. After passing through the sampling beam, a diffraction effect occurs to generate a propagation beam containing diffraction information. This propagation beam receives beam information through a linear array cloud particle detector, which is used to identify whether there are shadow changes and diffraction fringe structures in the propagation beam. Specifically:

[0008] The detection optical path includes a laser emitter, a lens group, and a reflector group. The lens group includes a first lens, a second lens, and a third lens, and the reflector group includes a first reflector and a second reflector. The laser emitter's emitting end is vertically upward and is used to emit a laser beam. The first reflector redirects the vertically upward beam emitted by the laser emitter into a horizontal beam, while the second reflector reflects the horizontal beam into a vertically downward direction. A movable, translucent, and fixed-width turntable is provided between the first and second reflectors. The first and second lenses are arranged between the laser emitter and the first reflector, with their centers located on the central axis of the laser beam. They function to produce a collimated, expanded, and fixed-width parallel light beam. A third lens is arranged between the second reflector and the linear cloud particle detector, with its center located on the central axis of the laser beam. The third lens provides a microscopic magnification function.

[0009] The imaging plane includes a sampling plane and an observation plane. A translucent fixed disk is placed on the sampling plane. The translucent fixed disk is driven by a precision servo motor to rotate at high speed and can move in the direction parallel to and perpendicular to the sampling beam. The translucent fixed disk uses a photolithography method to process a black solid cloud particle mask pattern representing cloud particles on its surface, representing cloud particles in the atmosphere. The horizontal distance from the translucent fixed disk to the vertical center axis of the third lens is the sampling distance Z1. A linear array cloud particle detector is placed on the observation plane as the receiving end of the device. The light beam carrying diffraction information is transmitted to the detection unit of the linear array cloud particle detector, and then the host computer observes and stores the real-time recorded particle mask diffraction image. The vertical distance from the linear array cloud particle detector to the horizontal center axis of the third lens is the propagation distance Z2.

[0010] The schematic diagram of the adjustable cloud particle diffraction image generating device is as follows Figure 2 As shown, there are three adjustment modes for this device:

[0011] Adjust the translucent fixed turntable on the sampling plane to a position parallel to the horizontal sampling beam. The movement direction of the translucent fixed turntable is perpendicular to the direction of the horizontal parallel sampling beam. This adjustment method can change the cloud particle mask pattern passing through the sampling beam, thereby obtaining diffraction images with different patterns.

[0012] Adjust the transparent fixed dial on the sampling plane to move horizontally left and right within the horizontal parallel sampling beam, that is, the moving direction is parallel to the beam direction, and change the sampling distance Z1. This method can obtain diffraction images of the same shape but different sampling distances;

[0013] By adjusting the position of the linear array cloud particle detector on the observation plane and moving the linear array cloud particle detector up and down in the vertical direction, that is, changing the diffraction propagation distance Z2, diffraction images of the same shape with different magnifications can be obtained; through one or more combinations of the above adjustment methods, sampling experiments with different shapes, different sampling distances and different magnifications can be realized, and experimental diffraction images of different particle masks can be obtained.

[0014] Furthermore, the first reflector and the second reflector are placed opposite to each other with an angle of 90°. The angle is the angle between the plane where the first reflector is located and the plane where the second reflector is located. The angles between the first reflector and the second reflector and the vertical axis are both 45°.

[0015] Furthermore, the light-transmitting fixed turntable is driven by a precision servo motor and fixed in the sampling area between the first reflector and the second reflector, and at the same time, the diameter position of the collimated parallel light beam passing vertically through the light-transmitting fixed turntable is adjusted. The diameter is a diameter parallel to the horizontal plane, ensuring that the light-transmitting fixed turntable complies with the vertical scanning principle in each movement of cutting the collimated parallel light beam.

[0016] A cloud particle diffraction image classification method based on a cloud particle diffraction image acquisition device is provided. The cloud particle diffraction image classification method includes acquiring a cloud particle diffraction image dataset, building a convolutional neural network classification model for cloud particle diffraction images, and classifying and identifying cloud particle diffraction images. The specific steps are as follows:

[0017] Step S1: Perform a cloud particle sampling simulation experiment using an adjustable cloud particle diffraction image generation device to obtain an experimental diffraction image of the particle mask, as follows:

[0018] Step S1.1, based on common cloud particle shapes, a black solid cloud particle mask pattern representing cloud particles is processed on the transparent fixed turntable using a photolithography method;

[0019] Step S1.2: Fix the device and determine the experimental parameters related to the sampled particle mask, the focal lengths of the first lens, L2, and L3, and the sampling distance Z1 and propagation distance Z2. Start the laser emitter, the linear array cloud particle detector, and the precision servo motor that controls the rotation of the transparent fixed turntable, so that the transparent fixed turntable rotates at high speed to provide a stable tangential speed for the particle mask. When the particle mask on the transparent fixed turntable passes through the collimated parallel light beam, a diffraction effect occurs. The light beam is reflected by the second reflector and propagates vertically downward. The light beam carrying the diffraction information is then magnified by the third lens and propagated to the linear array cloud particle detector. Finally, the host computer observes and stores the real-time recorded diffraction image of the particle mask.

