Plant seedling image classification method based on MFC neural network architecture search

Abstract

This invention proposes a plant seedling image classification method based on MFC neural network architecture search. The method comprises: constructing a dataset of plant seedling images; training a pre-defined model based on the MFC neural network architecture search using the dataset to obtain a plant seedling classification model; and classifying plant seedling images based on the plant seedling classification model. This invention can automatically design an effective convolutional neural network model for plant seedling image classification, demonstrating better competitiveness and advancement compared to manually designed classification models in recent years.

Description

Technical Field

[0001] This invention belongs to the field of image classification technology, and in particular relates to a method for classifying plant seedling images based on MFC neural network architecture search. Background Technology

[0002] Weeds hinder crop growth and generally need to be removed during the seedling stage to prevent damage. Manual weeding is typically used, but it is time-consuming and labor-intensive. Precision agriculture, based on information technology, aims to optimize agricultural inputs for maximum efficiency. The automatic collection of weed information primarily relies on computer image processing and pattern recognition technologies. Accurate identification of plant seedlings using computer vision technology is crucial, and using this technology for automated weed control also helps overcome the inconveniences of manual methods.

[0003] In the field of computer vision, Convolutional Neural Networks (CNNs) are end-to-end learning models capable of image extraction and classification. In recent years, plant seedling classification methods based on deep CNNs have been proposed. In research and applications of plant identification and weed detection, different plants often have leaves with similar shapes and colors, requiring differentiation based on leaf features. CNNs can extract features such as leaf shape and texture at different scales, enabling seedling identification through a classification network. However, current seedling identification techniques using classification networks suffer from low classification accuracy due to the lack of pre-training of CNN models on large datasets. While CNNs based on transfer learning can effectively improve classification accuracy, they require manual selection of pre-trained models, which is time-consuming and labor-intensive. Summary of the Invention

[0004] The purpose of this invention is to propose a plant seedling image classification method based on MFC neural network architecture search, which improves the performance of neural network architecture search for plant classification models and achieves accurate classification of plant seedlings.

[0005] To achieve the above objectives, this invention provides a plant seedling image classification method based on MFC neural network architecture search, comprising:

[0006] Construct a dataset of plant seedling images;

[0007] Construct a pre-defined search model based on the MFC neural network architecture;

[0008] Based on the dataset, a pre-defined model based on the MFC neural network architecture is trained to obtain a plant seedling classification model.

[0009] Based on the plant seedling classification model, plant seedling images are classified.

[0010] Optionally, constructing a pre-defined search model based on the MFC neural network architecture includes:

[0011] Construct an initial search space structure; wherein the initial search space structure includes: Transfer Cell network, Normal Cell network, Pooling Cell network and Dropout Cell network;

[0012] The controller is used to search the initial search space structure to obtain the optimal search space;

[0013] The optimal search spaces are stacked to obtain a model based on the MFC neural network architecture.

[0014] Optionally, the Transfer Cell network includes several pre-trained candidate sub-models, each of which includes several searchable layers and several different shared layers; wherein the searchable layers are set at the end of the different shared layers, and each shared layer includes several searchable layers.

[0015] The Transfer Cell network incorporates different shallow layers of different pre-trained candidate sub-models as weight transfer modules into the search space to automatically select shallow layers of different levels of different pre-trained candidate sub-models.

[0016] Optionally, the pre-trained candidate sub-models include: MobileNet V3-Large sub-network, EfficientNet sub-network, ResNet-50 sub-network, Xception sub-network, and Inception-v3 sub-network.

[0017] Optionally, the Normal Cell network includes a multi-branch structure composed of several preset first blocks, wherein each preset first block includes two OP candidate operation modules and one Combine connection candidate operation module;

[0018] The OP candidate operation module uses the DSC sub-module based on the attention mechanism as candidate operations;

[0019] The DSC submodule based on the attention mechanism uses DSC 3×3 and DSC 5×5 as candidates for the search space, and SE, CBAM and CA as candidates for the attention module.

[0020] Optionally, the Pooling Cell network includes: a preset second block, wherein the preset second block includes two Max Pooling candidate operation modules and a Combine connection candidate operation module.

