A deep learning-based method for image segmentation of adhered and stacked particles

Abstract

The present invention relates to a method for segmenting adhesion and stacked particle images based on deep learning, and belongs to the field of computer image segmentation. The method comprises: S1: annotating particle image data; S2: constructing a deep learning neural network model, using the particle image and the annotated image obtained in step S1 as a training set for network training, that is, adjusting the network parameters through a back propagation algorithm; S3: performing instance segmentation on the particle image using the deep learning neural network model constructed in step S2; S4: performing statistical analysis on the instance segmentation results of the particle image. The present invention can perform instance segmentation on adhesion and stacked particle images, obtain labeled images of different particles, and thus solve the problem of counting adhesion and stacked particles.

Description

Technical Field

[0001] The present invention belongs to the field of computer image segmentation and relates to a method for segmenting adhesion and stacked particle images based on deep learning. Background Art

[0002] Common image recognition methods include target detection, semantic segmentation, and instance segmentation. However, the implementation difficulty of the three methods varies: (1) Target detection is to encircle objects of different categories in the image with different rectangular frames. It does not need to accurately determine the boundaries of the objects, so it is not applicable in scenes that require refined object detection. (2) Semantic segmentation is to represent objects of different categories in the image with different labels, and assign a category label (such as car, building, person, etc.) to each pixel. It can achieve pixel-level boundary segmentation, but cannot distinguish different individuals. Therefore, this method cannot distinguish the count of adhesion and stacked particles. (3) Instance segmentation is to represent objects of different categories in the image with different labels, and assign a category label (such as car, building, person, etc.) to each pixel, and distinguish and judge different individuals of the same category.

[0003] Tasks such as counting biological cells in a microscopic image, counting nanomaterial particles, the number of people in a crowd under camera surveillance, and crop particles in crop monitoring are extremely difficult to count manually due to the huge number of particles in the image, so it is extremely important to develop a method to automatically detect particles.

[0004] At present, the instance segmentation method is often used in image particle recognition, mainly including: grayscale-based watershed algorithm, edge detection-based Hough transform algorithm, level set algorithm, feature space classification and other algorithms. However, these traditional algorithms can only be used in simple situations and are not suitable for the following situations:

[0005] (1) Various particle types and their different shapes and areas.

[0006] (2) Too much edge information of particles is lost and the edges are not clear.

[0007] (3) Mutual adhesion and overlap between particles.

[0008] The particle segmentation method based on deep learning has stronger recognition ability and better robustness. For example, the patent application with publication number CN111899274A discloses a particle size analysis method based on deep learning TEM image segmentation. This method uses an artificially labeled semantic segmentation neural network model to segment TEM nanoparticles, which can segment particles of different shapes. However, it can only achieve semantic segmentation and still cannot solve the problem of counting adhered and stacked particles. Therefore, particle image segmentation requires an instance segmentation method that can solve adhered and stacked particles. Summary of the invention

[0009] In view of this, the purpose of the present invention is to provide a method for segmenting adhesion and stacked particle images based on deep learning, which uses a deep learning network architecture to perform instance segmentation on adhesion and stacked particle images to obtain labeled images of different particles, thereby solving the problem of counting adhesion and stacked particles.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A method for segmenting an image of adhesion and stacked particles based on deep learning, specifically comprising the following steps:

[0012] S1: Prepare the dataset: manually annotate the particle image data;

[0013] S2: Build a deep learning neural network model, use the particle image and the labeled image obtained in step S1 as the training set to perform network training, that is, adjust the network parameters through the back propagation algorithm to achieve the best results for both the training set and the validation set;

[0014] S3: Perform instance segmentation on the particle image using the deep learning neural network model constructed in step S2;

[0015] S4: Perform statistical analysis on the instance segmentation results of the particle image, which can be displayed in different colors.

[0016] Further, in step S2, the constructed deep learning neural network model includes: a backbone feature extraction network based on the Convolutional Block Attention Module (CBAM) module, an anchor box generation network (Region Proposal Network, RPN), a ROI classification regression network and a Mask instance segmentation network;.

