Fine-grained flower image classification method based on neural network architecture search
By introducing an attention mechanism and a dense connection rule-based neural network architecture search into flower image classification, the network structure is optimized, solving the problems of low efficiency and poor accuracy in existing flower classification technologies, and achieving more efficient and accurate flower image classification.
Patent Information
- Application Number
- CN202310768604.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-27
- Publication Date
- 2026-02-24
- Estimated Expiration
- 2043-06-27
AI Technical Summary
Existing flower classification methods based on artificially designed features are inefficient, have poor accuracy, are easily affected by external light and artificially designed features, and are time-consuming and labor-intensive.
We employ a neural network architecture-based search method. By adding an attention mechanism to candidate convolution operations on top of DARTS, we construct a full attention-convolution search space and design dense connection rules for reduction cells. This optimizes the network structure to reduce information loss and fuse features at different scales.
On public flower classification datasets, it has fewer parameters and higher classification accuracy, outperforming traditional network models and network models based on weakly supervised methods, thus saving network design time and costs.
Smart Images
Figure CN116778248B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of neural network search, and more specifically, to a fine-grained flower image classification method based on neural network architecture search. Background Technology
[0002] Methods based on manually designed features and those based on deep learning feature extraction are currently the mainstream flower classification methods. Methods based on manually designed features are cumbersome, often requiring supervised segmentation of the original image to remove background information, thus obtaining regions containing only flower images. Then, texture features such as SIFT or HOG are extracted, and finally, a classifier is used to identify the flower species. Both flower image segmentation and feature extraction based on flower shape, texture, and color require manual design, which is very time-consuming and labor-intensive. Furthermore, the final classification performance is easily affected by external lighting and the quality of manually designed features. Therefore, methods based on manually designed features suffer from high subjectivity, low efficiency, and low accuracy. Summary of the Invention
[0003] To overcome the shortcomings and deficiencies of existing classification methods, this invention proposes a fine-grained flower image classification method based on Neural Architecture Search (NAS). Building upon Differentiable Architecture Search (DARTS), an attention mechanism is added before each candidate convolution operation in the search space. This constructs an attention-convolution module as a new candidate operation, forming a full attention-convolution search space that enhances the network's focus on salient features. Furthermore, a connection rule with more input nodes and denser connections between nodes is designed for the reduction cell, reducing information loss due to downsampling while fusing features at different scales.
[0004] To achieve the above objectives, the present invention provides the following technical solution:
[0005] A fine-grained flower image classification method based on neural network architecture search includes the following steps:
[0006] S1: Collection of flower images;
[0007] S2: Construct attention-convolution modules as candidate operations or design dense connection rules for reduction cells in neural network architectures;
[0008] S3: When constructing attention-convolution modules as candidate operations, the full attention-convolution search space is formed by various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed; when designing dense connection rules for reduction cells under the neural network architecture, the full attention-convolution search space is formed by various candidate operations, and the reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture.
[0009] S4: In the full attention-convolution search space, a gradient descent-based search strategy is used to search for the best cell in the reduction cell and normal cell respectively;
[0010] S5: Stack the two best cells found to build the final flower image classification model and obtain the final classification model;
[0011] S6: Use the final classification model to classify the flower image set.
[0012] Furthermore, in step S2, the construction of the attention-convolution module specifically involves: the input of the attention-convolution module passing through the attention mechanism and then through the ReLU activation function, and then through a 3×3 convolutional layer to the output of the BatchNorm regularization operation.
[0013] Furthermore, in step S2, the design of dense connection rules for the reduction cell under the neural network architecture specifically involves:
[0014] More input nodes and denser connection rules were designed for the reduction cell in the neural network architecture.
[0015] Furthermore, the reduction cell contains nine nodes, including four input nodes, four intermediate nodes, and one output node. Each intermediate node is connected to four forward nodes, which are either input nodes or output nodes.
[0016] Furthermore, in step S3, the method for constructing the full attention-convolution search space is specifically as follows:
[0017] When using attention-convolution modules, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation, average pooling operation, and attention-convolution modules;
[0018] When the attention-convolution module is not used, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation, and average pooling operation.