[0021] Optionally, the Dropout Cell network includes: a plurality of Dropout modules with independent dropout rates P; the Dropout Cell network uses a controller to adaptively select preset Dropout modules and selects the best prediction result through a voting mechanism.

[0022] Optionally, using the controller to search the initial search space structure to obtain the optimal search space includes:

[0023] The sampling probability of each pre-trained candidate sub-model is predicted using the controller and sampling is performed to obtain the operation code of each pre-trained candidate sub-model; wherein, the Transfer cell network is used to search the shallow layers of different levels of different pre-trained models, the Normal cell network is used to search for candidate operation for feature extraction, the Pooling cell network is used to search for candidate operation for downsampling, and the Dropout cell network is used to search for the dropout rate P.

[0024] Construct an MFC-NAS candidate model based on the opcode;

[0025] The system is trained using a randomly selected subset of training samples, and the controller parameters are updated using gradient descent based on the validation accuracy. The next search begins and continues until the search ends, at which point the optimal multifunctional cell, namely the optimal Transfer cell, Normal cell, Pooling cell, and Dropout cell, is obtained.

[0026] The present invention has the following beneficial effects:

[0027] This invention proposes a plant seedling image classification method based on MFC neural network architecture search. It designs Transfercell, Normal cell, Pooling cell, and Dropout cell to improve the performance of the plant classification model. The method can automatically search for the optimal model structure based on the plant seedling dataset, achieving high classification accuracy and a small number of model parameters. The constructed plant seedling classification model can automatically design an effective convolutional neural network model for classifying plant seedling numbers, demonstrating better competitiveness and advancement compared to manually designed classification models in recent years. Attached Figure Description

[0028] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0029] Figure 1 This is a schematic diagram of the MFC-NAS process according to an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of the searchable weight sharing layer structure according to an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the Normal cell structure according to an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the depthwise separable convolutional module structure based on the attention mechanism in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the Pooling cell structure according to an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the Dropout cell structure according to an embodiment of the present invention;

[0035] Figure 7 This is a schematic diagram of the MFC-NAS model construction process according to an embodiment of the present invention;

[0036] Figure 8 This is a schematic diagram of plant seedlings according to an embodiment of the present invention.

[0037] Figure 9 This is a schematic diagram comparing the model training convergence speed of different shared layers S with an average of 500 controller samplings in an embodiment of the present invention.

[0038] Figure 10 This is a schematic diagram of the optimal transfer cell according to an embodiment of the present invention;

[0039] Figure 11 This is a schematic diagram of the optimal Normal cell in an embodiment of the present invention;

[0040] Figure 12 This is a schematic diagram of the optimal pooling cell according to an embodiment of the present invention;

[0041] Figure 13 This is a schematic diagram of the optimal Dropout cell according to an embodiment of the present invention;

[0042] Figure 14 This is a schematic diagram of the optimal overall structure of an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] This embodiment proposes a plant seedling image classification method based on MFC neural network architecture search, including:

[0046] Construct a dataset of plant seedling images;

[0047] Construct a pre-defined search model based on the MFC neural network architecture;

[0048] Based on the dataset, a pre-defined model based on the MFC neural network architecture is trained to obtain a plant seedling classification model.

[0049] Based on the plant seedling classification model, plant seedling images are classified.

[0050] Furthermore, the pre-defined search model based on the MFC neural network architecture includes:

[0051] Construct the initial search space structure, which includes: Transfer Cell network, Normal Cell network, Pooling Cell network, and Dropout Cell network.

[0052] The controller is used to search the initial search space structure to obtain the optimal search space;

[0053] By stacking the best search spaces, a model based on the MFC neural network architecture can be obtained.

[0054] Specifically, to improve the performance of Neural Architecture Search (NAS) for plant classification models, this embodiment proposes a multifunctional cell-driven classification model based on NAS, namely MFC-NAS. This includes:

[0055] (1) A multi-functional cell (MFC) search space is proposed, defining four different types of cells: Transfer Cell, Dropout Cell, Normal Cell, and Pooling Cell. Transfer Cell improves the expressive power of shallow features by adaptively sharing pre-trained weights, Dropout Cell improves the generalization ability of the model by adaptively selecting the dropout rate, Normal Cell is used to improve the feature extraction ability of the model, and Pooling Cell is used for downsampling and reducing the number of model parameters.

