Crop disease recognition method based on multi-branch feature cascade
By using multi-branch grouping convolutional residual network and multi-layer feature normalized cascade network in crop disease identification, the problems of high network parameters and low recognition efficiency in the prior art are solved, and more efficient disease recognition is achieved.
Patent Information
- Application Number
- CN202311062446.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-22
- Publication Date
- 2025-04-25
- Estimated Expiration
- 2043-08-22
AI Technical Summary
The existing crop disease identification methods have shortcomings in reducing network parameters, improving model identification efficiency and utilizing feature information at different levels of samples, resulting in low recognition efficiency.
A crop disease recognition method based on multi-branch feature cascade is adopted. By constructing a multi-branch packet convolutional residual network and a multi-layer feature normalization cascade network, richer features are obtained and network parameters are reduced, and the recognition efficiency is improved.
It realizes the reduction of network parameters while deepening the depth of the network, alleviating the problem of gradient disappearance, digging up semantic information at different levels, and improving the efficiency and accuracy of crop disease recognition.
Smart Images
Figure CN117036822B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of crop disease identification, and in particular to a crop disease identification method based on multi-branch feature cascade. Background Art
[0002] In the process of promoting agricultural development, the production mode in the agricultural field has changed from manual to intelligent. In agricultural planting, disease is one of the key factors affecting the yield and quality of crops. At present, the diagnosis of crop diseases still mostly relies on the rich knowledge and experience of experts. This method is difficult to meet the needs of large-scale and rapid detection of plant diseases in scientific research and actual production in terms of recognition efficiency and recognition accuracy.
[0003] In recent years, deep learning has achieved great success in the field of computer vision, and smart agriculture using deep learning methods to identify crop diseases has shown broad application prospects. However, diseased crops account for a small proportion of total crops, and fewer effective labels can be collected. In reality, diseases are diverse, and the use of deep learning methods to generalize disease features based on limited disease data sets is prone to overfitting problems.
[0004] Argüeso et al. proposed a twin neural network with Triplet loss. They first fine-tuned the Inception V3 network to extract leaf image features, then trained a one-to-many class SVM classifier to learn the mapping of different plant leaves. The recognition accuracy of six diseases on the PlantVillage dataset exceeded 90%. However, this method requires adjusting the distance between different categories, and the use of Triplet loss will increase the number of network parameters. The calculation speed of Triplet loss is also slower than the cross entropy loss, resulting in low efficiency of model disease recognition. Chen et al. proposed a local feature matching conditional neural adaptive process based on meta-learning, which can adaptively adjust the difference between two categories. Experiments were conducted on the apple leaf disease dataset and the PlantVillage dataset, with an accuracy of 93.9%. However, this technology has the limitation of low recognition efficiency in the face of large-scale datasets. Pan et al. used Faster R-CNN and small sample learning to automatically estimate the severity of strawberry leaf scorch. Although this method increased the number of samples and had a relatively high accuracy rate for diseases, it did not fully utilize the feature information of different levels of samples in the feature extraction stage, and the detection speed was slow and the equipment requirements were high.
[0005] In summary, the above technologies still have a lot of room for improvement in reducing the number of network parameters, improving model recognition efficiency, and utilizing feature information at different levels of samples. Therefore, a new method for crop disease recognition is needed. Summary of the invention
[0006] In view of this, the purpose of the present invention is to overcome the defects in the prior art and provide a crop disease identification method based on multi-branch feature cascade, which can obtain richer features and at the same time reduce the amount of network parameters, thereby improving the efficiency of crop disease identification.
[0007] The crop disease identification method based on multi-branch feature cascade of the present invention comprises the following steps:
[0008] S1. Collecting crop image information; wherein the image information includes disease image information and non-disease image information;
[0009] S2. Input the image information into the crop disease image recognition model, train the recognition model, and obtain a trained recognition model;
[0010] S3. Input the crop image to be tested into the trained recognition model and output the crop disease recognition result.
[0011] Furthermore, the crop disease image recognition model includes a feature extraction network and a classification network;
[0012] The feature extraction network includes a multi-branch grouped convolutional residual network and a multi-layer feature normalization cascade network;
[0013] The classification network is used to calculate the category to which the input image data belongs.
