Lightweight tea disease identification method based on improved OfficientNet

Abstract

The invention discloses a lightweight tea disease identification method based on an improved OfficientNet. The lightweight tea disease identification method comprises the following steps: step 1, using a digital camera to collect image data of tea diseases in different time periods every day; 2, performing enhancement processing on the image data in the original data set A to generate an enhanced data set B; 3, the enhanced data set B is segmented according to the proportion of 8: 2, 20% of data is used as a test set, 80% of data is used as a training set, after training is completed through the data set, a weight file of the model is obtained, and intelligent recognition can be achieved through the weight file. According to the method, an efficient CFA channel focusing attention mechanism is provided, the focusing capacity of the model for important disease characteristics under complex background interference is enhanced, the purpose of improving the disease recognition accuracy is achieved, the model recognition accuracy is improved, meanwhile, the parameter quantity and the calculation quantity of the model are reduced, and the model is made to be lighter.

Description

Technical Field

[0001] The present invention relates to the technical field of tea disease identification, and particularly relates to a lightweight tea disease identification method based on improved EfficientNet. Background Art

[0002] Since tea is vulnerable to various factors such as climate, pests, and pathogens, it leads to the occurrence of diseases and a reduction in production. Therefore, it is crucial to promptly detect tea diseases and take targeted measures. Traditional tea disease identification mainly relies on manual observation. Due to the large variety of tea diseases, it is difficult to achieve manual identification. In recent years, with the widespread popularity of inexpensive edge computing devices with camera functions, artificial intelligence technology has been widely applied in the field of disease identification, making it possible to automatically identify tea diseases.

[0003] However, in order to improve the recognition accuracy of existing deep learning models, they usually have complex model structures. And complex model structures will consume a large amount of computing resources, resulting in slow operation on edge computing devices. Therefore, there is an urgent need for a method that can run quickly on edge computing devices and accurately identify tea diseases to improve the efficiency of field disease diagnosis. Summary of the Invention

[0004] In order to improve the recognition accuracy of deep learning models and solve the problem of their slow operation on edge computing devices, the present invention proposes a lightweight tea disease identification method based on improved EfficientNet.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A lightweight tea disease identification method based on improved EfficientNet, comprising the following steps:

[0007] Step 1: Use a digital camera to collect image data of tea diseases at different time periods every day. Based on the collected data, manual label annotation is carried out to form the original dataset A for model training;

[0008] Step 2: Perform enhancement processing on the image data in the original dataset A to generate the enhanced dataset B. This step aims to improve the robustness and generalization ability of the model, reduce the model's dependence on specific attributes, and effectively solve the problem of insufficient data samples;

[0009] Step 3: Divide the enhanced dataset B in a ratio of 8:2, where 20% of the data is used as the test set and 80% of the data is used for the training set;

[0010] Step 4: Based on the improved EfficientNet-Lite0, construct a lightweight model named TDN for tea disease recognition. The lightweight model for tea disease recognition includes a Stem feature reuse module layer, a FEMBCM feature enhancement mobile inverted bottleneck convolutional layer, a Head feature reuse module layer, and an average pooling layer;

[0011] Step 5: Utilize the dataset B obtained in Step 2 to obtain the optimal model weights during multiple trainings. Then, based on the obtained optimal model weights, conduct the recognition of tea diseases.

[0012] Furthermore, the tea diseases in Step 1 can be divided into 8 categories: algal leaf spot, tea white spot, tea black spot, tea rust, tea leafminer, red spider on tea, dodder on tea, and healthy tea leaf images. When collecting these disease images, select different time periods for collection to obtain disease images under different lighting conditions.

[0013] Furthermore, in Step 2, perform enhancement operations on the images in the original dataset A. This enhancement operation includes image flipping, changing the hue and saturation of the image, and grayscaling the picture. Among them, image flipping means vertically flipping the collected image to eliminate the model's dependence on a specific direction; changing the hue and saturation of the image adjusts the hue and saturation of the image to simulate images under different lighting and environmental conditions; grayscaling the picture converts a color image into a grayscale image to reduce the complexity of the model and improve the model's recognition ability for disease structural features.