[0019] Furthermore, in step S4, a gradient descent-based search strategy is used to search for the optimal cell in the full attention-convolution search space, specifically as follows:
[0020] S4.1 Assign weights to each edge connected to each node of the normal cell and the rededuction cell in the neural network architecture and to each candidate operation, and refer to the weights as architecture parameters;
[0021] S4.2 updates the architecture parameters using the gradient descent algorithm on a flower image dataset, with accuracy as the metric.
[0022] S4.3: The normal cell with the largest edge and the largest candidate operation in the normal cell is the best normal cell; the reduction cell with the largest edge and the largest candidate operation in the reduction cell is the best reduction cell.
[0023] Furthermore, the final flower image classification model includes network models constructed from the two optimal cells searched in step S4.3: the Attention-NAS-A network model or the Attention-NAS-B network model. The Attention-NAS-A network model and the Attention-NAS-B network model have different optimal reduction cell and normal cell stacking methods.
[0024] Furthermore, the Attention-NAS-A network model is constructed as follows:
[0025] We select the 12 best normal cells and the 2 best reduction cells. The first three layers are depthwise separable convolutional layers. We place reduction cells at layers 7 and 12 of the network model, and normal cells at the remaining layers. Finally, we place a global average pooling layer and a fully connected layer at the end of the Attention-NAS-A network model.
[0026] Furthermore, the Attention-NAS-B network model is constructed as follows: The Attention-NAS-B network model has 12 normal cells and 2 reduction cells. The first three layers are depthwise separable convolutional layers. Reduction cells are placed in the 12th and 17th layers of the network model, and normal cells are placed in the remaining layers. The Attention-NAS-B network model is finally filled with a global average pooling layer and a fully connected layer.
[0027] Furthermore, the feature is that, in step S2, attention-convolution modules are constructed simultaneously as candidate operations and dense connection rules are designed for reduction cells under the neural network architecture; in step S3, a full attention-convolution search space is constructed from various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture.
[0028] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:
[0029] The main contributions of this invention are twofold:
[0030] This invention first adds an attention mechanism before each candidate convolution operation in the search space, constructs an attention-convolution module as a new candidate operation, and forms a full attention-convolution search space. Then, it designs connection rules for the reduction cell with more input nodes and denser connections between intermediate nodes. Finally, it adjusts the position of the optimal reduction cell to obtain two network models with different parameter sizes, thereby deploying them to more terminal devices.
[0031] The two network models discovered by this invention have only 4.07M and 2.3M parameters, respectively. On two public flower classification datasets, they have advantages in both accuracy and parameter quantity compared with traditional network models. They also achieve an advanced level of accuracy compared with flower classification network models based on weak supervision methods. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the fine-grained flower image classification method based on neural network architecture search described in this invention.
[0033] Figure 2 This is a flowchart of the fine-grained flower image classification method based on neural network architecture search as described in an embodiment of the present invention;
[0034] Figure 3This is a schematic diagram of the candidate operation structure of the 3x3 attention-convolution module according to an embodiment of the present invention;
[0035] Figure 4 This refers to the reduction cell structure of the existing DARTS technology;
[0036] Figure 5 This refers to the reduction cell under the dense connection rule provided in the embodiments of the present invention;
[0037] Figure 6 This invention describes two different network architectures constructed using stacking methods.
[0038] Figure 7 This is an example of an image from the Flower17 dataset provided in an embodiment of the present invention;
[0039] Figure 8 This is an example of an image from the Oxford102 dataset provided in an embodiment of the present invention;
[0040] Figure 9 The optimal normal cell provided for the embodiments of the present invention;
[0041] Figure 10 The optimal reduction cell provided for embodiments of the present invention. Detailed Implementation
[0042] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.
[0043] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions.
[0044] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0046] Example 1
[0047] This embodiment provides a fine-grained flower image classification method based on neural network architecture search, such as... Figure 1 As shown, it includes the following steps:
[0048] S1: Collection of flower images;
[0049] S2: Construct attention-convolution modules as candidate operations or design dense connection rules for reduction cells in neural network architectures;
[0050] S3: When constructing attention-convolution modules as candidate operations, the full attention-convolution search space is formed by various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed; when designing dense connection rules for reduction cells under the neural network architecture, the full attention-convolution search space is formed by various candidate operations, and the reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture.