[0014] Furthermore, a multi-branch grouped convolutional residual network is constructed by the following method:
[0015] Based on the ResNeXt model, a single residual block is expanded to form multiple residual blocks;
[0016] Multiple residual blocks are aggregated, and the residual blocks of each layer of the network are adjusted to grouped convolutional residual blocks suitable for shallow networks.
[0017] Furthermore, a multi-layer feature normalization cascade network is constructed by the following method:
[0018] Extracting multi-layer feature maps in the neural network structure, and placing the feature maps into a linear layer for updating network parameters; the linear layer is used to perform horizontal extraction of effective parts in each layer that can improve the feature extraction effect;
[0019] The effective part of each layer is downsampled through P×P convolution and maximum pooling to form a same-channel structure of uniform size, and the weights of each layer are continuously adjusted, and finally all layers are cascaded and output.
[0020] Further, the recognition model is trained to obtain a trained recognition model, which specifically includes:
[0021] Construct training sample data based on meta-learning:
[0022] Randomly sample R epochs from the meta-training set. Each task contains S tasks. Each task consists of a support set and a query set. Both the support set and the query set contain N categories. Each category in the support set contains K samples, and each category in the query set contains M samples.
[0023] The training sample data is input into the recognition model for training to obtain a trained recognition model.
[0024] A crop disease recognition system based on multi-branch feature cascade, comprising a collection unit, a recognition model training unit and a disease recognition unit;
[0025] The acquisition unit is used to acquire crop image information;
[0026] The recognition model training unit is used to input the image information into the crop disease image recognition model, train the recognition model, and obtain a trained recognition model;
[0027] The disease recognition unit is used to input the image of the crop to be tested into the trained recognition model and output the crop disease recognition result.
[0028] Furthermore, the crop disease image recognition model includes a feature extraction network and a classification network;
[0029] The feature extraction network includes a multi-branch grouped convolutional residual network and a multi-layer feature normalization cascade network;
[0030] The classification network is used to calculate the category to which the input image data belongs.
[0031] Furthermore, a multi-branch grouped convolutional residual network is constructed by the following method:
[0032] Based on the ResNeXt model, a single residual block is expanded to form multiple residual blocks;
[0033] Multiple residual blocks are aggregated, and the residual blocks of each layer of the network are adjusted to grouped convolutional residual blocks suitable for shallow networks.
[0034] Furthermore, a multi-layer feature normalization cascade network is constructed by the following method:
[0035] Extracting multi-layer feature maps in the neural network structure, and placing the feature maps into a linear layer for updating network parameters; the linear layer is used to perform horizontal extraction of effective parts in each layer that can improve the feature extraction effect;
[0036] The effective part of each layer is downsampled through P×P convolution and maximum pooling to form a same-channel structure of uniform size, and the weights of each layer are continuously adjusted, and finally all layers are cascaded and output.
[0037] Further, the recognition model is trained to obtain a trained recognition model, which specifically includes:
[0038] Construct training sample data based on meta-learning:
[0039] Randomly sample R epochs from the meta-training set. Each task contains S tasks. Each task consists of a support set and a query set. Both the support set and the query set contain N categories. Each category in the support set contains K samples, and each category in the query set contains M samples.
[0040] The training sample data is input into the recognition model for training to obtain a trained recognition model.
[0041] The beneficial effects of the present invention are as follows: the present invention discloses a method for identifying crop diseases based on multi-branch feature cascades, and constructs a crop disease image recognition model, which applies grouped convolution to residual blocks and uses residual blocks to alleviate the gradient vanishing problem of the network. The input of the network is split into isomorphic multi-branch embeddings, and through feature supplementation between different branches, richer features are obtained and the number of network parameters is reduced at the same time. A multi-layer feature cascade backbone network is designed to mine information at different levels in an image, cascade output more positions and detail information in shallow features and stronger semantic information in deep features, thereby improving the generalization ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0043] Figure 1 This is a schematic diagram of the crop disease image recognition model framework of the present invention;
[0044] Figure 2 It is a conceptual diagram of the prototype network of the present invention;
[0045] Figure 3 is a schematic diagram of an existing residual module of the present invention;
[0046] Figure 4 Schematic diagram of the improved multi-branch grouped convolution residual module of the present invention;
[0047] Figure 5 It is a schematic diagram of the existing feature pyramid network of the present invention;
[0048] Figure 6 This is a schematic diagram of the improved multi-layer feature normalization cascade network of the present invention. DETAILED DESCRIPTION
[0049] The present invention is further described below in conjunction with the accompanying drawings, as shown in the drawings:
[0050] The crop disease identification method based on multi-branch feature cascade of the present invention comprises the following steps:
[0051] S1. Collecting crop image information; wherein the image information includes disease image information and non-disease image information;
[0052] S2. Input the image information into the crop disease image recognition model, train the recognition model, and obtain a trained recognition model;
[0053] S3. Input the crop image to be tested into the trained recognition model and output the crop disease recognition result.