[0020] Furthermore, in step S1.2, because the linear array cloud particle detector reconstructs the particle two-dimensional image by combining effective frames line by line, its working line frequency Lps needs to be adapted to the tangential velocity of the particle mask before reconstructing the target image to avoid image compression or stretching. The matching formula is:

[0021]

[0022] Where v is the tangential velocity of the cloud particle mask pattern on the optically transparent fixed turntable, and R is the equivalent physical size of a single detection pixel of the linear array cloud particle detector under the magnification.

[0023] Step S2, based on the cloud particle sampling simulation experiment conducted by the adjustable cloud particle diffraction image generation device, realizes optical diffraction simulation through the particle diffraction simulation algorithm to obtain the simulated diffraction image of the particle mask on the transparent fixed turntable. The specific steps are as follows:

[0024] Step S2.1, setting the parameters of the particle diffraction simulation algorithm to be consistent with the parameters of the adjustable cloud particle diffraction image generation device, including the focal lengths of the first lens, L2, and L3, the sampling distance Z1, and the propagation distance Z2; setting the simulated particle image to be consistent with the shape of the particle mask pattern on the translucent fixed turntable, both representing cloud particles, and binarizing the particle image generated by mapping software to generate a binary image as the sampling plane image;

[0025] Step S2.2: Perform a Fourier transform-based optical diffraction simulation calculation on the sampled plane image. First, according to the optical diffraction theory, calculate the complex amplitude of the diffraction field on the observation plane. The complex amplitude of the diffraction field is the convolution of the incident light and the optical transfer function. The optical transfer function is calculated based on the light wave propagation distance, wavelength, and boundary conditions. The calculation formula is:

[0026]

[0027] Wherein, wave number k = 2π / λ, λ is the wavelength; z is the diffraction propagation distance; (x0, y0) is a point on the diffraction screen; E(x0, y0) is the complex amplitude distribution of the point (x0, y0) on the sampling plane; U(X, Y) represents the complex amplitude intensity of the diffraction field at the corresponding point (X, Y) on the observation plane; i represents the basic unit of the complex number.

[0028] Using the convolution theorem of Fourier transform, the convolution operation is converted into the product operation of the incident light spectrum and the transfer function spectrum in the frequency domain. At the same time, the continuous form of the optical diffraction equation is numerically discretized. The observation plane image is divided into grids at fixed intervals along the horizontal and vertical directions, and discretized into an M×N matrix. Finally, the discrete complex amplitude distribution is obtained, which is formulated as follows:

[0029]

[0030] Among them, ΔX and ΔY represent the horizontal and vertical sampling intervals of the observation plane image, which must satisfy Δx0 and Δy0 represent the horizontal and vertical sampling intervals on the sampling plane; U(m,n) is the discrete complex amplitude intensity of the observation plane image, (m,n) is the sampling point of the observation plane image, (m0,n0) is the sampling point on the sampling plane image, and m and n are the row and column indices of the matrix. The above results are unified in the spatial domain to obtain the diffraction image G, which is represented by the two-dimensional matrix of the complex amplitude of the diffraction field. The logarithm of its modulus is used to display the simulated diffraction image. The formula is:

[0031] Display=log(1+|G|) (4)

[0032] Step S3, such as Figure 3 As shown in the reliability verification flow chart of the particle diffraction simulation algorithm, the experimental diffraction image obtained in step S1 and the simulated diffraction image obtained in step S2 are used as input, a grayscale value threshold is set, and the main feature map of the experimental diffraction image and the main feature map of the simulated diffraction image are extracted. Then, the structural similarity index of the main feature map of the experimental diffraction image and the main feature map of the simulated diffraction image are calculated, and a structural similarity index threshold T is set. When the structural similarity index of the two images is lower than the threshold T, the simulated diffraction image does not meet the requirements. The sampling interval in the particle diffraction simulation algorithm is adjusted and Gaussian noise is added to match the ambient noise of the experimental device. Step S2 is repeated until the threshold T meets the requirements.

[0033] Furthermore, because the main feature maps of the experimental diffraction image and the main feature maps of the simulated diffraction image ignore the brightness and contrast terms and focus on structural comparison, that is, the covariance term, the structural similarity formula used is:

[0034]

[0035] Where C is a constant, δ x and δ y are the standard deviations of images x and y, δ xy is the standard deviation of images x and y;

[0036] Furthermore, the structural similarity threshold T may be in the range of 0.7 to 0.85.

[0037] In step S4, after the structural similarity index verified in step S3 reaches the threshold value T, it is considered that the simulation environment and the experimental device environment are consistent, and the particle diffraction simulation algorithm can simulate the device well and can be used for simulation experiments in other situations. A large number of simulated diffraction images can be obtained by adjusting the algorithm parameters. Specifically:

[0038] Step S4.1, adjusting the type of graphics in the sampling plane image, including shape and size, to simulate characteristic parameters such as size, shape, and phase of cloud particles in the atmosphere;

[0039] Step S4.2, adjusting the sampling distance, simulating different sampling distances by adjusting the parameter Z1, and obtaining simulated diffraction images at different sampling distances;

[0040] Step S4.3, adjusting the propagation distance, simulating different propagation distances by adjusting the parameter Z2, and obtaining simulated diffraction images under different propagation distances.

