CN116721302A - A lightweight network-based ice and snow crystal particle image classification method - Google Patents

Abstract

The application relates to an ice and snow crystal particle image classification method based on a lightweight network, which belongs to the technical field of computer vision and comprises the following steps: training the built network model by using a training sample data set containing a plurality of classified ice and snow crystal particle images, classifying, testing and verifying the ice and snow crystal particle images of the network model by using a testing data set of the ice and snow crystal particle images, and selecting model parameters with the best performance for storage; inputting the obtained ice and snow crystal particle image into a trained and tested network model, and carrying out feature extraction and category reasoning calculation on the ice and snow crystal particle sub-image to obtain a predicted category so as to realize automatic classification of the ice and snow crystal particle image. The depth separable cavity convolution layer can effectively fuse global and local characteristics, and has good classification effect on detail characteristics of ice and snow crystal grains with small difference in scale and structure.

Description

A lightweight network-based ice and snow crystal particle image classification method

Technical field

The invention relates to the field of computer vision technology, and in particular to a lightweight network-based ice and snow crystal particle image classification method.

Background technique

Ice and snow crystal particles are an important atmospheric aerosol and have an important impact on climate and weather. Correctly classifying ice and snow crystal particle images helps researchers better understand and study weather patterns, thereby more accurately predicting weather changes and improving the accuracy of weather forecasts. Secondly, ice and snow crystal particles can change the optical properties in the atmosphere, sky color, human visual experience, etc. This information all affects people's observation ability and sensory experience. The correct classification of ice and snow crystal particles can help relevant workers better understand these influencing factors and formulate more reasonable meteorological policies to protect the public's visual experience and health.

In recent years, with the rapid development of deep learning, artificial intelligence technology has shown great application potential in the processing and analysis of images and big data. Its application in atmospheric science has also attracted the attention of researchers, especially artificial intelligence. Deep learning technology has been applied to the automatic classification of ice and snow crystal particle shapes. The classic one is to use the pre-trained residual network ResNet152 to identify the shape of ice and snow crystal particles. In the ICDC (Ice Crystals Database in China, ICDC) ice and snow crystal data set containing 7282 images in 10 categories, 96% recognition was achieved. accuracy, but the prediction effect for small sample data is poor. In order to improve the possible small sample prediction problems of residual convolutional networks, someone proposed a deep learning method based on the embedded hypergraph convolution layer. This method uses the structure of the embedded hypergraph to construct feature relationships from local and global graph structures. This improves its performance in small samples and achieves better classification results when the sample distribution is unbalanced.

However, the existing technology has the following problems: 1. Lack of effective extraction of morphological characteristics of ice and snow crystal particles: the existing ice and snow crystal particle classification methods based on deep convolutional networks have large intra-class differences and inter-class differences for ice and snow crystal particles. The ability to deal with small fine-grained and unbalanced samples is limited, making it difficult to fully extract and express the characteristic information in ice and snow crystal particle images, resulting in errors and reducing the accuracy and reliability of classification. 2. Slow reasoning speed: The existing ice and snow crystal particle classification method based on deep convolutional network has problems such as too deep depth and excessive computational IO overhead, which leads to slow reasoning speed. For the task of quickly identifying and classifying ice and snow crystal particles, it is required. , these methods still have a lot of room for optimization in terms of inference speed. 3. Large parameter scale: The existing ice and snow crystal particle classification method based on deep convolutional network has a large number of parameters, which directly makes it difficult to embed the algorithm into terminal equipment and limits the practical application of the algorithm.

It should be noted that the information disclosed in the above background section is only used to enhance understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to those of ordinary skill in the art.

Contents of the invention

The purpose of the present invention is to overcome the shortcomings of the existing technology, provide a lightweight network-based ice and snow crystal particle image classification method, and solve the shortcomings of the existing technology.

The purpose of the present invention is achieved through the following technical solutions: a lightweight network-based ice and snow crystal particle image classification method, the classification method includes:

Use the training sample data set containing multiple categories of ice and snow crystal particle images to train the built lightweight network UDiNet, and use the test data set of ice and snow crystal particle images to classify the lightweight network UDiNet. Test and verify, select the model parameters with the best performance to save;

Input the acquired ice and snow crystal particle images into the trained and tested lightweight network UDiNet, perform feature extraction and category inference calculation on the ice and snow crystal particle images, and obtain a predicted category to achieve automatic classification of the ice and snow crystal particle images. .