[0056] (2) Design a search strategy for MFC. Transfer Cells are constructed by searching for shared weights in different shallow layers of the pre-trained model. The search strategies for Pooling Cells, Dropout Cells, and Normal Cells are similar, searching for Pooling Cells of different pooling types and sizes, and Dropout Cells of different dropout rates. Finally, the best-performing cells are stacked to achieve the joint driving force of multiple functional cells. This method is named Multi-functional Cell based Neural Architecture Search (MFC-NAS).

[0057] Candidate MFC-NAS models are constructed based on opcodes. After the controller generates a list of opcodes for all cells, it constructs the corresponding cells according to the opcode information, such as... Figure 7 As shown, each opcode list corresponds to a cell. First, a T-cell is constructed, whose opcode list corresponds to the candidate options in the T-cell search space. Then, N-cell, P-cell, and D-cell are constructed sequentially. Similarly, the opcode lists of these cells also correspond to the candidate options in their respective search spaces. In the MFC-NAS model, a P-cell will pass through an N-cell again until it reaches the last P-cell before entering a D-cell. The number of times it passes through a D-cell is equal to the number of layers in the P-cell minus one, where the number of layers in the P-cell is equal to the number of layers in the N-cell (L).

[0058] (3) The MFC-NAS method was used in the plant seedling image classification experiment. Compared with the SOTA methods such as RepVGG and EfficientNet, MFC-NAS achieved higher test accuracy (mAP≈99.17%) and the number of model parameters was about 1.58M.

[0059] The MFC-NAS proposed in this embodiment mainly includes three stages: search space definition, searching for the best cell, stacking cells, and testing (e.g., ...). Figure 1 (As shown).

[0060] In the search space definition phase, four new search space structures were designed, including Transfer Cell, Normal Cell, Pooling Cell, and Dropout Cell.

[0061] In this embodiment, the Transfer Cell uses shallow blocks from multiple pre-trained models as searchable candidate weight sharing layers, which better expresses the universality of feature information and improves the model's generalization ability by automatically searching for suitable shallow blocks to adapt to different network structures. The Normal Cell is a multi-branch structure composed of multiple blocks, each including input / output nodes, candidate operations, and connection methods. The outputs of these blocks are connected using a concat operation. The Pooling Cell is a single-branch structure that uses multiple pooling operations as search candidates and consists of a single block, including input / output nodes and candidate operations. The Dropout Cell uses multiple different dropout probabilities P as search candidates, automatically searching for a suitable dropout rate P to improve the model's generalization ability.

[0062] Furthermore, the Transfer Cell network includes several pre-trained candidate sub-models, each of which includes several searchable layers and several different shared layers; wherein, the searchable layers are set at the end of different shared layers, and each shared layer includes several searchable layers; wherein, the pre-trained candidate sub-models include: MobileNet V3-Large sub-network, EfficientNet sub-network, ResNet-50 sub-network, Xception sub-network and Inception-v3 sub-network.

[0063] In the optimal cell search phase, search parameters are first initialized, including search strategy, search period, and controller parameters. Then, an RNN model is used as the controller to generate sampling probabilities for each cell candidate. Simultaneously, the controller is trained, and reinforcement learning policy gradients are used to further update the controller weights. A probabilistic sampling method is used to select suitable candidates from the search space. Through periodic training of the controller, optimal structures such as Transfer cell (T-cell), Normal cell (N-cell), Pooling cell (P-cell), and Dropout cell (D-cell) are obtained.

[0064] In the cell construction phase, the four search spaces of optimal T-cell, N-cell, P-cell and D-cell are stacked to obtain the MFC-NAS model. Finally, the model is tested, and the feasibility of this method is verified through comparison, ablation and other experiments.