[0041] In step S5, the small number of experimental diffraction images obtained in step S1 and the large number of simulated diffraction images obtained in step S4 are collected and classified and integrated according to particle morphological characteristics to construct a multi-category cloud particle diffraction image dataset. Next, the dataset is divided into a training set and a validation set according to a preset ratio, which are used for parameter optimization and performance evaluation of the neural network model, respectively. Finally, to eliminate dimensional differences and improve training stability, all images are normalized, with pixel values uniformly scaled to the range [0, 1], ensuring a balanced number of samples across categories and consistent image sizes. To address the random nature of the three-dimensional position and rotation angle of cloud particles in the atmosphere, data augmentation is used to apply random 0° to 360° rotations, horizontal or vertical translations, and mirror flips to the original images to enhance the model's generalization ability to changes in particle spatial distribution.

[0042] Step S6: Design a convolutional neural network classification model for cloud particle diffraction images. Combined with the diffraction fringe features of the diffraction image, the model must be able to extract multi-level information from low-level edge features to high-level semantic features while also ensuring a lightweight module to reduce computational effort and parameter count. The specific steps are as follows:

[0043] Step S6.1, diffraction feature extraction design, includes convolutional layers and pooling layers. In the convolutional layer, the convolution kernel is locally multiplied and added with the preprocessed diffraction image to extract local image features. Assuming the input diffraction image is x, the convolution kernel weight is w, and the bias is b, the convolution output y(i, j) at position (i, j) can be expressed as:

[0044]

[0045] Where K is the convolution kernel size. Each convolution block uses multiple convolution layers, each layer uses a different number of convolution kernels, and the convolution kernel slides on the input with a step size S. The size of the output feature map after convolution can be calculated by the following formula:

[0046]

[0047] Where I represents the size of the input image, and P is the number of pixels to pad the edges. The padding method is "same." A pooling layer is then introduced, using the max pooling method to reduce the dimensionality of the diffraction image while preserving the main diffraction features. All convolutional layers use ReLU as the activation function, which is expressed as:

[0048] ReLU(x)=max(0,x) (8)

[0049] Step S6.2, classification design, including the fully connected layer and the softmax layer. The feature map after convolution and pooling is flattened into a one-dimensional vector through the flatten operation, which serves as the input of the fully connected layer. Here, the output calculation formula of each neuron is:

[0050]

[0051] Among them, y j is the output of the jth neuron, x i is the i-th feature of the input, w ij is the weight of the ith input to the jth neuron, b j is the bias of the jth neuron. Multiple fully connected layers are set up, and a dropout mechanism is introduced to randomly discard some neurons to prevent overfitting. The ReLU activation function is used after the fully connected layers. The last fully connected layer uses the Softmax activation function, and its number of neurons is equal to the number of cloud particle categories. The output is converted into a probability distribution, indicating the probability that the input diffraction image belongs to each category. The category with the highest probability is selected as the final shape classification result.

[0052] Step S7: Training and optimizing the convolutional neural network model for cloud particle diffraction images. When building the model, an L2 regularization term is introduced to constrain the model's weights and biases, enhancing its generalization capabilities. The Adam optimizer is used to adaptively adjust the learning rate of each parameter. The network parameters are updated through multiple iterations until optimal performance is achieved. After each iteration, the model's performance on unseen data is evaluated against a validation set to prevent overfitting.

[0053] Specifically, the Adam optimizer algorithm combines the advantages of momentum and adaptive learning rate adjustment to accelerate the training process and stabilize convergence. The formula is:

[0054]

[0055] Among them, g t is the gradient, η is the learning rate, β1, β2 are momentum parameters, θ t are model parameters.

[0056] In step S8, the cloud particle diffraction image is identified and classified using the convolutional neural network classification model of the cloud particle diffraction image. The experimental diffraction images that are not used for neural network model training are generated in step S1 as a test set. The images are input into the trained cloud particle diffraction image classification model to obtain classification results. The classification results are further analyzed to obtain the shape, size, and phase information of the particles.

[0057] The beneficial effects of the present invention are:

[0058] (1) The present invention designs an adjustable cloud particle diffraction image generation device, which can obtain particle diffraction images through simple transformations, and can solve the problem of unclear particle diffraction images in different situations. At the same time, a particle diffraction simulation algorithm is proposed to realize the diffraction image simulation of particles of arbitrary shapes, and the reliability of the simulation is verified, thereby improving the efficiency and accuracy of obtaining diffraction images of cloud particles of arbitrary shapes;

[0059] (2) The dataset for training the convolutional neural network of the present invention is self-generated. By adjusting the parameters, a large amount of different data can be generated. It is no longer limited to a real in-situ measurement image of a certain instrument as a dataset. This is conducive to more comprehensive extraction of particle diffraction stripe structural feature information and helps to promote the classification research of particle images obtained by optical array imaging devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of a cloud particle diffraction image acquisition and classification method proposed in the present invention;

[0061] Figure 2 This is a schematic diagram of the adjustable cloud particle diffraction image generation device proposed in the present invention;

[0062] Figure 3 This is a flow chart for reliability verification of the particle diffraction simulation algorithm proposed in the present invention;

[0063] Figure 4 Graphs for a specific embodiment of the present invention: (a) shows the main features of a rectangular experimental diffraction image, the main features of a simulated diffraction image, and a difference between the two; (b) shows the main features of a hexagonal experimental diffraction image, the main features of a simulated diffraction image, and a difference between the two; (c) shows the main features of a circular experimental diffraction image, the main features of a simulated diffraction image, and a difference between the two.