[0051] S4: In the full attention-convolution search space, a gradient descent-based search strategy is used to search for the best cell in the reduction cell and normal cell respectively;
[0052] S5: Stack the two best cells found to build the final flower image classification model and obtain the final classification model;
[0053] S6: Use the final classification model to classify the flower image set;
[0054] like Figure 2 As shown, the fine-grained flower image classification method based on neural network architecture search can be divided into three steps: improving the search space, searching for the best cell, and the best cell stack.
[0055] Example 2
[0056] This embodiment, based on Embodiment 1, continues to disclose the following content:
[0057] like Figure 3 As shown, in step S2, in this embodiment, the construction of the 3x3 attention-convolution module is specifically as follows: the input of the attention-convolution module passes through the attention mechanism and then through the ReLU activation function, and then through a 3×3 convolutional layer to the output of the BatchNorm regularization operation.
[0058] Example 3
[0059] Based on Examples 1 and 2, this embodiment continues to disclose the following content:
[0060] More input nodes and denser connection rules were designed for the reduction cell in the neural network architecture;
[0061] like Figure 5As shown, the reduction cell contains nine nodes, including four input nodes, four intermediate nodes, and one output node. Each intermediate node is connected to four forward nodes, which are either input nodes or output nodes. As described above, it is consistent with... Figure 4 As shown, the existing DARTS reduction cell structure has denser connections between nodes compared to the newly designed connection rules.
[0062] Example 4
[0063] This embodiment, based on Embodiment 1, continues to disclose the following content:
[0064] In step S3, the method for constructing the full attention-convolution search space is as follows:
[0065] When using attention-convolution modules, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation, average pooling operation, and attention-convolution modules;
[0066] When the attention-convolution module is not used, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation and average pooling operation. Other candidate operations are included in the search space as shown in Table 1.
[0067] Table 1 Candidate Operation Set
[0068]
[0069]
[0070] Example 5
[0071] This embodiment, based on Embodiment 1, continues to disclose the following content:
[0072] In step S4, a gradient descent-based search strategy is used to search for the optimal cell in the full attention-convolution search space, specifically as follows:
[0073] S4.1 Assign weights to each edge connected to each node of the normal cell and the rededuction cell in the neural network architecture and to each candidate operation, and refer to the weights as architecture parameters;
[0074] S4.2 updates the architecture parameters using the gradient descent algorithm on a flower image dataset, with accuracy as the metric.
[0075] S4.3: The optimal normal cell is the one with the largest edge and the largest candidate operation in the normal cell. Figure 9 As shown; the optimal reduction cell is the one with the largest edge and the largest candidate operation in the reduction cell. Figure 10 As shown, the final flower image classification model includes network models constructed from the two optimal cells searched in step S4.3: the Attention-NAS-A network model or the Attention-NAS-B network model, as follows: Figure 6 As shown, the Attention-NAS-A and Attention-NAS-B network models have different optimal reduction cell and normal cell stacking methods. Both stacking methods ultimately use global average pooling and fully connected layers for classification.
[0076] The Attention-NAS-A network model is constructed as follows:
[0077] We select 12 optimal normal cells and 2 optimal reduction cells. The first three layers are depthwise separable convolutional layers. We place reduction cells at layers 7 and 12 of the network model and normal cells at the remaining layers. Finally, we place a global average pooling layer and a fully connected layer in the Attention-NAS-A network model.
[0078] The Attention-NAS-B network model is constructed as follows: The Attention-NAS-B network model has 12 normal cells and 2 reduction cells. The first three layers are depthwise separable convolutional layers. Reduction cells are placed in the 12th and 17th layers of the network model, and normal cells are placed in the remaining layers. The Attention-NAS-B network model is finally filled with a global average pooling layer and a fully connected layer.
[0079] In step S2, attention-convolution modules are constructed simultaneously as candidate operations, and dense connection rules are designed for reduction cells under the neural network architecture. In step S3, a full attention-convolution search space is constructed from various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture.