[0065] Transfer cell

[0066] Inspired by the fact that shallow layers of CNNs can better represent general features, different shallow layers of different pre-trained models are added to the search space as weight transfer cells (T-cells) to automatically select shallow layers of different levels of different models, share the general feature information of pre-training, accelerate model convergence and improve generalization ability.

[0067] Five pre-trained candidate models were selected: MobileNet V3-Large, EfficientNet, ResNet-50, Xception, and Inception-v3. As shown in Table 1, each candidate model contains three searchable layers. For example, opcode 2,1 indicates that the candidate model is ResNet-50, and its searchable layer 2 is selected. To fully search different shallow layers of the candidate models, searchable layers are set at different shared layers S of the candidate models. i The end. For example. Figure 2 As shown, the shared layer consists of shallow layers of different levels, each shared layer S i Contains several searchable layers L j .

[0068] Table 1 Candidates for the Transfer Cell Search Space

[0069]

[0070] The shared layers of the candidate model MobileNet V3-Large are shown in Table 2. conv2d is a convolutional layer with a kernel size of 3×3 and a stride of 2, while bneck is a depthwise separable convolution (DSC) layer. This model contains three shared layers, using S... i,i∈{1,2,3} represents the first searchable layer of the first shared layer, i.e., the first bneck_3×3 of the searchable layer 3×bneck_3×3, and so on. In the EfficientNet shared layer, MBConv1 is composed of multiple bneck modules from MobileNetV3 (as shown in Table 3). In the ResNet-50 shared layer, conv1 is a convolutional layer with a kernel size of 7×7 and a stride of 2. conv2_x has a total of 3 sets of convolutional blocks, each containing a 1×1 convolution, a 3×3 convolution, and a 1×1 convolution (as shown in Table 4). The Xception shared layer (as shown in Table 5) contains 2 3×3 convolutions and 15 DSC_3×3. The Inception-v3 shared layer (as shown in Table 6) contains 6 3×3 convolutions and 3 Inception blocks, where the Inception block is a convolutional module for multi-scale feature fusion.

[0071] Table 2 Searchable Weight Sharing Layers for MobileNetV3-Large Candidate Models

[0072]

[0073] Table 3. Searchable Weight Sharing Layers for EfficientNet Candidate Models

[0074]

[0075] Table 4 shows the searchable weight-sharing layers for ResNe-50 candidate models.

[0076]

[0077] Table 5 shows the searchable weight-sharing layers for Xception candidate models.

[0078]

[0079] Table 6 shows the searchable weight-sharing layer for Xception candidate models.

[0080]

[0081]

[0082] Furthermore, the Normal Cell network includes a multi-branch structure composed of several preset first blocks, wherein each preset first block includes two OP candidate operation modules and one Combine connection candidate operation module;

[0083] The OP candidate operation module uses the DSC sub-module based on the attention mechanism as candidate operations;

[0084] The DSC submodule based on the attention mechanism considers DSC 3×3 and DSC 5×5 as candidates for the search space, and SE, CBAM and CA as candidates for the attention module.

[0085] In this embodiment, the Normal Cell network uses an improved Normal Cell;

[0086] The structure of the improved Normal cell is basically the same as that of NASNet. For example... Figure 3 As shown, the improved Normal cell is a multi-branch structure composed of N blocks. Each block contains two candidate operations (OPs) and a Combine connection candidate operation. The Combine operation includes two searchable connection candidates: Add and Concat. It contains five sequence blocks, each including two H operations, two OPs, and one Combine connection operation. Similar to the Normal cell, these are generated by a controller. H represents the input, and OPs are the operation candidates in the search space. The Combine operation includes two searchable connection candidates: Add and Concat. Unlike the NASNet search space, it primarily uses DSC and attention-based DSC as the main candidate operations. As shown in Table 7, DSC can reduce the number of parameters, while adding an attention mechanism can extract more important features. Table 8 lists the candidate operations for the Combine operation.