[0064] Figure 5 In the specific embodiment of the present invention, (a) is a rectangular experimental diffraction image; (b) is a rectangular simulated diffraction image; (c) is a hexagonal experimental diffraction image and a simulated diffraction image; (d) is a snowflake-shaped simulated diffraction image;

[0065] Figure 6Schematic diagram of a convolutional neural network classification model for cloud particle diffraction images in a specific embodiment of the present invention;

[0066] Figure 7 The test results of the first 10 training runs in a specific embodiment of the present invention are shown in Figure 1. (a) shows the accuracy of the training set and the validation set, and (b) shows the loss of the training set and the validation set.

[0067] Figure 8 This is a diagram showing the recognition results of 7 groups of data in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0069] The present invention provides a cloud particle diffraction image acquisition device and classification method. An adjustable cloud particle diffraction image generation device is provided. In combination with a cloud particle diffraction simulation algorithm, the efficiency of acquiring cloud particle diffraction images is improved. At the same time, a cloud particle diffraction image classification model is established to achieve rapid and accurate recognition of cloud particle shapes. The method includes the following steps:

[0070] Step S1: Build an adjustable cloud particle diffraction image generation device. A laser emitter emits light through the first lens, generating a collimated beam expansion path through the L1 and L2 lenses, generating a parallel beam of fixed width. The area between the two reflectors serves as the particle sampling area. Z1 is the sampling distance, the horizontal distance from the translucent fixed-point turntable to the vertical center axis of the second lens. The L3 lens primarily magnifies the diffraction image within the propagation distance Z2, magnifying it to a certain magnification before displaying it on the linear cloud particle array detector.

[0071] In the embodiment of the present invention, the wavelength of the laser transmitter selected is 520 nm, and the Gaussian beam radius is 11 mm;

[0072] In the embodiment of the present invention, the first lens L2 is selected to generate a parallel collimated light beam, and the third lens has a focal length of 40 mm;

[0073] In the embodiment of the present invention, the selected sampling distance Z1 is 45 mm and the propagation distance Z2 is 300 mm;

[0074] In the embodiment of the present invention, the selected linear array cloud particle detector has a resolution size of 2048×1 and a pixel size of 14 μm.

[0075] Step S2, performing a cloud particle sampling simulation experiment using an adjustable cloud particle diffraction image generating device, includes the following steps:

[0076] Step S2.1: engraving a cloud particle mask pattern on the translucent fixed turntable, using a two-dimensional plane pattern to represent the shape of cloud particles, such as a circle representing spherical water droplets in natural clouds, a rectangle representing columnar ice crystals, a hexagon representing plate-like ice crystals, an H-shaped pattern representing capped columnar ice crystals, or other special-shaped patterns to represent complex ice crystal shapes;

[0077] Step S2.2: Drive the light-transmitting fixed turntable to rotate, adjust the operating line frequency of the linear array cloud particle detector to match the turntable rotation speed, and then collect the diffraction image. The matching formula is:

[0078]

[0079] Where v is the tangential velocity of the cloud particle mask pattern on the optically transparent fixed turntable, and R is the equivalent physical size of a single detection pixel of the linear array cloud particle detector under the magnification;

[0080] In the embodiment of the present invention, the selected translucent positioning turntable is an optical glass circular turntable with a diameter of 200 mm, a thickness of 4 mm, a smooth surface and high light transmittance; in the embodiment of the present invention, the selected sampling particle mask pattern is a rectangle with a length of 60 μm and a width of 30 μm, a regular hexagon with a circumscribed circle diameter of 60 μm and a circle with a diameter of 60 μm; the turntable tangential speed is 1 m / s, and the detector working line frequency is 460 kHz.

[0081] Step S3: Determine the independent variable parameters of the particle diffraction simulation algorithm based on the parameters of the experimental device, determine the sampling plane pattern input into the algorithm based on the cloud particle mask pattern on the turntable, and then calculate the complex amplitude of the diffraction field on the observation plane based on the optical diffraction theory. The complex amplitude of the diffraction field is the convolution of the incident light and the optical transfer function. The optical transfer function is calculated based on the light wave propagation distance, wavelength, and boundary conditions. The calculation formula is:

[0082]

[0083] Wherein, wave number k = 2π / λ, where λ is the wavelength; z is the diffraction propagation distance; (x0, y0) is a point on the diffraction screen; E(x0, y0) is the complex amplitude distribution of point (x0, y0) on the sampling plane; U(X, Y) represents the complex amplitude intensity of the diffraction field at the corresponding point (X, Y) on the observation plane; and i represents the basic unit of the complex number. The observation plane image is divided into a grid at fixed intervals in the horizontal and vertical directions, discretized into a 512×512 matrix. The discrete complex amplitude distribution formula is as follows:

[0084]

[0085] Among them, ΔX and ΔY represent the horizontal and vertical sampling intervals of the observation plane image, which must satisfy Δx0 and Δy0 represent the horizontal and vertical sampling intervals on the sampling plane; U(m,n) is the discrete complex amplitude intensity of the observation plane image, (m,n) is the sampling point of the observation plane image, (m0,n0) is the sampling point on the sampling plane image, and m and n are the row and column indices of the matrix. The above results are unified in the spatial domain to obtain the diffraction image G, which is represented by the two-dimensional matrix of the complex amplitude of the diffraction field. The logarithm of its modulus is used to display the simulated diffraction image. The formula is:

[0086] Display=log(1+|G|) (4)

[0087] Step S4: Compare the diffraction image of the experimental device with the simulated diffraction image, and calculate the structural similarity of the main diffraction feature images of the experiment and the simulation to see whether they meet the threshold T. The structural similarity formula used is:

[0088]

[0089] Where C is a constant, δ x and δ y are the standard deviations of images x and y, δ xy is the standard deviation of the images x and y.