[0080] The following section will use the Flowers17 flower classification dataset and the Oxford102 flower dataset as examples to illustrate this point. Figure 7 As shown, the Flowers17 flower classification dataset contains 17 common flowers from the UK, including sunflowers, hyacinths, daffodils, and chrysanthemums. Each category has 80 images with different poses, sizes, and angles. The training and test sets are divided in an 8:2 ratio. Figure 8 As shown, the Oxford102 flower dataset contains 102 flower species, with 40 to 258 images for each species, totaling 40,258 images. The dataset is divided in an 8:2 ratio. The experiments were implemented using PyTorch 1.10.0, with a system environment of Win10, Intel i7 CPU, 16GB DDR4 RAM, and GTX1070Ti GPU with 8GB of RAM. The experiments included search and validation phases. In the search phase, the best cell was first searched using a search space constructed with three different attention mechanisms on the Flowers17 flower dataset. Experiments were then conducted on the Flowers17 dataset to determine the best attention mechanism. Then, ablation experiments were performed to evaluate the effectiveness of each improved method. In the validation phase, the searched classification network model was compared with traditional artificial neural network models and related flower classification neural network models on multiple flower classification datasets.
[0081] During the search phase, a 224x224 image was used as input, the initial learning rate of the controller was 0.1, the search epochs were 50, the optimization algorithm was SGD, the batch size was 8, and other parameters were the same as DARTS. During the validation phase, a 224x224 image was used as input, the optimization algorithm was SGD, the batch size was 16, the initial learning rate was 0.1, and the training lasted for 500 epochs.
[0082] Using the Flower17 flower image dataset as a proxy dataset, the best cell was searched. To introduce a more suitable attention mechanism, three attention mechanisms—SE, CBAM, and ECA—were introduced into the search space to search for the best cell and construct a neural network model, as shown in Table 2. Compared with the neural network model searched in the original search space, the neural network model searched in the search space with the ECA attention mechanism decreased the Flower17 classification accuracy by 0.83%, the neural network model searched in the search space with the CBAM attention mechanism only improved the Flower17 classification accuracy by 0.07%, while the neural network model searched in the search space with the SE attention mechanism improved the Flower17 classification accuracy by up to 1.48%. Therefore, the search space with the SE attention mechanism can find a better neural network model, with an accuracy 1.41% and 2.31% higher than the neural network models searched with the other two attention mechanisms, respectively.
[0083] Table 2 Results of introducing different attention mechanisms into the search space
[0084]
[0085] Figure 10 This shows the optimal reduction cell found on the Flower17 dataset in the search space constructed by introducing the SE attention mechanism. Figure 9 The optimal normal cell is shown. Hi-1 and Hi-2 represent the outputs of the previous two cells, which serve as the two inputs to the current cell. Hi represents the output of the current cell. The optimal normal cell contains two skip connection operations and four attention-convolution operations.
[0086] The optimal reduction cells, Hi-1, Hi-2, Hi-3, and Hi-4, represent the outputs of the previous four cells and serve as the four inputs to the current cell. Hi represents the output of the current cell. The optimal reduction cell contains seven max pooling operations, two skip connection operations, and six attention-convolution operations. Because the reduction cell has more inputs, it retains more shallow features, reduces the information loss caused by downsampling, and forms denser connections between intermediate nodes through more different candidate operations, thus fusing features of more different scales. Therefore, when the new reduction cell participates in the construction of the network model, it has a higher classification accuracy in the flower classification task.
[0087] As shown in Table 3, to verify the effectiveness of the two improved methods proposed in this invention, four sets of comparative experiments were conducted on the Flower17 flower image dataset. When only the attention mechanism was introduced to construct the search space (SE), the final accuracy was 91.92%. When only dense connection rules were designed for the reduction cell, the final accuracy was 92.64%. When both improved methods were retained, the final classification result improved to 94.12%, which is higher than the 90.44% without any improved methods. Therefore, the ablation experiment proved the effectiveness of the improved methods of this invention.
[0088] Table 3 evaluates the effectiveness of different improvement methods.
[0089]
[0090] Model Comparison
[0091] Two optimal cells were searched using the search space with an introduced SE attention mechanism, and two different stacking methods were used to construct two network models, SE-NAS-A and SE-NAS-B. The effectiveness of the proposed method was verified by comparing it with traditional neural network models and flower classification network models based on weak supervision methods on the Oxford102 flower dataset and the Flower17 flower dataset.