The backbone network of the lightweight network UDiNet is a channel reorganization network. The lightweight network UDiNet includes a channel reorganization convolution unit, a lightweight attention mechanism layer and a depth-separable atrous convolution layer;

The channel reorganization convolution unit is used to construct a channel reorganization network, and rearranges the information of each channel group in the feature map during group convolution to form a channel rearrangement mechanism and increase information fusion between channels; the lightweight The level-level attention mechanism layer is used to weight the importance of feature maps. The learned attention weights allow the lightweight network UDiNet to automatically select and focus on features that are more important to the current task;

The channel reorganization network is constructed using basic units composed of grouped convolution and pointwise convolution, and a grouped channel rearrangement mechanism is introduced in the grouped convolution to make the output feature map of the last layer of basic unit of the channel reorganization network as depth separable Input to atrous convolutional layer.

The depth-separable atrous convolution layer uses a residual network structure, which includes a depth-separable atrous convolution branch and short connection analysis. The two branches are respectively based on the feature map and channel number generated by the last basic unit of the channel reorganization network. Standard, the depth-separable atrous convolution branch is used to extract global features of the input feature map, and the short connection branch is used to collect local information of the input feature map, and the output feature maps of the two branches are connected in the channel dimension to achieve channel Extension.

The global feature extraction of the input feature map through the depth-separable atrous convolution branch, and the collection of local information on the input feature map through the short connection branch specifically include:

A1. Use a deep convolution with a convolution kernel size of 3×3 to learn the feature information of the ice and snow crystal particle image details on a single feature map, and then use a point-by-point convolution with a convolution kernel size of 1×1 to learn the feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images;

A2. Upgrade the dimension through point-by-point convolution with a convolution kernel size of 1×1 to obtain more feature information of different channels of the ice and snow crystal particle image, and then fill two layers of pixels with a value of 0 around the feature map to While expanding the feature map of ice and snow crystal particles, no noise will interfere with the feature information;

A3. Use deep convolution with a convolution kernel size of 3×3 to learn the overall feature information of the ice and snow crystal particle feature image on a single feature map, and use point-by-point convolution with a convolution kernel size of 1×1 to learn the overall feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images;

A4. Connect the calculation results of steps A1 and A2 along the channel dimension as the output feature map of the depth-separable atrous convolution layer.

The training of the lightweight network UDiNet includes the following:

The parameters of the convolutional neural network pre-trained on the large image data set ImageNet are used for the initialization of the convolutional layer in the backbone network of the lightweight network UDiNet;

Using the Pytorch framework, Epochs is set to 100, and Batch Size is set to 128;

Use the AdamW optimization algorithm and set the learning rate to 0.0001;

In the public data set ICDC containing multiple categories of ice and snow crystal particle images, 80% of the image data for each category were randomly selected for training.

The test data set of ice and snow crystal particle images is used to classify and test the ice and snow crystal particle images of the lightweight network UDiNet, and the model parameters with the best performance are selected for storage, which specifically includes the following:

At the end of each Epoch training, use the same test sample data set to test the lightweight network UDiNet, and evaluate the classification results obtained from the test;

by formula Calculate the classification accuracy of the lightweight network UDiNet for all test sample data sets, where ŷ _i represents the category of the i-th image predicted by the lightweight network UDiNet, y _i is the corresponding true category value, and M is the test time The total number of images input to the network model, eq(y _i, ŷ _i ) is an equality function. If and only if y _i approaches ŷ _i , the better the prediction performance of the lightweight network UDiNet is, the closer the accuracy is to 1 ;

The data used for testing is the remaining 20% of the image data except the training sample data set in the public ICDC data set containing multiple categories of ice and snow crystal particle images;

Save the lightweight network UDiNet with the highest calculated accuracy in multiple tests.