[0090] In the embodiment of the present invention, the structural similarity threshold T is set to 0.75, and the comparison shapes selected are rectangle, hexagon and circle. Figure 4 As shown, (a) shows the main features of the rectangular experimental diffraction image, the main features of the simulated diffraction image, and the difference between the two; (b) shows the main features of the hexagonal experimental diffraction image, the main features of the simulated diffraction image, and the difference between the two; (c) shows the main features of the circular experimental diffraction image, the main features of the simulated diffraction image, and the difference between the two. When the calculated value is less than T, it indicates that the main diffraction features are significantly different. You can adjust the sampling interval in the particle diffraction simulation algorithm, add Gaussian noise to match the ambient noise of the experimental device, and increase the threshold T to meet the requirements.

[0091] Step S5: after the threshold is met, adjusting the parameters of the particle diffraction simulation algorithm to obtain a large number of simulated diffraction images;

[0092] In the embodiment of the present invention, the algorithm parameters can be adjusted to generate diffraction images of three types of graphics: rectangle, regular hexagon, and circle. Each type of image has 9324 images. The specific operations are as follows:

[0093] Step S5.1: Adjust the diffraction screen pattern to simulate the shape of cloud particles. The patterns are rectangular, hexagonal, and circular. The particle size (the length of the long side for rectangles, the diameter of the circumscribed circle for hexagons, and the diameter of the circumscribed circle for circles) can be adjusted from 10 μm to 190 μm in 5 μm increments.

[0094] Step S5.2, adjusting the sampling distance Z1, the adjustment range is 20mm-75mm, and the adjustment step is 5mm;

[0095] Step S5.3, adjusting the propagation distance Z2, with an adjustment range of 150 mm to 350 mm and an adjustment step of 10 mm;

[0096] In the embodiment of the present invention, Figure 5 Shown are some particle diffraction images, including experimental and simulated rectangular and hexagonal diffraction images of different particle sizes, as well as simulated snowflake-shaped diffraction images. The specific diffraction characteristics of each type of image are significantly different. For example, circular particles exhibit alternating light and dark circular stripes, with the center believed to be a Poisson bright spot; hexagonal particles exhibit hexagonal stripes and hexagonal bright spots; rectangular particles have stripes and bright spots arranged along their long edges and a tendency to split; and snowflake-shaped particles exhibit complex diffraction stripe patterns, with a certain similarity between their outline and diffraction structure and those of snowflakes.

[0097] Step S6, combining a large number of particle diffraction images obtained by the above method with diffraction images generated by the experimental device as a data set, and obtaining a classification model for diffraction images of different shapes through convolutional neural network training;

[0098] In the embodiment of the present invention, three two-dimensional images with typical features, namely, a circle, a regular hexagon, and a rectangle, are used to train a convolutional neural network, and a well-trained model is finally obtained. The specific steps are as follows:

[0099] Step S6.1, data loading and preprocessing. Diffraction images of the same data volume for three patterns, circular, regular hexagonal, and rectangular, are obtained from the experimental device and simulation. The image size is unified to 192×192 pixels, and the original pixel value (0, 255) is converted to the interval (0, 1). 80% of the data set is divided into a training set to train the network model, and the remaining 20% is used as a validation set. The test data comes from samples that have not been seen by the neural network generated by the experimental device. Since the angle and position of cloud particles in nature are constantly changing, the collected images are randomly rotated, translated, and flipped to enhance the generalization ability of the model and expand the data set.

[0100] Step S6.2: Construct a convolutional neural network classification model for cloud particle diffraction images to automatically extract image features and classify cloud particle diffraction images. This network classification model mainly includes a convolutional layer, an activation function, a pooling layer, a dropout layer, a fully connected layer, and an output layer. The specific steps are as follows:

[0101] In the embodiment of the present invention, Figure 6 As shown in the figure, the convolutional neural network model consists of 10 convolutional layers and 4 pooling layers. The convolutional layer performs convolution operations on the input image through the convolution kernel and uses a 3×3 convolution kernel to extract local features:

[0102]

[0103] Where I(x,y) is the input image, K(i,j) is the convolution kernel, and O(x,y) is the convolution result. The network structure is organized in a block manner. Each convolution block uses multiple convolution layers, each layer uses a different number of convolution kernels, and the convolution kernel slides on the input with a step size S. The size of the output feature map after convolution can be calculated by the following formula:

[0104]