[0092] Comparison with traditional neural network models
[0093] Table 4 shows the neural network models (SE-NAS-A and SE-NAS-B) searched by the method proposed in this invention on two flower image datasets, compared with multiple traditional neural networks.
[0094] Table 4 Comparison with traditional neural network models
[0095]
[0096] As shown in Table 4, the neural network model searched by the method proposed in this invention is far superior to traditional network models in terms of both parameter count and accuracy. On the Oxford102 dataset, the SE-NAS-A network model achieved the highest accuracy (96.13%), and the SE-NAS-B network model achieved the second highest accuracy (94.85%). Compared with the Vgg16 network model with the largest number of parameters, the classification accuracy was improved by 15.23% and 13.94% respectively, while the number of parameters was only 0.31% and 0.15% of that of the Vgg16 network model. Compared with the traditional network model ResNet50 with the highest accuracy, the accuracy was improved by 10.83% and 9.54% respectively, and the number of model parameters was only 17.16% and 9.7% of that of the ResNet50 network.
[0097] Compared to MobileNet, the traditional network model with the fewest parameters, the SE-NAS-A and AIDA-NAS-B network models not only have fewer parameters, but also improve accuracy by 17.84% and 16.55%, respectively.
[0098] The SE-NAS-A and SE-NAS-B network models also achieved similar classification results (94.12% and 92.64%) on the Flower17 dataset as on the Oxford102 flower dataset, respectively, improving upon traditional network models by 8.47%-17.65% and 6.99%-16.17%. The comparative results show that the proposed method achieves higher classification accuracy in flower classification with fewer parameters compared to traditional network models. A comparison with flower classification network models based on weakly supervised methods is also provided.
[0099] Table 5 shows the comparison results of the proposed method and the flower classification network model based on the weak supervision method on two flower image classification datasets. References 1-6 are all manually designed network models for flower classification tasks.
[0100] Table 5 Comparison with Flower Classification Network Model
[0101]
[0102] As shown in Table 5, Reference 1 (Xiong Juju, Xu Yang, Fan Runze, et al. Lightweight visual Transformer-based flower recognition[J]. Journal of Graphics, 1-9 (in Chinese) introduces the PoolFrome module based on moving windows and the dual-channel attention mechanism into the Swin-T network to improve the network's ability to extract global features, achieving good classification results; Reference 2 (Yin H, Fu Xiang, Zeng Jixian, et al. Selective convolutional feature fusion for floral image classification[J]. Chinese Journal of Graphics, 2019, 24(05):762-72 (in Chinese) al. Selective convolutional feature fusion for flower image classification [J]. Journal of Image and Graphics, 2019, 24(05): 762-72)) The method of fusion of shallow and deep features effectively improves the accuracy of flower image classification compared with the traditional neural network model. Reference 3 (Yang Wanggong, Huai Yongjian. Multi-layer feature fusion and floral image classification with regions of interest [J]. Journal of Harbin Engineering University, 2021, 42(04): 588-94)) By fusing multi-layer image features and using a joint uniformly distributed cross-entropy loss function, the classification accuracy was effectively improved on multiple flower classification datasets, reaching 93.57% and 85.47% respectively. Reference 4 (XIAOLING X, CUI X, BINGN)Inception-v3 for flower classification; procedures of the 2017 2nd International Conference on Image, Vision and Computing (ICIVC), F 2-4 June 2017, 2017 [C]) achieved classification accuracy of 95% and 94% respectively on two flower classification datasets using a transfer learning-based method. Reference 5 (C1BUK M, BUDAK U, GUO Y, et al. Efficient deep features selections and classification for flower species recognition [J]. Measurement, 2019, 137(7-13)) achieved higher classification accuracy on the Flower17 dataset than the method proposed in this invention, and also achieved similar classification accuracy on the Oxford102 classification dataset. However, the network structure in Reference 5 was designed by researchers based on experience, which took a lot of time, while the network model of this invention is automatically constructed based on the flower classification task, greatly saving the time cost of designing the network model. Reference 6 (PEREIRA-FERRERO VH, VALEM LP, PEDRONETTE D CG. Feature augmentation based on Manifold ranking and LSTM for image classification[J]. Expert Systems with Applications, 2023, 213(118995). Although it achieved the highest classification accuracy on the Flower17 dataset, it only achieved 84.64% accuracy on the Oxford102 dataset.