Multiple classifications of the ice and snow crystal particle images include rosettes, rosettes, long columns, short columns, hollow columns, spheres, small irregularities, plates, hexagonal snowflakes and complex shapes; depth-separable hollow volumes The multilayer output feature map is input into the convolution classifier, and then the softmax function is used to normalize the output probability to complete the task of classifying 10 categories of ice and snow crystal particle images.

The present invention has the following advantages: a lightweight network-based ice and snow crystal particle image classification method, which can effectively integrate global and local features through a depth-separable hole convolution layer. For ice and snow crystal particles with tiny differences in scale and structure, Different detailed features can have better classification effect. Secondly, by using depthwise separable convolution, ordinary convolution is divided into two steps, which greatly reduces the calculation amount and parameter amount of the model; the depthwise separable atrous convolution branch uses The inverse bottleneck structure can map low-dimensional feature maps to high dimensions when the number of input feature map channels is small, and extract features from the data in high dimensions, thereby reducing losses in the process of extracting feature map information; using lightweight The advanced model architecture design greatly reduces the amount of model parameters and speeds up model inference, and is expected to be deployed and applied in various edge computing terminals; by designing the convolution classifier, the spatial information is retained while the translation of the input is not affected. Variability has a certain degree of robustness. Compared with the fully connected classifier, the convolution classifier not only improves the parameter efficiency of the model but also enhances the generalization ability of the model.

Description of the drawings

Figure 1 is a schematic flow diagram of the present invention;

Figure 2 is a schematic structural diagram 1 of the lightweight network UDiNet of the present invention;

Figure 3 is a schematic structural diagram 2 of the lightweight network UDiNet of the present invention;

Figure 4 is a schematic structural diagram 3 of the lightweight network UDiNet of the present invention;

Figure 5 is a schematic structural diagram 4 of the lightweight network UDiNet of the present invention;

Figure 6 is a schematic diagram of a depthwise separable atrous convolution layer.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only These are part of the embodiments of this application, but not all of them. The components of the embodiments of the present application generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Accordingly, the following detailed description of the embodiments of the application provided in connection with the appended drawings is not intended to limit the scope of the application as claimed, but rather to merely represent selected embodiments of the application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without any creative work shall fall within the scope of protection of this application. The present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention specifically relates to a method for classifying ice and snow crystal particle images based on the lightweight network UDiNet. This network actively reduces the amount of parameters and calculations required by the network through deep separable volumes, and solves the pain points of the existing ice and snow crystal particle classification algorithm's excessive parameter scale and high calculation amount. Secondly, UDiNet uses a depth-separable atrous convolution to achieve rapid fusion of spatial information and positional information in the feature map, which increases the receptive field of the network and enables the network to quickly integrate classification information while also having more accurate classification capabilities. Specifically include the following:

Use the training sample data set containing multiple categories of ice and snow crystal particle images to train the built lightweight network UDiNet, and use the test data set of ice and snow crystal particle images to classify the lightweight network UDiNet. Test and verify, select the model parameters with the best performance to save;

Input the acquired ice and snow crystal particle images into the trained and tested lightweight network UDiNet, perform feature extraction and category inference calculation on the ice and snow crystal particle images, and obtain a predicted category to achieve automatic classification of the ice and snow crystal particle images. .

The input of the lightweight network UDiNet is a 224×224×3 RGB color ice and snow crystal particle image. Its backbone network is a channel reorganization network. The lightweight network UDiNet includes a channel reorganization convolution unit, a lightweight attention mechanism layer and depth Separable atrous convolutional layers;

Among them, the depth-separable hole convolution layer is the last layer of the convolutional network of the network. ReLU6 is used for activation after each lightweight attention mechanism layer of the lightweight network UDiNet, and the exponential smoothing is a batch of momentum=0.9 Normalization BatchNormalization accelerates training and improves performance.

As shown in Figures 2 to 5, first, ordinary convolution with a convolution kernel size of 3×3 is used to perform preliminary feature extraction at a shallow level, and then the channel reorganization unit and the attention mechanism module are combined to perform an attention mechanism. Channel reorganization feature learning, then use depth-separable atrous convolution in the last layer of the feature extractor part of the network to fully extract and integrate all features between and within channels, and finally based on the output of the network feature extractor , using a fully connected layer to predict the category of the model.