[0087] Table 7. Candidate Operations in the Normal Cell Search Space

[0088]

[0089]

[0090] Table 8. Candidate Operations for Combine

[0091]

[0092] The DSC module structure based on the attention mechanism is as follows: Figure 4As shown, this module uses DSC 3×3 and DSC 5×5 as candidates for the search space, and SE, CBAM, and CA as candidates for the attention module. First, the feature map is input into a Conv 1×1, then batch normalization (BatchNorm, BN) and ReLU activation (BN+ReLU) are performed, then input into a DSC 3×3 or DSC 5×5, followed by BN+ReLU, and then input into the candidate attention module to obtain the attention weights. The input is multiplied by the attention weights to obtain the weighted feature map, and finally input into a Conv 1×1+BN to output the module result.

[0093] Furthermore, the Pooling Cell network includes: a preset second block, wherein the preset second block includes two Max Pooling candidate operation modules and a Combine connection candidate operation module.

[0094] To adaptively reduce the number of parameters, this embodiment proposes a Pooling cell. The Pooling cell has a similar structure to the Normal cell, both consisting of blocks. However, the Pooling cell's search space only includes two operations: MaxPooling and Avg Pooling (as shown in Table 9), and it has only one block. The structure of the Pooling cell is as follows: Figure 5 As shown.

[0095] Table 9. Candidate Operations in the Pooling Cell Search Space

[0096]

[0097]

[0098] Furthermore, the Dropout Cell network includes: several Dropout modules with independent dropout rates P; the Dropout Cell network uses a controller to adaptively select preset Dropout modules and selects the best prediction result through a voting mechanism.

[0099] In this embodiment, a Dropout cell is designed to adaptively select the dropout rate to prevent model overfitting. For example... Figure 6 As shown in Table 10, the Dropout cell adds multiple Dropouts with independent dropout rates P to the search space, allowing the controller to adaptively select a Dropout and finally select the best prediction result through a voting mechanism, thereby improving the model's generalization ability.

[0100] Table 10 Candidate dropout rate P in the Dropout cell search space

[0101]

[0102] Furthermore, the controller is used to search the initial search space structure to obtain the optimal search space, including:

[0103] The sampling probability of each pre-trained candidate sub-model is predicted by the controller and sampling is performed to obtain the opcode of each pre-trained candidate sub-model. The corresponding opcodes are as follows: the Transfer cell network is used to search for shallow layers of different levels of different pre-trained models, the Normal cell network is used to search for candidate operations for feature extraction, the Pooling cell network is used to search for candidate operations for downsampling, and the Dropout cell network is used to search for the dropout rate P.

[0104] Construct MFC-NAS candidate models based on opcodes;

[0105] The system is trained using randomly selected training samples, and the controller parameters are updated using gradient descent based on the validation accuracy. The next search begins and continues until the search ends, obtaining the optimal search space. The Transfer Cell network is used to search for shallow layers at different levels of different pre-trained models, the Normal Cell network is used to search for candidate operations for feature extraction, the Pooling Cell network is used to search for candidate operations for downsampling, and the Dropout Cell network is used to search for the dropout rate P.

[0106] This embodiment is based on reinforcement learning to search for the optimal cell. For example... Figure 1 As shown, firstly, the controller predicts the sampling probability of each cell and performs sampling to obtain the opcode for each cell; then, an MFC-NAS candidate model is constructed based on the opcode. Next, the model is trained using randomly selected training samples, and the controller parameters are updated using policy gradient descent based on the validation accuracy, thus starting the next search, until the search ends. Specifically, the Transfer cell is used to search for shallow layers at different levels of different pre-trained models. The Normal cell is used to search for candidate operations for feature extraction. The Pooling cell is used to search for candidate operations for downsampling. The Dropout cell is used to search for the dropout rate P.

[0107] In this embodiment, the candidate MFC-NAS model is constructed based on opcodes. After the controller generates a list of opcodes for all cells, it constructs the corresponding cell according to the opcode information, such as... Figure 7As shown, each opcode list corresponds to a cell. First, the Transfer cell is constructed, and its opcode list corresponds to the candidate options in the Transfer cell's search space. Then, the Normal cell, Pooling cell, and Dropout cell are constructed sequentially. Similarly, the opcode lists of these cells also correspond to the candidate options in their respective search spaces. In the MFC-NAS model, the Pooling cell will pass through the Normal cell again until it reaches the last Pooling cell before entering the Dropout cell. The number of times it passes through the Normal cell is equal to the number of layers of the Pooling cell minus 1, where the number of layers of the Pooling cell is equal to the number of layers of the Normal cell (L).