[0017] Further, in step S2, a backbone feature extraction network based on a CBAM module is used to extract backbone features, which specifically includes the following steps:

[0018] (1) The input image is processed by the CBAM module once to obtain the feature layer of the image;

[0019] (2) Then perform a 7×7 convolutional layer with a step size of 2;

[0020] (3) Then perform a 3×3 maximum pooling layer with a step size of 2;

[0021] (4) Then perform a CBAM module process;

[0022] (5) Then perform the ResNetBlock module processing three times;

[0023] (6) Then perform another CBAM module processing and then a 1×1 convolution to obtain the feature layer C2;

[0024] (7) Then perform four ResNetBlock module processing and one CBAM module processing to obtain feature layer C3;

[0025] (8) Then, the processing result of the CBAM module in (7) is processed by the ResNetBlock module six times, and then processed by the CBAM module once more to obtain the feature layer C4;

[0026] (9) Then, the processing result of the CBAM module in (8) is processed by the ResNetBlock module three times, and then processed by the CBAM module once to obtain the feature layer C5;

[0027] (10) The feature layer C5 undergoes a 1×1 convolution to obtain the feature layer P5;

[0028] (11) Downsample the feature layer P5 in (10) once, and then obtain the feature layer P6 through downsampling;

[0029] (12) Upsample the feature layer P5 in (10) to obtain the feature layer X5, and perform weighted addition of the feature layer obtained by 1×1 convolution of the feature layer C4 and X5 to obtain the feature layer P4;

[0030] (13) Upsampling the feature layer P4 in (12) once to obtain the feature layer X4, and performing weighted addition of the feature layer obtained by 1×1 convolution of the feature layer C3 and X4 to obtain the feature layer P3;

[0031] (14) The feature layer P3 in (13) is upsampled once to obtain a feature layer X3 of size (256, 256, 256), and the feature layer C2 obtained by 1×1 convolution with a channel number of 256 is added to the corresponding elements of X3 to obtain a feature layer P2 of size (256, 256, 256).

[0032] Further, in step S2, the structure of the CBAM module is as follows: the input image or feature is subjected to global MaxPool pooling and global AvgPool pooling respectively, and then the obtained global MaxPool pooling features and global AvgPool pooling features are sent to the same fully connected layer to obtain MaxPool weights and AvgPool weights respectively, the two sets of weights are multiplied to obtain channel attention weights, and then the channel attention weights are multiplied with the input image or feature to obtain the channel attention processing result;

[0033] Then, the results of channel attention processing are used to compress the values ​​of multiple channels into a single channel through MaxPool pooling and AvgPool pooling to obtain spatial features. The spatial features are then fused through 7×7 convolution to obtain spatial attention weights. The spatial attention weights are multiplied by the input image or features to obtain the processing results of the CBAM module.

[0034] Furthermore, in step S2, the structure of the ResNetBlock module is: the input feature layer is subjected to 3×3 convolution, and then BatchNormal and Relu activation functions are performed, and then 3×3 convolution is performed again, and BatchNormal and Relu are performed again to obtain the processed feature layer, and the feature layer is short-circuited with the input feature layer to perform a weighted addition operation. The purpose of adaptively adjusting the network to make the convolution operation in the module effective is achieved, thereby deepening the network without causing problems such as gradient explosion and gradient disappearance.

[0035] Further, in step S2, RPN is used to obtain a suggestion box, which specifically includes the following steps:

[0036] (1) Feature layers P2 to P6 are input, and according to the sizes of the five feature layers, three candidate boxes are obtained for each pixel point of each feature layer on the original image. The number of candidate boxes for each layer is related to the size of the feature layer.

[0037] (2) Perform a 1×1 convolution on the feature layers P2 to P6 and then perform softmax activation function classification to obtain the confidence that each candidate box is the foreground;

[0038] (3) After performing a 1×1 convolution on the feature layers P2 to P6, an adjustment parameter for each candidate box is obtained. The adjustment parameter adjusts the candidate box so that the object is in the center of the candidate box.