The channel reorganization convolution unit is used to construct a channel reorganization network. This unit rearranges the information of each channel group in the feature map during group convolution, forming a channel rearrangement mechanism to increase information fusion between channels. In order to further increase the recognition accuracy, as shown in Figure 3 and Figure 4, a lightweight attention mechanism layer is introduced after steps 2 and 3 of the channel reorganization network to implement feature map importance weighting, and the attention weight obtained through learning , the model can automatically select and focus on features that are more important to the current task. This helps to improve the model's perception of key features and reduce the interference of redundant information, thereby improving the model's performance and generalization ability. Finally, the output feature map of the last layer of channel reorganized convolution unit is used as the input of the depth-separable atrous convolution layer.

The channel reorganization network is constructed using basic units composed of grouped convolution and pointwise convolution, and a grouped channel rearrangement mechanism is introduced in the grouped convolution to make the output feature map of the last layer of basic unit of the channel reorganization network as depth separable Input to atrous convolutional layer.

The depth-separable atrous convolution layer uses a residual network structure, which includes a depth-separable atrous convolution branch and short connection analysis. The two branches are respectively based on the feature map and channel number generated by the last basic unit of the channel reorganization network. Standard, the depth-separable atrous convolution branch is used to extract global features of the input feature map, and the short connection branch is used to collect local information of the input feature map, and the output feature maps of the two branches are connected in the channel dimension to achieve channel Extension.

As shown in Figure 6, the global feature extraction of the input feature map is performed through the depth-separable atrous convolution branch, and the local information collection of the input feature map through the short connection branch specifically includes:

A1. Use a deep convolution with a convolution kernel size of 3×3 to learn the feature information of the ice and snow crystal particle image details on a single feature map, and then use a point-by-point convolution with a convolution kernel size of 1×1 to learn the feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images;

A2. Upgrade the dimension through point-by-point convolution with a convolution kernel size of 1×1 to obtain more feature information of different channels of the ice and snow crystal particle image, and then fill two layers of pixels with a value of 0 around the feature map to While expanding the feature map of ice and snow crystal particles, no noise will interfere with the feature information;

A3. Use deep convolution with a convolution kernel size of 3×3 to learn the overall feature information of the ice and snow crystal particle feature image on a single feature map, and use point-by-point convolution with a convolution kernel size of 1×1 to learn the overall feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images;

A4. Connect the calculation results of steps A1 and A2 along the channel dimension as the output feature map of the depth-separable atrous convolution layer.

The loss function of the lightweight network UDiNet selects the cross-entropy loss function, which is defined as the following formula:

Where N is the number of samples, _xi is the input ice and snow crystal particle image, the probability distribution p(xi ₎ is the expected output, that is, the true category of the ice and snow crystal particle image _xi , and the probability distribution q(xi ₎ is the actual output, That is, the network model predicts the category of the ice and snow crystal particle image x _i . The smaller the loss, the closer the distribution of the two is, which means the classification result is better.

The training of the lightweight network UDiNet includes the following:

The parameters of the convolutional neural network pre-trained on the large image data set ImageNet are used for the initialization of the convolutional layer in the backbone network of the lightweight network UDiNet;

Using the Pytorch framework, Epochs is set to 100, and Batch Size is set to 128;

Use the AdamW optimization algorithm and set the learning rate to 0.0001;