[0108] This embodiment proposes a plant seedling image classification method based on MFC neural network architecture search, including:

[0109] Construct a dataset of plant seedling images;

[0110] In this embodiment, the dataset used is a plant seedling dataset from Aarhus University and the University of Southern Denmark. This dataset contains 12 plant seedling types, including 3 common crops and 9 weeds. For example... Figure 8 As shown in Table 11, the dataset contains images of 12 plant seedlings, including three common crops (4), Maize (7), and Sugar Beet (11), and nine weeds (0), Charlock (1), Cleavers (2), Common Chickweed (3), Fat Hen (5), Loose Silky-bent (6), Scentless Mayweed (8), Shepherd's Purse (9), and Small-flowered Cranesbill (10). The images were scaled, randomly flipped horizontally, randomly flipped vertically, randomly rotated, and standardized. The dataset was then divided into training, validation, and test sets in an 8:1:1 ratio.

[0111] Table 1. Distribution of seedling quantity for 1112 plant species

[0112]

[0113] The pre-defined MFC neural network architecture search model is trained based on the dataset to obtain a plant seedling classification model.

[0114] Plant seedling images are classified based on a plant seedling classification model.

[0115] Parameter configuration in this embodiment

[0116] During the search phase, the normal cell has 3 blocks, 2 layers, and 5 dropouts. The controller has 1000 training epochs, a learning rate of 0.001, and uses Adam as the optimizer. During the training phase of the classification model, the training epochs are set to 50, the batch size to 12, and the learning rate to 0.001. Cross-entropy loss is used as the training loss function, AdamW is used as the optimizer, and ReduceLROnPlateau is employed to automatically adjust the learning rate based on changes in validation accuracy. The image size is set to 224×224.

[0117] Evaluation indicators

[0118] In this embodiment, the mean average precision (mAP) is used as the evaluation index of the experiment, as shown in Formula (1), where Q represents the number of predicted samples and AP represents the average precision, as shown in Formula (2). It is represented by the PR curve to show the weighted average of the precision corresponding to n thresholds, where R represents the recall rate and P represents the precision rate.

[0119]

[0120]

[0121] Search for the best Cell

[0122] First, during the search phase, the three shared layers S1, S2, and S3 of the Transfer cell were experimentally compared to verify the impact of different values ​​of S on search performance. Figure 9 This represents the model training convergence speed of different shared layers S over an average of 500 controller samplings within 50 epochs. All three searchable layers contained in each shared layer S are sampled. When the shared layer is S1, the searched model converges the slowest, and when the shared layer is S3, the searched model converges the fastest.

[0123] As shown in Table 12, a deeper shared layer S has a better impact on search performance. When the shared layer is S1, the searched model has the worst performance, while when the shared layer is S3, the searched model has the best performance. Table 13 shows the experimental results for S3-L1, S3-L2, and S3-L3. It can be seen that an excessively large search layer leads to a decrease in model performance. When the shared layer is S3-L1, the experimental results are the best, with an optimal mAP of 99.17%. When the shared layer is S3-L2 and S3-L3, the highest mAP found is 98.84%, which is 0.33% lower than that of S3-L1. The number of parameters for the model searched in S3-L3 is approximately 10 times that of S3-L1.

[0124] Table 12 Experimental results for different shared layers S

[0125]

[0126]

[0127] Table 13 Experimental results for different searchable layers in the shared layer S3

[0128]

[0129] Based on multiple search experiments, the optimal transfer cell is EfficientNet (S3-L1), as follows: Figure 10 As shown, the optimal Normal cell is as follows: Figure 11 As shown, the optimal pooling cell is as follows: Figure 12 As shown, the optimal Dropout cell is as follows: Figure 13 As shown, and the optimal overall structure is as follows: Figure 14 As shown.