[0054] The present invention has certain performance improvements in the effectiveness of crop disease identification, and achieves the goal of deepening the network while reducing the number of network parameters as much as possible, alleviating gradient disappearance, and mining semantic information at different levels, effectively solving the problem of fewer effective crop disease samples.
[0055] In this embodiment, the crop disease image recognition model includes a feature extraction network and a classification network; the feature extraction network includes a multi-branch grouped convolutional residual network and a multi-layer feature normalization cascade network; the classification network is used to calculate the category to which the input image data belongs. The classification network can adopt an existing prototype network.
[0056] As the network depth increases, the test error and training error both increase. Research and analysis show that deeper networks will cause the gradient vanishing problem, thus hindering the convergence of the network. Therefore, in order to solve this problem, the ResNet network can be used. The ResNet network adds direct edges to the nonlinear convolutional layer, that is, directly uses convolution with a step size of 2 for downsampling, replaces the fully connected layer with global average pooling, and adds a residual module to improve the efficiency of information propagation. In the case of not deepening the network as much as possible, the classification accuracy of the recognition model is improved. However, the ResNet network model is not enough to improve the recognition efficiency of the model; although the existing technology has improved the ResNet network model to form the ResNeXt network model, the ResNeXt network model is not suitable for the recognition of small sample crop diseases.
[0057] The multi-branch grouped convolutional residual network is constructed by the following method:
[0058] Based on the ResNeXt model, a single residual block is expanded to form multiple residual blocks;
[0059] Aggregate multiple residual blocks and adjust the residual blocks of each layer of the network into grouped convolution residual blocks suitable for shallow networks. Aggregate multiple residual blocks means connecting multiple residual networks (ResNet) or other similar blocks together to build a deeper and more complex network structure. This aggregation can increase the depth and representation ability of the model to better capture the complex features and patterns in the input data.
[0060] like Figure 3 , 4 As shown, through Figure 3 The residual module in is optimized and improved to obtain the following suitable for the present invention: Figure 4 The multi-branch grouped convolution residual module shown in FIG. Wherein, C represents the number of extracted feature map channels, H represents the height of the feature map, W represents the width of the feature map, and the present invention divides the feature map into 16 groups of 3×3 convolutional splicing feature maps.
[0061] In this embodiment, identifying objects of different scales has always been a challenge in the field of computer vision. Facing the problem of multi-scale target detection, scholars often use feature pyramids to solve it. However, due to the large amount of computation of feature pyramids, the overall detection speed will be reduced. If only high-level features are used for prediction in order to increase the detection speed, although the semantic information is enriched, the resolution will be reduced, and it will be difficult to accurately save the location information of the object.
[0062] In order to balance the detection speed and object location information, after obtaining features with a high degree of semantics at the high level, we continuously upsample them and then fuse them with the shallow features in the network. In this way, the fused features have both high-level semantic information and faster detection speed, while retaining higher resolution and object location information.
[0063] The present invention improves the feature pyramid and proposes a multi-layer feature normalization cascade backbone network. The multi-layer feature normalization cascade network (backbone network) is constructed by the following method:
[0064] First, a multi-layer feature map in a neural network structure is extracted, and then the feature map is placed in a linear layer for updating network parameters; the linear layer is used to horizontally extract the effective part of each layer that can improve the feature extraction effect;
[0065] The effective part of each layer is downsampled through 1×1 convolution and maximum pooling to form a same-channel structure of uniform size, and the weights of each layer are continuously adjusted, and finally all layers are cascaded and output.
[0066] like Figure 5 , 6 As shown, through Figure 5 The existing feature pyramid in is optimized and improved to obtain the following suitable for the present invention: Figure 6The multi-layer feature normalization cascade network shown.