[0039] (4) Select the candidate boxes with confidence in the top 6000, and adjust the candidate boxes using adjustment parameters to obtain the final recommended boxes.

[0040] Further, in step S3, ROI classification regression network prediction is adopted, which specifically includes the following steps:

[0041] (1) Input feature layers P2 to P5 and suggestion boxes, and obtain the local feature layer of each layer through the input feature layers and suggestion boxes;

[0042] (2) Then perform BatchNormal, and then perform 1×1 convolution and ReLU activation function processing;

[0043] (3) Then, the result of (2) is input into the fully connected layer and classified using the softmax activation function to obtain the confidence of each proposal box;

[0044] (4) Then, the result of (2) is input into the fully connected layer and a linear activation function is used to obtain the adjustment parameters of each suggestion box;

[0045] (5) Select the top 200 suggestion boxes with the highest confidence, and use the adjustment parameters in (4) to adjust the suggestion boxes to obtain the predicted boxes.

[0046] Further, in step S2, a Mask instance segmentation network is used to perform instance segmentation, which specifically includes the following steps:

[0047] (1) Input feature layers P2 to P5 and the prediction box;

[0048] (2) Obtain the local feature layer through the feature layer and the prediction box;

[0049] (3) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing;

[0050] (4) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing;

[0051] (5) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing;

[0052] (6) Then perform a 3×3 convolution and then perform a ReLU function processing;

[0053] (7) Then, a 2×2 deconvolution is performed, and the sigmoid activation function is used to classify each pixel to obtain the category of each pixel, and then the instance segmentation labeling map is drawn.

[0054] The beneficial effects of the present invention are:

[0055] (1) The present invention combines the CBAM module into the instance segmentation model for the first time, thereby improving the deep learning network and enhancing the accuracy of particle segmentation.

[0056] (2) The present invention can effectively separate the situations where particles are stuck together or stacked, thereby solving the cardinality problem.

[0057] (3) The segmentation method of the present invention is applicable to instance segmentation in the fields of biological cells, nanoparticles, crop particles, etc.

[0058] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0060] Figure 1 This is the original image of stacked particles;

[0061] Figure 2 This is the original image of the adhesion particles;

[0062] Figure 3 Label the graph for stacked particles;

[0063] Figure 4 Label the diagram for the adhesion particles;

[0064] Figure 5 A diagram of the deep learning neural network model architecture constructed for the present invention;

[0065] Figure 6 This is the network structure diagram for backbone feature extraction based on the CBAM module;

[0066] Figure 7 This is the CBAM module structure diagram;

[0067] Figure 8 This is the ResNetBlock module structure diagram;

[0068] Fig. 9 It is the RPN structure diagram;

[0069] Fig.10 This is the ROI classification regression network structure diagram;

[0070] Fig.11 This is the Mask instance segmentation network structure diagram;

[0071] Fig.12 Segmentation labeling map for stacked particle instances;

[0072] Fig.13 Segmentation labeling for the adhesion particle instance. DETAILED DESCRIPTION

[0073] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0074] See also Figures 1 to 13 ,This paper proposes an improved Fast R-CNN algorithm, which mainly utilizes a deep learning network ,architecture segmentation of particle images, ,obtains labeled images of different particles, ,realizes instance segmentation of particle images, and solves the problem of sticky and stacked particles.

[0075] Embodiment 1: A method for segmenting an instance of an image of adhesion and stacked particles based on deep learning, comprising the following steps:

[0076] S1: Prepare dataset;

[0077] For particle image data (such as Figure 1 , 2 As shown) for manual annotation (such as Figure 3 , 4 As shown), that is, different labels are marked on different individuals (which can be represented by different colors).

[0078] S2: Build a deep learning neural network model, use the particle image and the labeled image obtained in step S1 as the training set for network training, that is, adjust the network parameters through the back propagation algorithm to achieve the best results for both the training set and the validation set.

[0079] like Figure 5 As shown in the figure, the architecture of the constructed deep neural network, in addition to the input layer and the output layer, mainly includes: backbone feature extraction network, anchor box generation network (RPN), ROI classification regression network, Mask instance segmentation network, and its characteristic is the integration of attention module (CBAM).