[0014] Furthermore, the Stem feature reuse module layer in Step 4 is composed of an FRM feature reuse module and a CFA channel focus attention mechanism in a serial structure. In the Stem feature reuse module layer, first use a feature reuse module with a convolutional kernel size of 3 and a stride of 2 to extract the overall feature information of the input image. Compared with the traditional convolutional layer, the feature reuse module has lower computational costs and the number of parameters, which can make the model more lightweight. Then, we use the CFA channel focus attention mechanism after the feature reuse module to further focus on the key features of the disease and enhance the model's expression ability.

[0015] Furthermore, the FEMBCM feature-enhanced mobile inverted bottleneck convolutional layer described in step 4 is composed of 16 FEMBCM feature-enhanced mobile inverted bottleneck convolutional modules. Different from traditional convolutional layers with only feature extraction structures, the FEMBCM consists of an FEB feature extraction branch and a DCB detailed feature compensation branch. Among them, the FEB feature extraction branch uses MBConv (mobile inverted bottleneck convolution) to extract disease features, aiming to control the model parameters and prevent the increase of model parameters. The DCB detailed feature compensation branch is used to further compensate for detailed features such as disease edge textures, making up for the loss of detailed features caused by downsampling. At the end of the FEMBCM, the feature information extracted by the above two branches is fused and then passed as input to the next module.

[0016] Furthermore, the Head feature reuse module layer and the Stem feature reuse module layer described in step 4 have the same structure. The difference is that the Head feature reuse module layer combines the feature information extracted from the previous two layers into richer feature information through the FRM (feature reuse module) with a convolutional kernel of 7. The purpose of this is to further improve the generalization ability of the model, enabling the model to better adapt to different datasets and recognition tasks. At the same time, using a convolution with a convolutional kernel of 7 can represent the same-sized receptive field with relatively fewer parameters, helping to reduce the model's parameter quantity and making the model more lightweight.

[0017] Furthermore, the CFA (channel attention mechanism) described in step 4 consists of two parallel modules: the KFEM (core feature extraction module) and the FFM (feature focusing module). The CFA attention mechanism can improve the model's focusing ability on important disease features, suppress the interference of unimportant features, and enable the model to adapt to natural environments with various complex background interferences.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. In the TDN model, an efficient CFA channel focusing attention mechanism is proposed, which enhances the model's focusing ability on important disease features under complex background interferences, achieving the purpose of improving the disease recognition accuracy.

[0020] 2. In the TDN model, an FRM feature reuse module is proposed, and FRC is used to replace traditional convolution. FRC improves the model's recognition accuracy while reducing the model's parameter quantity and computational complexity, making the model more lightweight.

[0021] 3. In the TDN model, the FEMBCM feature-enhanced mobile flip bottleneck convolution module is proposed. This module consists of a feature extraction branch and a local detail feature compensation branch. Among them, by using the local detail feature compensation branch, the loss of important detail features caused by the convolution downsampling operation is compensated, further improving the accuracy of the model in tea disease recognition.

[0022] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly and implement it according to the content of the specification, the following takes the preferred embodiments of the present invention and combines the drawings to describe in detail as follows. The specific implementation manner of the present invention is given in detail by the following embodiments and their drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present invention, and constitute a part of this application. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0024] Figure 1 is a flowchart of a lightweight tea disease recognition method based on improved EfficientNet;

[0025] Figure 2 is a structural diagram of a lightweight model for tea disease recognition using the TDN lightweight tea disease recognition method constructed by the present invention;

[0026] Figure 3 is a structural diagram of the Stem feature reuse module and the Head feature reuse module constructed by the present invention;

[0027] Figure 4 is a structural diagram of the CFA channel attention mechanism constructed by the present invention;

[0028] Figure 5 is a structural diagram of the FEMBCM feature-enhanced mobile inversion bottleneck convolution module constructed by the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] The principles and features of the present invention are described below with reference to the drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention. In the following paragraphs, the present invention is described more specifically by way of example with reference to the drawings. It should be noted that the drawings are all in a very simplified form and use non-precise scales, only for the purpose of conveniently and clearly assisting in explaining the embodiments of the present invention.

[0030] As Figure 1 shown: A lightweight tea disease recognition method based on improved EfficientNet, the key of which is to include the following steps:

[0031] Step 1: First, use a camera device to collect the original images of tea leaf diseases to form an image dataset A, and pass the dataset A to the next step for data preprocessing and enhancement operations;

[0032] Step 2: Receive the original image dataset A from Step 1, and perform image enhancement operations on this dataset, such as image flipping, changing image hue and saturation, and grayscaling the pictures, to expand the richness of dataset A and improve the generalization of the model. Among them, there are a total of 1771 images in the original dataset A, and the number of images in dataset B after data enhancement operations increases to 7080.