[0105] Here, I represents the size of the input image, and P is the number of pixels padded to the edges. A pooling layer is then introduced, using the max pooling method to reduce the dimensionality of the diffraction image while retaining the main diffraction features. The first convolutional block consists of two 3×3 convolutional layers, using 64 convolution kernels, a stride of 1, a padding of "same", and a ReLU activation function; the second convolutional block consists of two 3×3 convolutional layers, using 128 convolution kernels; the third convolutional block consists of three 3×3 convolutional layers, using 256 convolution kernels; and the fourth convolutional layer consists of three 3×3 convolutional layers, using 512 convolution kernels. Each convolutional block is followed by a max pooling layer with a pooling window size of 2×2 and a stride of 2. This is used to downsample and gradually reduce the size of the feature map, thereby extracting more discriminative features. All convolutional and fully connected layers use ReLU as the activation function, and its mathematical expression is:

[0106] ReLU(x)=max(0,x) (8)

[0107] The feature map after convolution and pooling is flattened into a one-dimensional vector through the Flatten operation and used as the input of the fully connected layer. The output calculation formula of each neuron here is:

[0108]

[0109] Among them, y j is the output of the jth neuron, x i is the i-th feature of the input, wij is the weight of the ith input to the jth neuron, b j is the bias of the jth neuron. Multiple fully connected layers are set up, and a dropout mechanism is introduced with a dropout ratio of 0.5, randomly discarding some neurons to prevent overfitting. The number of neurons in the final fully connected layer is equal to the number of cloud particle categories. The final output layer uses a softmax activation function to convert the score into a probability distribution. After the fully connected layer, the features are mapped to three output nodes, and the probability distributions of three different shapes are calculated. The probability of each category is calculated, with a hexagonal probability of 0.78, a circular probability of 0.19, and a rectangular probability of 0.03, identifying the diffraction image as a hexagon.

[0110] Step S6.4: Model training and optimization. When building the model, an L2 regularization term is introduced to constrain the model's weights and biases, enhancing its generalization capabilities. The Adam optimizer is used to adaptively adjust the learning rate of each parameter. The network parameters are updated through multiple iterations until optimal performance is achieved. After each epoch, the model's performance on unseen data is evaluated against a validation set to prevent overfitting.

[0111] Specifically, the Adam optimizer algorithm combines the advantages of momentum and adaptive learning rate adjustment to accelerate the training process and stabilize convergence. The formula is:

[0112]

[0113] Among them, g t is the gradient, η is the learning rate, β1, β2 are momentum parameters, θ t are model parameters. Figure 7 The following figure shows the test results of the training set and validation set in the first 10 cycles. (a) is the accuracy result graph of the training set and validation set, and (b) is the loss result graph of the training set and validation set.

[0114] Step S6.5: Model Saving and Prediction. The trained model is saved and used. Images of particles under certain conditions are randomly generated using the experimental apparatus. The model is used to classify new cloud particle diffraction images, outputting the probability of each shape category, and determining the final predicted shape based on the maximum probability value.

[0115] Step S7: statistics of classification results: the diffraction images generated by the experimental device and not participating in the neural network learning are input into the diffraction classification model, and the statistical output results are obtained.

[0116] In the embodiment of the present invention, some pictures are randomly selected from the diffraction images and divided into 7 groups of data. Each group contains three types of graphics: circle, hexagon and rectangle. The number of diffraction images of each graphic is 20, 40, 60, 80, 100, 120 and 140 respectively. The classification model accurately predicts the number of images as shown in the following figure. Figure 8 As shown in the figure, particle phases can be analyzed based on their shape: circles are considered water droplets, while hexagons and rectangles are considered ice crystals. Then, a contour extraction algorithm is used to extract the contours of correctly identified and classified particles. Further, particle size parameters can be measured: the length of the long side for rectangles, the diameter of the circumscribed circle for hexagons, and the diameter of the circle. Statistical analysis of large amounts of data and spectral analysis can ultimately predict the state of the entire cloud particle dataset.

[0117] Those skilled in the art will appreciate that the above-described embodiments merely represent several embodiments of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the appended claims.

Claims

1. A cloud particle diffraction image acquisition device, characterized in that: The cloud particle diffraction image acquisition device is an adjustable cloud particle diffraction image generation device, comprising a detection light path and an imaging plane; The detection optical path includes a laser emitter, a lens group, and a reflector group; the imaging plane includes a sampling plane and an observation plane. A transparent fixed turntable is placed on the sampling plane, and a linear array cloud particle detector is placed on the observation plane. The linear array cloud particle detector serves as a receiving end. The detection optical path generates a sampling beam and propagates diffraction information through the beam. After the transparent fixed turntable passes through the sampling beam, a diffraction effect occurs to generate a propagation beam containing diffraction information. The propagation beam receives the beam information through the linear array cloud particle detector, which is used to identify whether there is any shadow change and diffraction fringe structure in the propagation beam. The specific structure is as follows: In the detection optical path: the emitting end of the laser emitter is vertically upward and is used to emit a laser beam; the lens group includes a first lens, a second lens, and a third lens; the reflector group includes a first reflector and a second reflector, wherein a movable light-transmitting fixed dial is provided between the first reflector and the second reflector, the first lens and the second lens are arranged between the laser emitter and the first reflector, and the first lens is close to the laser emitter; the function of the first reflector is to change the direction of the vertical upward light beam emitted by the laser emitter into a horizontal light beam, and the second reflector reflects the horizontal light beam into a vertical downward direction; the third lens is arranged between the second reflector and the linear array cloud particle detector, and the third lens realizes a microscopic magnification function; A light-transmitting rotating disk is placed on the sampling plane, and the light-transmitting rotating disk can move in a direction parallel to the sampling beam and in a direction perpendicular to the sampling beam; a black solid cloud particle mask pattern representing cloud particles is processed on the surface of the light-transmitting rotating disk by photolithography, representing cloud particles in the atmosphere; The horizontal distance from the transparent optical positioning dial to the vertical center axis of the third lens is the sampling distance Z1; A linear array cloud particle detector is placed on the observation plane, and the vertical distance from the linear array cloud particle detector to the horizontal central axis of the third lens is the propagation distance Z2.