[0103] The method proposed in this invention achieved the highest classification accuracy of 96.14% on the Oxford102 dataset and 94.12% on the Flower17 dataset, reaching an advanced level compared with manually designed network models; the results show that the flower image classification method based on neural network architecture search proposed in this invention is effective.
[0104] The same or similar labels correspond to the same or similar parts;
[0105] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.
[0106] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A fine-grained flower image classification method based on neural network architecture search, characterized in that, Includes the following steps: S1: Collect flower images; S2: Construct attention-convolution modules as candidate operations or design dense connection rules for reduction cells in neural network architectures; S3: When constructing attention-convolution modules as candidate operations, the full attention-convolution search space is formed by various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed; when designing dense connection rules for reduction cells under the neural network architecture, the full attention-convolution search space is formed by various candidate operations, and reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture. S4: In the full attention-convolution search space, a gradient descent-based search strategy is used to search for the best cell in the reduction cell and normal cell respectively; S5: Stack the two best cells found to build the final flower image classification model and obtain the final classification model; S6: Use the final classification model to classify the flower image set; In step S3, the method for constructing the full attention-convolution search space is as follows: When using attention-convolution modules, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation, average pooling operation, and attention-convolution modules; When the attention-convolution module is not used, the full attention-convolution search space also includes: depthwise separable convolution, skip connection operation, max pooling operation and average pooling operation; In the full attention-convolution search space, a gradient descent-based search strategy is used to search for the optimal cell, specifically: S4.1 Assign weights to the edges connected to each node of the normal cell and the rededuction cell in the neural network architecture and to each candidate operation, and refer to the weights as architecture parameters; S4.2 updates the architecture parameters using the gradient descent algorithm on a flower image dataset, with accuracy as the metric. S4.3: The normal cell with the largest edge and the largest candidate operation in the normal cell is the best normal cell; the reduction cell with the largest edge and the largest candidate operation in the reduction cell is the best reduction cell. The final flower image classification model includes network models constructed from the two optimal cells searched in step S4.3: Attention-NAS-A network model or Attention-NAS-B network model. The Attention-NAS-A network model and the Attention-NAS-B network model have different optimal reduction cell and normal cell stacking methods. The Attention-NAS-A network model is constructed as follows: We select the 12 best normal cells and the 2 best reduction cells. The first three layers are depthwise separable convolutional layers. We place reduction cells at the 7th and 12th layers of the network model and normal cells at the remaining layers. The Attention-NAS-A network model is finally filled with global average pooling layers and fully connected layers. The Attention-NAS-B network model is constructed as follows: The Attention-NAS-B network model consists of 12 normal cells and 2 reduction cells. The first three layers are depthwise separable convolutional layers. Reduction cells are placed on the 12th and 17th layers of the network model, and normal cells are placed on the remaining layers. The Attention-NAS-B network model is finally filled with a global average pooling layer and a fully connected layer.
2. The fine-grained flower image classification method based on neural network architecture search according to claim 1, characterized in that, In step S2, the construction of the attention-convolution module specifically involves: The input to the attention-convolution module passes through the attention mechanism, then the ReLU activation function, and finally through a 3×3 convolutional layer to the BatchNorm regularization operation for output.
3. The fine-grained flower image classification method based on neural network architecture search according to claim 1, characterized in that, In step S2, the design of dense connection rules for the reduction cell under the neural network architecture specifically involves: More input nodes and denser connection rules were designed for the reduction cell in the neural network architecture.
4. The fine-grained flower image classification method based on neural network architecture search according to claim 3, characterized in that, The reduction cell contains nine nodes, including four input nodes, four intermediate nodes, and one output node. Each intermediate node is connected to four forward nodes, which are either input nodes or output nodes.
5. The fine-grained flower image classification method based on neural network architecture search according to any one of claims 1 to 4, characterized in that, In step S2, attention-convolution modules are constructed simultaneously as candidate operations, and dense connection rules are designed for reduction cells under the neural network architecture. In step S3, a full attention-convolution search space is constructed from various candidate operations including attention-convolution modules, and reduction cells and normal cells are constructed using the dense connection rules designed for reduction cells under the neural network architecture.