[0033] Step 3: Divide dataset B into a training set and a validation set according to a ratio of 8:2. Among them, the training set has 5664 images and the validation set has 1416 images.

[0034] Step 4: Build a lightweight model TDN for tea leaf disease recognition, and iteratively train the model based on the training set and validation set input in Step 3; then use the validation set to verify the training results of the tea leaf disease recognition model, evaluate the model performance using performance evaluation indicators, and adjust the model structure and training strategy according to the evaluation results; finally, save the optimal weights of the lightweight model TDN for tea leaf disease recognition for subsequent disease recognition work.

[0035] Step 5: Use the optimal weights of TDN saved in Step 4 to detect tea leaf disease images and output the final tea leaf disease classification results.

[0036] In Step 4, the lightweight tea leaf disease recognition model TDN is used to identify tea leaf disease images to obtain the classification results of the diseases. As Figure 2 shown, the input tea leaf disease feature map needs to go through the feature extraction and focusing operations of the Stem feature reuse module layer, the feature enhancement mobile reverse bottleneck convolution layer, the Head feature reuse module layer, and the CFA channel attention mechanism module in sequence, and finally the classification results of the diseases are output by the average pooling layer. When the above modules perform tea leaf disease recognition, the specific steps are as follows:

[0037] Step A1 (Stem): Extract the feature information of the diseases from the input feature map through the Stem feature reuse module to achieve the purpose of improving the model recognition accuracy and reducing the model calculation amount. As Figure 3 shown, in the Stem feature reuse module, the input feature map first extracts the feature information of the diseases by the FRM feature reuse module, and then the CFA channel focusing attention mechanism module further focuses on the key feature information of the diseases in the disease feature map.

[0038] Specifically, first, the input feature map is first subjected to feature extraction through an FRM feature reuse module with a convolution kernel size of 3. As Figure 3 shown, the structure of the FRM feature reuse module consists of an FCSS feature selection sub-module, an FES feature extraction sub-module, and an FRS feature reuse sub-module. The specific operation steps are as follows:

[0039] First, the feature selection sub-module first calculates the attention weights of each channel in the input feature map through an average pooling layer, a 1*1 convolution, and a Relu activation function. Then, the key feature channels of the lesions are screened from all channels of the input feature map. Thus, the noise feature channels caused by the complex background are discarded to reduce the interference of noise on the model recognition accuracy and reduce the computational amount. The selected key feature channels will be fed into the next sub-module for feature extraction.

[0040] Second, the feature extraction sub-module extracts a global feature map of the lesions with m channels from the key feature channels of the lesions by using a convolution with a kernel size of 2 to enhance the feature expression ability of the model. The extracted global feature map is input into the next FRS feature reuse sub-module for further feature extraction.

[0041] Third, the FRS feature reuse sub-module obtains the global feature map of the lesions with m channels provided by the upstream, while the downstream module requires the input channel number to be C. To meet this requirement, the module needs to generate an additional n (n = C - m) channels of feature map Y' based on the existing m channels of the key feature map Y of the lesions using a linear transformation function. Then, the two feature maps are concatenated into a complete output feature map with a channel number of C.

[0042] Secondly, the resulting feature map output by the FRM feature reuse module is input into the CFA attention mechanism module for further feature focusing operations. As Figure 4As shown in the figure, the CFA attention mechanism module consists of two parallel modules: the KFEM core feature extraction module and the FFM feature focusing module. The specific operation process is as follows: First, the input feature maps are respectively input into the KFEM core feature extraction module and the FFM feature focusing module to extract the key features of the disease image and the focused features of the lesion. In the KFEM core feature extraction module, the maximum pooling layer Maxpool is used to extract the disease core features of the input feature map; in the FFM feature focusing module, the global information is extracted by using the average pooling layer Avgpool to generate the feature representation in the channel dimension, and then the 1*1 convolutional layer is used to further perform feature mapping on these channel features to generate the channel feature weights. Finally, the channel feature weights are multiplied by the input feature map to obtain the focused feature map. Second, the extracted key feature map and the focused feature map are fused by element-wise addition, and finally the output feature map after feature focusing by the CFA attention mechanism module is obtained.