[0067] In this embodiment, the recognition model is trained to obtain a trained recognition model, which specifically includes:
[0068] Construct training sample data based on meta-learning:
[0069] Randomly sample 60 epochs from the meta-training set. Each task contains 200 tasks. Each task consists of a support set and a query set. Both the support set and the query set contain N categories. Each category in the support set contains K samples, and each category in the query set contains M samples.
[0070] The training sample data is input into the recognition model for training to obtain a trained recognition model.
[0071] Among them, the training process of small sample learning based on meta-learning is divided into meta-training, meta-validation, and meta-testing, and different categories of sample data are used in these three stages.
[0072] Meta-learning uses Task as the basic unit. The data set in a Task includes a support set and a query set. In the meta-training phase, multiple Tasks are trained to ensure that the model has good generalization capabilities when facing different data.
[0073] Meta-validation is performed after each round of meta-training to determine whether to save the parameters learned in this training. The meta-validation and meta-testing phases use the parameters learned in the meta-training phase, and the constructed tasks are also the same. After meta-training and meta-validation, the meta-test uses the model trained in the meta-training phase to classify the target class samples.
[0074] In the small sample classification task, the training set contains a large number of labeled samples, called base class samples, which are used for model training; the validation set and test set contain unlabeled samples that are difficult to obtain in reality, called novel class samples, which are used to predict labels.
[0075] In the training phase, the Epoch of training is to randomly select N categories from the training set, select K samples from each category, and the selected N×K samples are the support set, denoted as {(x1,y1),...,(x (N×K) ,y (N×K) )}, where x is the vectorized representation of the original data in D dimensions, and y is its corresponding category. This model is called N-way K-shot, and then M samples are randomly selected from the N categories to form a query set of N×M samples.
[0076] In the meta-training phase, a large number of tasks are constructed for learning. Each task includes meta-training, meta-validation, and meta-testing. Their respective datasets are denoted as D train , D val , D test and Each data set also contains its own support set and query set. Through the above learning method, the network model can show good learning ability when facing unknown new class samples, instead of only being able to distinguish specific categories.
[0077] In the prototype network, each category has a prototype (Prototype), and the prototype point C of each category k is the mean of the class in the embedding space. The closer the sample is to the prototype point in the data set, the greater the probability that its label is the same as the label corresponding to the prototype point. Therefore, the classification problem of the sample becomes the nearest neighbor problem of the target sample in the embedding space. Figure 2 shown.
[0078] C1, C2, and C3 are the prototype points of the three categories respectively, and the target samples are embedded through an embedding function f φ : The D-dimensional samples are mapped to the M-dimensional space, and then the distances with the three center points are calculated to obtain the category of x.
[0079] A task D for meta-training train The support set D train_support For example, the prototype point C k The calculation formula is shown in formula 1:
[0080]
[0081] In the formula D train _ support The sample set of the category number k to be marked, x i Denotes the support set D train_support Each sample in .
[0082] For the query set D train_query The data in the same way are mapped, given the distance function d, the similarity between the support set and the query set data is calculated, and the softmax is used to output the category distribution of the sample x to be detected, as shown in Formula 2:
[0083]
[0084] Finally, the cross entropy loss is used to calculate the difference between the predicted value and the true value, and then the model parameters are updated by back propagation, as shown in Formula 3:
[0085]
[0086] Where i is the number of samples in the query set, N is the number of categories in the query set, and y ik The value of is 0 or 1, indicating whether the label of the i-th sample is k. If it is true, it is 0, otherwise it is 1, and p φ (y i =k|x i ) represents the probability value of the i-th sample label being k.
[0087] like Figure 1 As shown in Figure 1, the improved backbone network can obtain feature information of image samples at different levels and calculate the category to which the sample belongs through the prototype network.
[0088] The feature extraction stage of the present invention uses a backbone network of multi-branch grouped convolution residual modules and multi-layer feature normalization cascade, that is, f φ The robustness of the recognition model is improved, and then the prototype network classifier is used for classification. The model loss is calculated using the cross entropy loss, and then back-propagated to the encoder part. After the training is completed, the recognition model is transferred and tested using the test set to finally obtain the classification accuracy. Among them, knowledge transfer can be carried out at multiple levels, including feature transfer, model transfer and task transfer, which will not be repeated here.