2. The cloud particle diffraction image acquisition device according to claim 1, characterized in that: The centers of the first lens and the second lens are located on the central axis of the laser beam emitted by the laser emitter, and their function is to generate a collimated, expanded and fixed-width parallel light beam as a sampling beam; the center of the second reflector is located on the central axis of the laser beam.

3. The cloud particle diffraction image acquisition device according to claim 1, characterized in that: The adjustable cloud particle diffraction image generating device has three adjustment modes: Adjust the translucent fixed turntable on the sampling plane to a position parallel to the horizontal sampling beam. The movement direction of the translucent fixed turntable is perpendicular to the direction of the horizontal parallel sampling beam. This adjustment method can change the cloud particle mask pattern passing through the sampling beam, thereby obtaining diffraction images with different patterns. Adjust the transparent fixed dial on the sampling plane to move horizontally left and right in the horizontal parallel sampling beam, that is, the moving direction is parallel to the beam direction, and change the sampling distance Z1. This method can obtain diffraction images of the same shape but different sampling distances; Adjust the position of the linear cloud particle detector on the observation plane and move it up and down in the vertical direction, that is, change the diffraction propagation distance Z2, and obtain diffraction images of the same shape with different magnifications; By using one or more combinations of the above-mentioned adjustment methods, sampling experiments with different shapes, different sampling distances and different magnifications can be realized, and experimental diffraction images of different particle masks can be obtained.

4. The cloud particle diffraction image acquisition device according to claim 1, characterized in that: The first reflector and the second reflector are placed opposite to each other with an angle of 90°. The angle is the angle between the plane where the first reflector is located and the plane where the second reflector is located. The angles between the first reflector and the second reflector and the vertical axis are both 45°.

5. The cloud particle diffraction image acquisition device according to claim 1, characterized in that: The light-transmitting fixed turntable is driven to rotate by a servo motor.

6. A cloud particle diffraction image classification method implemented based on the cloud particle diffraction image acquisition device according to any one of claims 1 to 5, characterized in that: The cloud particle diffraction image classification method includes obtaining a cloud particle diffraction image dataset, building a convolutional neural network classification model for the cloud particle diffraction images, and classifying and identifying the cloud particle diffraction images, including the following steps: Step S1, performing a cloud particle sampling simulation experiment using an adjustable cloud particle diffraction image generating device to obtain an experimental diffraction image of a particle mask; Step S2, based on a cloud particle sampling simulation experiment conducted by the adjustable cloud particle diffraction image generation device, optical diffraction simulation is realized by using a particle diffraction simulation algorithm to obtain a simulated diffraction image of the particle mask on the transparent fixed turntable; Step S3: using the experimental diffraction image obtained in step S1 and the simulated diffraction image obtained in step S2 as input, setting a grayscale value threshold, extracting the main feature map of the experimental diffraction image and the main feature map of the simulated diffraction image, and then calculating the structural similarity index of the two main feature maps; setting a structural similarity index threshold T; when the structural similarity index of the two main feature maps is lower than the threshold T, the simulated diffraction image does not meet the requirements, adjusting the sampling interval in the particle diffraction simulation algorithm and adding Gaussian noise to match the ambient noise of the experimental device, and repeating step S2 until the threshold T meets the requirements; In step S4, after the structural similarity index verified in step S3 reaches a threshold value T, it is considered that the simulation environment and the experimental device environment are consistent, and a large number of simulated diffraction images are obtained by adjusting the simulation algorithm parameters in step S2; a large number of simulated diffraction images are obtained by the following method: Step S4.1, adjusting the type of graphics in the sampling plane image, including shape and size, to simulate characteristic parameters such as size, shape, and phase of cloud particles in the atmosphere; Step S4.2, adjusting the sampling distance, simulating different sampling distances by adjusting the parameter Z1, and obtaining simulated diffraction images at different sampling distances; Step S4.3, adjusting the propagation distance, simulating different propagation distances by adjusting the parameter Z2, and obtaining simulated diffraction images at different propagation distances; In step S5, the experimental diffraction images obtained in step S1 and the simulated diffraction images obtained in step S4 are first collected and classified and integrated according to the particle morphological characteristics to construct a multi-category cloud particle diffraction image dataset. Next, the multi-category cloud particle diffraction image dataset is divided into a training set and a validation set according to a preset ratio, which are used for parameter optimization and performance evaluation of the neural network model, respectively. Then, the training set and the validation set are normalized and data augmented. Step S6, designing a convolutional neural network classification model for cloud particle diffraction images, combining the diffraction fringe features of the diffraction images; Step S7: Training and optimizing a convolutional neural network classification model for cloud particle diffraction images. When constructing the convolutional neural network classification model, an L2 regularization term is introduced to constrain the weights and biases of the convolutional neural network classification model, and the learning rate of each parameter is adaptively adjusted using the Adam optimizer. The convolutional neural network classification model parameters are updated through multiple iterations until optimal performance is achieved. After each cycle, the convolutional neural network classification model is evaluated on unseen data based on a validation set to prevent overfitting. In step S8, the cloud particle diffraction image is identified and classified using the convolutional neural network classification model of the cloud particle diffraction image. The experimental diffraction image that was not used for training is generated in step S1 as a test set. The experimental diffraction image is input into the trained cloud particle diffraction image classification model to obtain the classification result. The classification result is further analyzed to obtain the shape, size, and phase information of the particles.