[0043] Step A2 (FEMBCM): The input image is passed through the FEMBCM feature enhancement convolutional module to extract the feature information of the disease, aiming to improve the recognition accuracy of the model and reduce the computational burden. As Figure 5 shown in the figure, in the FEMBCM feature enhancement convolutional module, the input image extracts the global features and enhances the local detail features in parallel through the FEB feature extraction branch and the DCB detail feature compensation branch. Among them, the FEMBCM feature enhancement mobile inverted bottleneck convolutional layer consists of 16 FEMBCM feature enhancement mobile inverted bottleneck convolutional modules, and the specific operation steps are as follows:

[0044] First, the FEB feature extraction branch uses a 1*1 convolutional layer, a depthwise separable convolutional layer, and a 1*1 convolutional layer in a serial structure to extract the global features. Compared with the traditional convolution, the mobile inverted bottleneck convolution has fewer parameter numbers and can effectively reduce the computational amount of the model.

[0045] Second, the DCB detail feature compensation branch uses an average pooling layer, a 1*1 convolutional layer, a Sigmoid activation function, and a maximum pooling layer in a serial structure to extract the local detail features of the input image. The DCB detail feature compensation branch can enhance the local detail feature information such as texture and edge, and effectively compensates for the loss of important detail features caused by the convolutional downsampling operation.

[0046] Finally, the feature map output by the FEB feature extraction branch and the feature map output by the DCB detail feature compensation branch are added to obtain the final output feature map.

[0047] Step A3 (Head): Pass the input feature map through the Head feature reuse module layer to integrate the feature information extracted in Steps A1 and A2, so as to further enhance the model's feature expression ability and improve its generalization performance.

[0048] Step A4 (CFA): Input the feature map output by Step A3 into the CFA attention mechanism module to improve the model's focusing ability on important disease features, suppress the interference of unimportant features, and enable the model to adapt to the natural environment with various complex background interferences.

[0049] Step A5 (Average Pooling Layer): Finally, input the feature map output by Step A4 into the average pooling layer to convert the individual probabilities of each prediction result into probabilities in the overall prediction, so as to obtain the final tea disease classification result. Through the above Steps A1 to A5, a lightweight model TDN for tea disease recognition is constructed. The lightweight model TDN for tea disease recognition is used to iteratively train the lightweight tea disease network model with the training set described in Step 3 and save the optimal weights for tea disease classification and recognition.

[0050] In Step 4, the performance evaluation metrics include accuracy, recall, number of parameters, and floating-point operation count. Among them, the higher the accuracy and recall values, the better the model's performance. The number of parameters and floating-point operation count are used to measure the complexity of the model, and the smaller the value, the more lightweight the model and the faster its running speed on edge computing devices.

[0051] During the model training process, the training set and validation set are divided into multiple batches (batch) using the batch training method, where the batch size is set to 72 and the number of iteration rounds is 100. The AdaBelief optimization algorithm is used to optimize the model, with the learning rate set to 0.001, the exponential decay rates of the first moment and second moment being 0.9 and 0.999 respectively, and the calculated adaptive learning rate adjustment value being 1e-16. For 7 types of diseased tea leaves and healthy tea leaves, the accuracy of the tea disease recognition model reaches 98.03%, and the number of model parameters is only 2.914M.

[0052] Table 1 compares the experimental results of the lightweight tea disease recognition model TDN and 6 existing lightweight convolutional neural network models (ShuffleNet, MobileNet V1, MobileNet V2, MobileNet V3, SqueezeNet, and EfficientNetV2-S), 3 classic convolutional neural network models (AlexNet, VGG16, and ResNet101), and 4 Transformer models (ViTB16, DeiT, Swin Transformer, and UniFormer) on the self-built tea disease dataset B.

[0053]

[0054]

[0055] Table 1

[0056] As can be seen from Table 1, firstly, compared with the three classic CNN models (AlexNet, VGG16, and ResNet101), the recognition accuracy of TDN has increased by 7.12%, 4.87%, and 6.18% respectively. Secondly, compared with the six lightweight CNN models (ShuffleNet V2, MobileNet V1, MobileNet V2, MobileNet V3-Large, EfficientNet V2-S, and SqueezeNet), the recognition accuracy of TDN has increased by 8.23%, 7.38%, 5.40%, 3.82%, 4.89%, and 21.30% respectively. Finally, compared with the four popular Transformer models (ViTB16, DeiT, Swin Transformer, and UniFormer), the recognition accuracy of TDN has increased by 6.63%, 3.86%, 4.03%, and 3.75% respectively. In summary, the recognition accuracy of TDN has reached 98.03%, which is better than the above benchmark models.