[0080] like Figure 6 As shown in the figure, the process of training deep neural networks mainly includes: backbone feature extraction part, RPN suggestion box part, ROI Align prediction part, Mask instance segmentation part. Each part is introduced in detail below.

[0081] 1) Main feature extraction part, such as Figure 6 As shown, the specific steps include:

[0082] (1) The input image is processed by the CBAM module once to obtain the feature layer of the image.

[0083] Among them, the structure processed by the CBAM module is as follows Figure 7 As shown in the figure: the input image or feature undergoes global MaxPool pooling and global AvgPool pooling respectively, and then the obtained global MaxPool pooling features and global AvgPool pooling features are sent to the same fully connected layer to obtain MaxPool weights and AvgPool weights respectively, and the two sets of weights are multiplied to obtain the channel attention weights, and then the channel attention weights are multiplied with the input image or feature to obtain the channel attention processing result.

[0084] The results of channel attention processing are compressed into a single channel through MaxPool pooling and AvgPool pooling to obtain spatial features, and then the spatial features are fused through 7x7 convolution to obtain spatial attention weights. The spatial attention weights are multiplied by the input image or features to obtain the CBAM module to obtain the processing result.

[0085] (2) Then a 7×7 convolutional layer is performed with a step size of 2.

[0086] (3) Then perform a 3×3 maximum pooling layer with a step size of 2.

[0087] (4) Then, the CBAM module is processed.

[0088] (5) Then perform the ResNetBlock module processing three times.

[0089] Among them, the structure of the ResNetBlock module is as follows Figure 8 As shown in the figure: the input feature layer is convolved by 3x3, then BatchNormal and Relu activation functions are performed, then 3x3 convolution is performed again, and BatchNormal and Relu are performed again to obtain the processed feature layer, and this feature layer is short-circuited with the input feature layer to perform weighted addition operations. The purpose of adaptively adjusting the network to make the convolution operation in this module effective is achieved, thereby deepening the network without causing problems such as gradient explosion and gradient disappearance.

[0090] (6) Then, the CBAM module is processed again, and then a 1×1 convolution is performed to obtain the feature layer C2.

[0091] (7) Then, the ResNetBlock module is processed four times and the CBAM module is processed once to obtain the feature layer C3.

[0092] (8) Then, the processing result of the CBAM module in (7) is processed by the ResNetBlock module six times and then processed by the CBAM module once to obtain the feature layer C4.

[0093] (9) Then, the processing result of the CBAM module in (8) is processed by the ResNetBlock module three times and then processed by the CBAM module once to obtain the feature layer C5.

[0094] (10) The feature layer C5 is subjected to a 1×1 convolution to obtain the feature layer P5.

[0095] (11) Downsample the feature layer P5 in (10) once, and then obtain the feature layer P6 through downsampling.

[0096] (12) Upsample the feature layer P5 in (10) to obtain the feature layer X5, and perform weighted addition of the feature layer obtained by 1×1 convolution on the feature layer C4 and X5 to obtain the feature layer P4.

[0097] (13) Upsample the feature layer P4 in (12) to obtain the feature layer X4, and perform weighted addition of the feature layer obtained by 1×1 convolution on the feature layer C3 and X4 to obtain the feature layer P3.

[0098] (14) The feature layer P3 in (13) is upsampled once to obtain a feature layer X3 of size (256, 256, 256), and the feature layer C2 obtained by 1×1 convolution with a channel number of 256 is added to the corresponding elements of X3 to obtain a feature layer P2 of size (256, 256, 256).

[0099] 2)RPN suggestion box part, such as Fig. 9 As shown, the specific steps include:

[0100] (1) The obtained feature layers P2 to P6 are input, and according to the sizes of the five feature layers, three candidate boxes are obtained on the original image for each pixel point of each feature layer. The number of candidate boxes of each layer is related to the size of the layer.

[0101] (2) Perform a 1×1 convolution on the feature layers P2 to P6 and then perform softmax activation function classification to obtain the confidence that each candidate box is the foreground.