[0089] The present invention also relates to a crop disease identification system based on multi-branch feature cascade, which corresponds to the above-mentioned crop disease identification method based on multi-branch feature cascade, and can be understood as a system for implementing the above-mentioned method, and the system includes a collection unit, a recognition model training unit and a disease identification unit;
[0090] The acquisition unit is used to acquire crop image information;
[0091] The recognition model training unit is used to input the image information into the crop disease image recognition model, train the recognition model, and obtain a trained recognition model;
[0092] The disease recognition unit is used to input the image of the crop to be tested into the trained recognition model and output the crop disease recognition result.
[0093] Wherein, the crop disease image recognition model includes a feature extraction network and a classification network;
[0094] The feature extraction network includes a multi-branch grouped convolutional residual network and a multi-layer feature normalization cascade network;
[0095] The classification network is used to calculate the category to which the input image data belongs.
[0096] Among them, the multi-branch grouped convolutional residual network is constructed by the following method:
[0097] Based on the ResNeXt model, a single residual block is expanded to form multiple residual blocks;
[0098] Multiple residual blocks are aggregated, and the residual blocks of each layer of the network are adjusted to grouped convolutional residual blocks suitable for shallow networks.
[0099] A multi-layer feature normalization cascade network is constructed by the following method:
[0100] Extracting multi-layer feature maps in the neural network structure, and placing the feature maps into a linear layer for updating network parameters; the linear layer is used to perform horizontal extraction of effective parts in each layer that can improve the feature extraction effect;
[0101] The effective part of each layer is downsampled through P×P convolution and maximum pooling to form a same-channel structure of uniform size, and the weights of each layer are continuously adjusted, and finally all layers are cascaded and output.
[0102] The present invention provides an objective and scientific crop disease identification method and system. Aiming at the problems of complex models and high training overhead, a single residual block is expanded on the basis of ResNeXt to form a multi-branch structure, and a group convolution strategy is adopted for it, so that the model can be trained on multiple GPUs in a parallel manner, allowing the network to process more images at each step, thereby achieving the purpose of reducing the number of network parameters and improving model efficiency. Aiming at the problem that sample feature information is not fully utilized, a multi-layer feature normalization cascade backbone network is proposed. Based on the feature pyramid, in the face of the multi-scale change problem of objects in the feature extraction process, each layer of feature map information is not fully utilized, and only the part of each layer that can improve the feature extraction effect is extracted and horizontal extraction is performed. With a small increase in the amount of calculation, feature maps with strong low-resolution semantic information and feature maps with weak high-resolution semantic information but rich spatial information are fused.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A method for identifying crop diseases based on multi-branch feature cascade, characterized in that: The steps include: S1. Collecting crop image information; wherein the image information includes disease image information and non-disease image information; S2. Input the image information into the crop disease image recognition model, construct training sample data based on meta-learning to train the recognition model, and obtain a trained recognition model; The crop disease image recognition model includes a feature extraction network and a classification network; The feature extraction network includes a multi-branch grouped convolutional residual network and a multi-layer feature normalization cascade network; The classification network uses a prototype network classifier to perform classification and calculate the category to which the input image data belongs; The multi-branch grouped convolutional residual network is constructed by the following method: Based on the ResNeXt model, a single residual block is expanded to form multiple residual blocks; Aggregating multiple residual blocks and adjusting the residual blocks of each layer of the network to grouped convolution residual blocks suitable for shallow networks; wherein aggregating multiple residual blocks refers to connecting multiple residual networks or other similar blocks together; A multi-layer feature normalization cascade network is constructed by the following method: Extracting multi-layer feature maps in the neural network structure, and placing the feature maps into a linear layer for updating network parameters; the linear layer is used to perform horizontal extraction of effective parts in each layer that can improve the feature extraction effect; The effective part of each layer is downsampled through P×P convolution and maximum pooling to form a same-channel structure of uniform size, and the weights of each layer are continuously adjusted, and finally all layers are cascaded and output; S3. Input the crop image to be tested into the trained recognition model and output the crop disease recognition result.
2. The crop disease identification method based on multi-branch feature cascade according to claim 1, characterized in that: The recognition model is trained to obtain a trained recognition model, which specifically includes: Construct training sample data based on meta-learning: Randomly sample R epochs from the meta-training set. Each task contains S tasks. Each task consists of a support set and a query set. Both the support set and the query set contain N categories. Each category in the support set contains K samples, and each category in the query set contains M samples. The training sample data is input into the recognition model for training to obtain a trained recognition model.