7. The cloud particle diffraction image classification method implemented by the cloud particle diffraction image acquisition device according to claim 6 is characterized in that: The step S1 is specifically as follows: Step S1.1, based on the common cloud particle shape, a black solid cloud particle mask pattern representing cloud particles is processed on a transparent fixed turntable using a photolithography method; Step S1.2, fix the device and determine the sampled particle mask, the focal length of the first lens, L2, and L3, and the experimental parameters related to the sampling distance Z1 and the propagation distance Z2; start the laser emitter and the linear cloud particle detector, and make the transparent fixed turntable rotate at high speed to provide a stable tangential speed for the particle mask. When the particle mask on the transparent fixed turntable passes through the collimated parallel light beam, a diffraction effect occurs, and the light beam is reflected by the second reflector and propagates vertically downward. Then, the light beam carrying the diffraction information is magnified by the third lens and propagated to the linear cloud particle detector. Finally, the diffraction image of the particle mask is observed and stored in real time on the host computer.

8. The cloud particle diffraction image classification method implemented by the cloud particle diffraction image acquisition device according to claim 6 is characterized in that: The step S2 is specifically as follows: Step S2.1: Setting the parameters of the particle diffraction simulation algorithm to be consistent with the parameters of the adjustable cloud particle diffraction image generation device, including the focal lengths of the first lens, L2, and L3, the sampling distance Z1, and the propagation distance Z2; setting the simulated particle image to be consistent with the shape of the particle mask pattern on the translucent fixed turntable, both representing cloud particles; generating the particle image using mapping software and then binarizing it, and using the generated binary image as the sampling plane image; Step S2.2, performing optical diffraction simulation calculation based on Fourier transform on the sampling plane image; First, the complex amplitude of the diffraction field on the observation plane is calculated. The complex amplitude of the diffraction field is the convolution of the incident light and the optical transfer function. The optical transfer function is calculated based on the light wave propagation distance, wavelength and boundary conditions. The calculation formula is: Wherein, wave number k = 2π / λ, λ is the wavelength; z is the diffraction propagation distance; (x0, y0) is a point on the diffraction screen; E(x0, y0) is the complex amplitude distribution of the point (x0, y0) on the sampling plane; U(X, Y) represents the complex amplitude intensity of the diffraction field at the corresponding point (X, Y) on the observation plane; i represents the basic unit of the complex number; Secondly, using the convolution theorem of Fourier transform, the convolution operation is converted into the product operation of the incident light spectrum and the transfer function spectrum in the frequency domain. At the same time, the continuous form of the optical diffraction equation is numerically discretized. The observation plane image is divided into grids at fixed intervals along the horizontal and vertical directions, and discretized into an M×N matrix. Finally, the discrete complex amplitude distribution is obtained, which is formulated as follows: Among them, ΔX and ΔY represent the horizontal and vertical sampling intervals of the observation plane image, which must satisfy Δx0 and Δy0 represent the horizontal and vertical sampling intervals on the sampling plane; U(m,n) is the discrete complex amplitude intensity of the observation plane image, (m,n) is the sampling point of the observation plane image, (m0,n0) is the sampling point on the sampling plane image, and m and n are the row and column indices of the matrix; the above results are unified into the spatial domain to obtain the diffraction image G, which is represented as a two-dimensional matrix of the complex amplitude of the diffraction field, and the logarithm of its modulus is used to display the simulated diffraction image.

9. The cloud particle diffraction image classification method implemented by the cloud particle diffraction image acquisition device according to claim 6, characterized in that: In the step S3: The calculation formula of the structural similarity is: Where C is a constant, δ x and δ y are the standard deviations of images x and y, δ xy is the standard deviation of images x and y; The structural similarity threshold T ranges from 0.7 to 0.

85.

10. The cloud particle diffraction image classification method implemented by the cloud particle diffraction image acquisition device according to claim 6, characterized in that: The step S6 is specifically as follows: Step S6.1, design the diffraction feature extraction part, including the convolution layer and the pooling layer; In the convolution layer, the convolution kernel is locally multiplied and added with the preprocessed diffraction image to extract local image features; Then the pooling layer is introduced and the maximum pooling method is selected; all convolutional layers and fully connected layers use ReLU as the activation function; Step S6.2, design the classification part; The feature map after convolution and pooling operations is flattened into a one-dimensional vector through the Flatten operation and used as the input of the fully connected layer. Multiple fully connected layers are set up, and the Dropout mechanism is introduced to randomly discard some neurons to prevent overfitting. The ReLU activation function is used after the fully connected layer. The last fully connected layer adopts the Softmax activation function, and its number of neurons is equal to the number of cloud particle categories. The output is converted into a probability distribution, indicating the probability that the input diffraction image belongs to each category. The category with the highest probability is selected as the final shape classification result.

Images