[0102] (3) After performing a 1×1 convolution on the P feature layers 2 to P6, an adjustment parameter for each candidate box is obtained. The adjustment parameter adjusts the candidate box so that the object is in the center of the candidate box.

[0103] (4) Select the candidate boxes with confidence in the top 6000, and adjust the candidate boxes using adjustment parameters to obtain the final recommended boxes.

[0104] 3) ROI Align prediction part, such as Fig.10 As shown, the specific steps include:

[0105] (1) Input feature layers P2 to P5 and suggestion boxes, and obtain the local feature layer of each layer through the input feature layers and suggestion boxes.

[0106] (2) Then perform BatchNormal, and then perform 1×1 convolution and Relu activation function processing.

[0107] (3) Then, the result of (2) is input into the fully connected layer and classified using the softmax activation function to obtain the confidence of each proposal box.

[0108] (4) Then, the result of (2) is input into the fully connected layer and a linear activation function is used to obtain the adjustment parameters of each suggestion box.

[0109] (5) Select the top 200 suggestion boxes with the highest confidence, and use the adjustment parameters in (4) to adjust the suggestion boxes to obtain the predicted boxes.

[0110] 4) Mask instance segmentation part, such as Fig.11 As shown, the specific steps include:

[0111] (1) Input feature layers P2~P5 and the prediction box.

[0112] (2) Obtain the local feature layer through the feature layer and the prediction box.

[0113] (3) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing.

[0114] (4) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing.

[0115] (5) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing.

[0116] (6) Then perform a 3×3 convolution and then perform a ReLU function processing.

[0117] (7) Then, a 2×2 deconvolution is performed, and the sigmoid activation function is used to classify each pixel to obtain the category of each pixel, and then the instance segmentation labeling map is drawn.

[0118] S3: Use the deep learning neural network model obtained in step S2 to perform instance segmentation on the particle image.

[0119] S4: Statistical analysis of instance segmentation results of particle images can be performed using different colors, such as Fig.12 , Fig.13 shown.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for segmenting adhesion and stacked particle images based on deep learning, characterized in that: The method specifically comprises the following steps: S1: Prepare the dataset: label the particle image data; S2: Construct a deep learning neural network model, use the particle image and the labeled image obtained in step S1 as the training set to perform network training, that is, adjust the network parameters through the back propagation algorithm; the constructed deep learning neural network model includes a backbone feature extraction network based on the CBAM module, RPN, ROI classification regression network and Mask instance segmentation network; among which, the CBAM module is an attention module, and the RPN is an anchor box generation network; The backbone feature extraction network based on the CBAM module is used to extract the backbone features, which specifically includes the following steps: (1) The input image is processed by the CBAM module once to obtain the feature layer of the image; (2) Then perform a 7×7 convolutional layer with a step size of 2; (3) Then perform a 3×3 maximum pooling layer with a step size of 2; (4) Then, the CBAM module is processed once; (5) Then perform three ResNetBlock module processes; (6) Then perform another CBAM module processing, and then undergo a 1×1 convolution to obtain the feature layer C2; (7) Then perform four ResNetBlock module processing and one CBAM module processing to obtain feature layer C3; (8) Then, the processing result of the CBAM module in (7) is processed by the ResNetBlock module six times, and then processed by the CBAM module once more to obtain the feature layer C4; (9) Then, the processing result of the CBAM module in (8) is processed by the ResNetBlock module three times, and then processed by the CBAM module once to obtain the feature layer C5; (10) The feature layer C5 is subjected to a 1×1 convolution to obtain the feature layer P5; (11) Downsample the feature layer P5 in (10) once, and then obtain the feature layer P6 through downsampling; (12) Upsample the feature layer P5 in (10) to obtain the feature layer X5, and perform weighted addition of the feature layer obtained by 1×1 convolution of the feature layer C4 and X5 to obtain the feature layer P4; (13) Upsample the feature layer P4 in (12) to obtain the feature layer X4, and perform weighted addition of the feature layer obtained by 1×1 convolution of the feature layer C3 and X4 to obtain the feature layer P3; (14) Upsample the feature layer P3 in (13) to obtain the feature layer X3, and add the corresponding elements of the feature layer obtained by 1×1 convolution of the feature layer C2 to X3 to obtain the feature layer P2; The structure of the CBAM module is as follows: the input image or feature undergoes global MaxPool pooling and global AvgPool pooling respectively, and then the obtained global MaxPool pooling features and global AvgPool pooling features are sent to the same fully connected layer to obtain MaxPool weights and AvgPool weights respectively, and the two sets of weights are multiplied to obtain channel attention weights, and then the channel attention weights are multiplied with the input image or feature to obtain the channel attention processing result; then, the results of the channel attention processing are compressed into a single channel through MaxPool pooling and AvgPool pooling to obtain spatial features, and then the spatial features are fused through 7×7 convolution to obtain spatial attention weights, and the spatial attention weights are multiplied with the input image or feature to obtain the processing result of the CBAM module; S3: Use the deep learning neural network model constructed in step S2 to perform instance segmentation on the particle image.