[0130] Plant seedling classification results

[0131] Table 15 compares the classification results of MFC-NAS with those of various other methods used in recent years for plant seedlings. MFC-NAS improves the model's feature extraction and generalization abilities through Transfer cell, Dropout cell, and Normal cell, thereby increasing the model's classification accuracy. Pooling cell reduces the number of model parameters. Therefore, compared with other methods, MFC-NAS has higher classification accuracy and fewer parameters. These experimental results demonstrate the feasibility of the proposed method for classifying plant seedlings.

[0132] Table 15 Comparison of experimental results between MFC-NAS and 10 other methods in recent years

[0133]

[0134]

[0135] This embodiment proposes a multi-functional cell-driven plant classification method based on NAS. This method focuses on designing transfer cells, normal cells, pooling cells, and dropout cells to improve the performance of the plant classification model. It can automatically search for the optimal model structure based on a plant seedling dataset to achieve high classification accuracy and a small number of model parameters. Furthermore, experimental results show that setting a larger number of blocks N1 in the transfer model and N2 in the normal cell, and a smaller number of layers L in the normal cell, can positively impact the classification results. This method can automatically design an effective convolutional neural network model for classifying plant seedling numbers, demonstrating better competitiveness and advancement compared to manually designed classification models in recent years.

[0136] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A plant seedling image classification method based on MFC neural network architecture search, characterized in that, include: Construct a dataset of plant seedling images; Construct a pre-defined search model based on the MFC neural network architecture; including: Construct an initial search space structure; wherein the initial search space structure includes: Transfer Cell network, Normal Cell network, Pooling Cell network and Dropout Cell network; The controller is used to search the initial search space structure to obtain the optimal search space; The optimal search spaces are stacked to obtain a model searched based on the MFC neural network architecture; The Transfer Cell network includes several pre-trained candidate sub-models, each of which includes several searchable layers and several different shared layers; wherein, the searchable layers are set at the end of the different shared layers, and each shared layer includes several searchable layers. The Transfer Cell network incorporates different shallow layers of different pre-trained candidate sub-models as weight transfer modules into the search space to automatically select shallow layers of different levels of different pre-trained candidate sub-models. The Normal Cell network includes a multi-branch structure composed of several preset first blocks, wherein each preset first block includes two OP candidate operation modules and one Combine connection candidate operation module; The OP candidate operation module uses the DSC sub-module based on the attention mechanism as candidate operations; The DSC submodule based on the attention mechanism uses DSC 3×3 and DSC 5×5 as candidates for the search space, and SE, CBAM and CA as candidates for the attention module. The Pooling Cell network includes: a preset second block, wherein the preset second block includes two Max Pooling candidate operation modules and one Combine connection candidate operation module; The Dropout Cell network includes: several Dropout modules with independent dropout rates P; the Dropout Cell network uses a controller to adaptively select preset Dropout modules and selects the best prediction result through a voting mechanism; Based on the dataset, a pre-defined model based on the MFC neural network architecture is trained to obtain a plant seedling classification model. Based on the plant seedling classification model, plant seedling images are classified.

2. The plant seedling image classification method based on MFC neural network architecture search according to claim 1, characterized in that, The pre-trained candidate sub-models include: MobileNet V3-Large sub-network, EfficientNet sub-network, ResNet-50 sub-network, Xception sub-network, and Inception-v3 sub-network.

3. The plant seedling image classification method based on MFC neural network architecture search according to claim 1, characterized in that, The process of using the controller to search the initial search space structure to obtain the optimal search space includes: The sampling probability of each pre-trained candidate sub-model is predicted using the controller and sampling is performed to obtain the operation code of each pre-trained candidate sub-model; wherein, the Transfer cell network is used to search the shallow layers of different levels of different pre-trained models, the Normal cell network is used to search for candidate operation for feature extraction, the Pooling cell network is used to search for candidate operation for downsampling, and the Dropout cell network is used to search for the dropout rate P. Construct an MFC-NAS candidate model based on the opcode; The system is trained using a randomly selected subset of training samples, and the controller parameters are updated using policy gradient descent based on the validation accuracy. The next search begins, and this process continues until the search ends, yielding the best multifunctional cell, namely the best Transfer cell, Normal cell, Pooling cell, and Dropout cell.

Images