In the public data set ICDC containing multiple categories of ice and snow crystal particle images, 80% of the image data for each category were randomly selected for training. At the same time, due to the small number of samples in the existing data set, in order to improve the generalization ability of the model, preprocessing operations were performed on the data, such as random aspect ratio cropping, random horizontal flipping and other operations for data enhancement, achieving a certain degree of Expanding the data set; and standardizing the image, and centralizing the image data through demeaning. According to the convex optimization theory and knowledge about data probability distribution, data centralization conforms to the law of data distribution, making it easier to obtain the results after training. Generalization effect; In addition, in order to adapt to the input size of the UDiNet model, random cropping is used to unify the image size to 224×224. The specific preprocessing operations are as follows: realize random horizontal flipping and vertical flipping of the image with a probability of 0.8; perform 0 padding according to the size of the image to make the image size a square, and then use the traditional interpolation algorithm bilinear interpolation method to unify it to 256×256 size; 224×224 random cropping is performed to fit the model UDiNet input; pixel-by-pixel normalization is performed, that is, the pixel value of each channel in the image is subtracted from the mean of the pixel value of the corresponding channel and then divided by the standard deviation to achieve Data centralization. Among them, when performing pixel-by-pixel normalization, the average values of the three RGB channels in the training set are 0.056, 0.331, and 0.666, and the standard deviations are 0.080, 0.218, and 0.303 respectively.

The test data set of ice and snow crystal particle images is used to classify and test the ice and snow crystal particle images of the lightweight network UDiNet, and the model parameters with the best performance are selected for storage, which specifically includes the following:

At the end of each Epoch training, use the same test sample data set to test the lightweight network UDiNet, and evaluate the classification results obtained from the test;

by formula Calculate the classification accuracy of the lightweight network UDiNet for all test sample data sets, where ŷ _i represents the category of the i-th image predicted by the lightweight network UDiNet, y _i is the corresponding true category value, and M is the test time The total number of images input to the network model, eq(y _i, ŷ _i ) is an equality function. If and only if y _i approaches ŷ _i , the better the prediction performance of the lightweight network UDiNet is, the closer the accuracy is to 1 ;

The data used for testing is the remaining 20% of the image data except the training sample data set in the public data set ICDC containing multiple categories of ice and snow crystal particle images for testing; and the following preprocessing operations are performed on it before inputting it into the network : Use bilinear interpolation method to adjust the image to 256×256 in proportion; randomly crop the image to 224×224 size; perform pixel-by-pixel normalization, that is, the pixel value of each channel in the image minus the pixel value of the corresponding channel The mean is divided by the standard deviation to centralize the data. Among them, when performing pixel-by-pixel normalization processing, the mean values of RGB channels in the test set are 0.056, 0.331, and 0.666, and the standard deviations are 0.080, 0.218, and 0.303 respectively;

Save the lightweight network UDiNet with the highest calculated accuracy in multiple tests.

Multiple classifications of the ice and snow crystal particle images include rosettes, rosettes, long columns, short columns, hollow columns, spheres, small irregularities, plates, hexagonal snowflakes and complex shapes; depth-separable hollow volumes The multilayer output feature map is input into the convolution classifier, and then the softmax function is used to normalize the output probability to complete the task of classifying 10 categories of ice and snow crystal particle images.

The above are only preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and Modifications can be made within the scope of the ideas described herein through the above teachings or technology or knowledge in related fields. Any modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention shall be within the protection scope of the appended claims of the present invention.

Claims (7)