[0057] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0058] The above is only the preferred embodiment of the present invention and does not impose any form of limitation on the present invention; any ordinary technician in the industry can smoothly implement the present invention as shown in the accompanying drawings of the specification and as described above; however, any minor changes, modifications, and equivalent variations made by those skilled in the art within the scope of the technical solution of the present invention by using the technical content disclosed above are all equivalent embodiments of the present invention; at the same time, any equivalent changes, modifications, and variations made to the above embodiments based on the essential technology of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A lightweight tea disease recognition method based on improved EfficientNet, characterized in that: The following steps are involved: Step 1: Use a digital camera to collect image data of tea diseases at different time periods every day; Step 2: Enhance the image data in the original data set A to generate an enhanced data set B; Step 3: Split the enhanced dataset B into a ratio of 8:2, where 20% of the data is used as a test set and 80% of the data is used as a training set; Step 4: Based on the improved EfficientNet-Lite0, a lightweight model named TDN for tea disease identification is constructed. The lightweight model TDN for tea disease identification includes a Stem feature reuse module layer, a FEMBCM feature enhanced mobile inversion bottleneck convolution layer, a Head feature reuse module layer, and an average pooling layer; Step 5: Using the data set B obtained in step 2, multiple trainings are performed to obtain the optimal model weights during the training process, and then tea diseases are identified based on the obtained optimal model weights.

2. A lightweight tea disease identification method based on improved EfficientNet according to claim 1, characterized in that: The tea diseases in step 1 are divided into 8 categories: algal spot disease, tea white spot disease, tea black spot disease, tea rust disease, tea green leaf moth, red spider tea, tea dodder and healthy tea leaf images; When collecting these disease images, different time periods are selected to obtain disease images under different lighting conditions.

3. A lightweight tea disease identification method based on improved EfficientNet according to claim 2, characterized in that: In step 2, the image in the original data set A is enhanced, and the enhancement operation includes image flipping, changing the image hue and saturation, and graying the image. Image flipping refers to vertically flipping the collected image to eliminate the model's dependence on a specific direction; changing the image hue and saturation adjusts the image hue and saturation to simulate images under different lighting and environmental conditions; and graying the image is to convert a color image into a grayscale image to reduce the complexity of the model and improve the model's ability to recognize the structural characteristics of the disease.

4. The lightweight tea disease identification method based on improved EfficientNet according to claim 3 is characterized in that: The Stem feature reuse module layer described in step 4 is composed of an FRM feature reuse module and a CFA channel focused attention mechanism in a serial structure.

5. The lightweight tea disease identification method based on improved EfficientNet according to claim 4 is characterized in that: In the Stem feature reuse module layer, a feature reuse module with a convolution kernel size of 3 and a stride of 2 is first used to extract the overall feature information of the input image.

6. The lightweight tea disease identification method based on improved EfficientNet according to claim 1, characterized in that: The FEMBCM feature enhanced moving inverted bottleneck convolution layer described in step 4 is composed of 16 FEMBCM feature enhanced moving inverted bottleneck convolution modules.

7. The lightweight tea disease identification method based on improved EfficientNet according to claim 1, characterized in that: The Head feature reuse module layer and the Stem feature reuse module layer described in step 4 have the same structure. The difference between the Head feature reuse module layer and the Stem feature reuse module layer is that the Head feature reuse module layer combines the feature information extracted by the first two layers into richer feature information by using the FRM feature reuse module with a convolution kernel of 7. At the same time, the Head feature reuse module layer uses convolution with a convolution kernel of 7, which can use relatively fewer parameters to represent the receptive field of the same size, which helps to reduce the number of parameters of the model and make the model more lightweight.

8. The lightweight tea disease identification method based on improved EfficientNet according to claim 1, characterized in that: The CFA channel attention mechanism described in step 4 consists of two parallel modules: the KFEM core feature extraction module and the FFM feature focusing module.

Images