2. The method for segmenting an image of adhesion and stacked particles according to claim 1, characterized in that: In step S2, the structure of the ResNetBlock module is as follows: the input feature layer is subjected to a 3×3 convolution, and then a BatchNormal and Relu activation function is performed, and then a 3×3 convolution is performed, and then a BatchNormal and Relu are performed to obtain the processed feature layer, and the feature layer is short-circuited with the input feature layer to perform a weighted addition operation.

3. The method for segmenting an image of adhesion and stacked particles according to claim 1, characterized in that: In step S2, RPN is used to obtain a suggestion box, which specifically includes the following steps: (1) Input feature layers P2 to P6, and according to the sizes of the five feature layers, obtain three candidate boxes for each pixel point in each feature layer on the original image. The number of candidate boxes in each layer is related to the size of the feature layer. (2) Perform a 1×1 convolution on the feature layers P2~P6 and then perform softmax activation function classification to obtain the confidence that each candidate box is the foreground; (3) After performing a 1×1 convolution on the feature layers P2~P6, the adjustment parameters for each candidate box are obtained. The adjustment parameters adjust the candidate box so that the object is in the center of the candidate box; (4) Select the candidate boxes with confidence in the top 6000, and adjust the candidate boxes using adjustment parameters to obtain the final recommended boxes.

4. The method for segmenting an image of adhesion and stacked particles according to claim 3, characterized in that: In step S2, ROI classification regression network prediction is adopted, which specifically includes the following steps: (1) Input feature layers P2~P5 and suggestion boxes, and obtain the local feature layer of each layer through the input feature layers and suggestion boxes; (2) Then perform BatchNormal, and then perform 1×1 convolution and Relu activation function processing; (3) Then, the result of (2) is input into the fully connected layer and classified using the softmax activation function to obtain the confidence of each proposed box; (4) Then, the result of (2) is input into the fully connected layer and a linear activation function is used to obtain the adjustment parameters of each suggestion box; (5) Select the top 200 suggested boxes with the highest confidence and adjust the suggested boxes using the adjustment parameters in (4) to obtain the predicted boxes.

5. The method for segmenting an image of adhesion and stacked particles according to claim 4, characterized in that: In step S2, a Mask instance segmentation network is used to perform instance segmentation, which specifically includes the following steps: (1) Input feature layers P2~P5 and prediction boxes; (2) Obtain the local feature layer through the feature layer and prediction box; (3) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing; (4) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing; (5) Then perform a 3×3 convolution, and then perform BatchNormal and Relu function processing; (6) Then perform a 3×3 convolution and then perform a ReLU function processing; (7) Then, perform 2×2 deconvolution, and use the sigmoid activation function to classify each pixel to obtain the category of each pixel, and then draw the instance segmentation labeling map.

6. The method for segmenting an image of adhesion and stacked particles according to any one of claims 1 to 5, characterized in that: The method further comprises step S4: performing statistical analysis on the instance segmentation results of the particle image.

Images