1. A lightweight network-based ice and snow crystal particle image classification method, characterized in that: the classification method includes: Use the training sample data set containing multiple categories of ice and snow crystal particle images to train the built lightweight network UDiNet, and use the test data set of ice and snow crystal particle images to classify the lightweight network UDiNet. Test and verify, select the model parameters with the best performance to save; Input the acquired ice and snow crystal particle images into the trained and tested lightweight network UDiNet, perform feature extraction and category inference calculation on the ice and snow crystal particle images, and obtain a predicted category to achieve automatic classification of the ice and snow crystal particle images. . 2. A kind of ice and snow crystal particle image classification method based on lightweight network according to claim 1, characterized in that: the backbone network of the lightweight network UDiNet is a channel reorganization network, and the lightweight network UDiNet includes channels Reorganized convolutional units, lightweight attention mechanism layers and depthwise separable atrous convolutional layers; The channel reorganization convolution unit is used to construct a channel reorganization network, and rearranges the information of each channel group in the feature map during group convolution to form a channel rearrangement mechanism and increase information fusion between channels; the lightweight The level-level attention mechanism layer is used to weight the importance of feature maps. The learned attention weights allow the lightweight network UDiNet to automatically select and focus on features that are more important to the current task; The channel reorganization network is constructed using basic units composed of grouped convolution and pointwise convolution, and a grouped channel rearrangement mechanism is introduced in the grouped convolution to make the output feature map of the last layer of basic unit of the channel reorganization network as depth separable Input to atrous convolutional layer. 3. A lightweight network-based ice and snow crystal particle image classification method according to claim 2, characterized in that: the depth-separable hole convolution layer uses a residual network structure, which includes a depth-separable hole convolution layer. Integral branch and short connection analysis. The two branches respectively use the feature map and the number of channels generated by the last basic unit of the channel recombination network as standards, and perform global feature extraction on the input feature map through the depth-separable hole convolution branch. Through the short connection The branch collects local information from the input feature map, and connects the output feature maps of the two branches in the channel dimension to achieve channel expansion. 4. A lightweight network-based ice and snow crystal particle image classification method according to claim 3, characterized in that: the input feature map is extracted globally through the depth-separable hole convolution branch, and short connections are used to extract the global features of the input feature map. The branch collects local information from the input feature map and specifically includes: A1. Use a deep convolution with a convolution kernel size of 3×3 to learn the feature information of the ice and snow crystal particle image details on a single feature map, and then use a point-by-point convolution with a convolution kernel size of 1×1 to learn the feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images; A2. Upgrade the dimension through point-by-point convolution with a convolution kernel size of 1×1 to obtain more feature information of different channels of the ice and snow crystal particle image, and then fill two layers of pixels with a value of 0 around the feature map to While expanding the feature map of ice and snow crystal particles, no noise will interfere with the feature information; A3. Use deep convolution with a convolution kernel size of 3×3 to learn the overall feature information of the ice and snow crystal particle feature image on a single feature map, and use point-by-point convolution with a convolution kernel size of 1×1 to learn the overall feature information of the ice and snow crystal particles. Integrate information from multiple feature maps of particle images; A4. Connect the calculation results of steps A1 and A2 along the channel dimension as the output feature map of the depth-separable atrous convolution layer. 5. A kind of ice and snow crystal particle image classification method based on lightweight network according to claim 1, characterized in that: the training of the lightweight network UDiNet includes the following content: The parameters of the convolutional neural network pre-trained on the large image data set ImageNet are used for the initialization of the convolutional layer in the backbone network of the lightweight network UDiNet; Using the Pytorch framework, Epochs is set to 100, and Batch Size is set to 128; Use the AdamW optimization algorithm and set the learning rate to 0.0001; In the public data set ICDC containing multiple categories of ice and snow crystal particle images, 80% of the image data for each category were randomly selected for training. 6. A kind of ice and snow crystal particle image classification method based on lightweight network according to claim 1, characterized in that: the test data set of ice and snow crystal particle images is used to perform ice and snow crystal particle image classification on the lightweight network UDiNet. Carry out classification testing and verification, and select the model parameters with the best performance to save, including the following: At the end of each Epoch training, use the same test sample data set to test the lightweight network UDiNet, and evaluate the classification results obtained from the test; by formula Calculate the classification accuracy of the lightweight network UDiNet for all test sample data sets, where ŷ _i represents the category of the i-th image predicted by the lightweight network UDiNet, y _i is the corresponding true category value, and M is the test time The total number of images input to the network model, eq(y _i, ŷ _i ) is an equality function. If and only if y _i approaches ŷ _i , the better the prediction performance of the lightweight network UDiNet is, the closer the accuracy is to 1 ; The data used for testing is the remaining 20% of the image data except the training sample data set in the public ICDC data set containing multiple categories of ice and snow crystal particle images; Save the lightweight network UDiNet with the highest calculated accuracy in multiple tests. 7. A lightweight network-based ice and snow crystal particle image classification method according to claim 1, characterized in that: the multiple classifications of the ice and snow crystal particle images include rosette-shaped, rose-shaped, long columnar, Short columnar, hollow columnar, spherical, small irregular, plate, hexagonal snowflake and complex shapes; the output feature map of the depth-separable atrous convolution layer is input to the convolution classifier, and then the softmax function is used to normalize the output probability ization, completing the task of classifying 10 categories of ice and snow crystal particle images.