Corn disease severity detection method based on confidence learning and fine-grained feature extraction

Abstract

The present invention relates to a method for detecting the degree of corn disease based on confidence learning and fine-grained feature extraction, including: constructing corn disease samples; establishing a training sample set; constructing a corn disease degree detection model; inputting the training sample set into the corn disease degree detection model, iteratively training to obtain the best corn disease degree detection model; corn disease degree detection: inputting the sample to be detected into the best corn disease degree detection model, and identifying the type and degree of corn disease in the sample to be detected. The present invention improves the robustness of the disease degree recognition model to noise data through a noise learning framework, and can have good recognition accuracy for various complex samples to be tested in actual scenarios; in view of the problem of high similarity between different disease degrees of corn, a transform coding module of enhanced residual is used to enrich the extracted features, and at the same time, a feature selection module is added to solve the problem that fine-grained features are difficult to identify.

Claims

1. A corn disease severity detection method based on confidence learning and fine-grained feature extraction, Features: The method comprises the following steps in order: (1) Constructing corn disease samples: Collecting pictures of various disease degrees in actual corn fields for preprocessing, marking different disease categories and disease degrees, and obtaining corn disease samples; (2) Establishing a training sample set: Select corn disease samples and divide them into a training set and a test set. The ratio of corn disease samples in the training set to the test set is 7:

3. Perform data enhancement on the training set to obtain a training sample set. (3) Construct a model for detecting the severity of corn diseases; (4) sending the training sample set to the corn disease severity detection model, and iteratively training to obtain the best corn disease severity detection model; (5) Corn disease severity detection: The sample to be tested is input into the optimal corn disease severity detection model to identify the type and severity of corn diseases in the sample to be tested; Step (3) specifically includes the following steps: (3a) A noise learning framework is constructed based on confidence learning. The noise learning framework consists of three modules: confidence learning module, main learning module and consistency learning module. There are noise labels in the initial corn disease samples. After the confidence learning module, the noise samples in the training sample set are filtered using confidence learning to obtain noise samples and clean samples. Then, the consistency regularization method and the cross entropy loss function L are used in the main learning module and the consistency learning module. ce Complete information learning from noise samples; (3b) Based on the sample noise learning framework, a fine-grained feature extraction model is constructed: the original visual transform coding module is improved by adding an enhanced residual transform coding module, and the multi-layer enhanced residual transform coding module extracts a high-dimensional feature map; then the feature selection module is used to select feature vectors that effectively distinguish highly similar disease degrees from the high-dimensional feature map; finally, the selected feature vectors are concatenated and sent to the fully connected layer to output the final image category, and the noise learning framework is combined with the fine-grained feature extraction model to obtain the final corn disease degree detection model.

2. The corn disease degree detection method based on confidence learning and fine-grained feature extraction according to claim 1, Features: Step (1) specifically includes the following steps: (1a) Corn disease image sampling: collecting images of common corn diseases in actual corn fields, including corn gray spot, corn rust and corn leaf spot; (1b) Classification of corn disease severity: For the collected corn disease images, the disease severity of different corn disease types is classified and labeled based on the ratio of the diseased area in the corn leaf. The ratio of the number of pixels in the diseased area to the number of pixels in the entire corn leaf image is defined as K1. That is, the ratio K1 is used to classify the corn disease level. The calculation formula of K1 is as follows: Among them, A and A 1 respectively represent the total area of the maize disease leaf and the area of the disease region, N 1 and N are the number of pixels in the disease region and the number of pixels in the entire maize disease leaf image, respectively; when K1 = 0, the maize disease image is labeled as a healthy maize; when 0 < K1 < 0.35, the disease degree of the maize disease image is labeled as general; when 0.35 ≤ K1 ≤ 1, the disease degree of the maize disease image is calibrated as severe.

3. The corn disease degree detection method based on confidence learning and fine-grained feature extraction according to claim 1, Features: In step (3a), the confidence learning module regards the corn disease samples in the training set with lower confidence as noise samples, and regards the corn disease samples in the training set with higher confidence as clean samples, removes the labels of the noise samples and converts them into unlabeled data, and sends them to the consistency learning module for processing. In the consistency learning module, the unlabeled data is subjected to two different data augmentations, namely, image angle rotation and image color transformation. Based on the consistency regularization method, the KL divergence calculation formula is used to predict the similarity of the two augmented data; the main learning module extracts the features of the clean samples and calculates the loss value L(θ), and at the same time refers to the loss value L calculated in the consistency learning module. u (θ), to jointly update the parameters of the corn disease severity detection model; A mini-batch of images in the training set is processed by the confidence learning module to obtain a clean sample, denoted as (X L ,Y L ), where X L =x 1 ,x 2 ,…,x k , represents k clean samples, Y L =y 1 ,y 2 ,…,y k , represents the corresponding true label, then, X L Input it into the neural network F in the main learning module to obtain the corresponding prediction result F(X L ; θ), and finally the following two parts are used to represent the total loss value: The first part uses the cross entropy loss function L ce , which is expressed as follows: Among them, f i represents a mapping from the input to the softmax layer of c categories; y ij represents the sample label; θ represents the training parameter, and c represents the category; The second part is the loss function Lu in the consistency learning phase: By combining the losses of the two parts, we use α to represent the loss function L u The weights occupied by the final total loss value in the main learning module are expressed as follows: L(θ)=(1-α)L ce (θ)+αL u (i) The main learning module uses the loss value L(θ) to perform back propagation to update the parameters of corn disease degree detection; The unlabeled data selected by the confidence learning module is first subjected to two different data augmentations, image angle rotation and image color transformation, in the consistency learning module, and the distribution of these two augmentations is predicted, denoted by p W ,p S They are the distribution prediction of image angle rotation enhanced data and the distribution prediction of image color transformation strong enhanced data. The KL divergence calculation formula is as follows: Among them, A w (·) and A S (·) represents different data augmentation strategies, n represents the number of samples of unlabeled data, and u i represents unlabeled data, and τ represents the setting probability threshold parameter.

4. The corn disease degree detection method based on confidence learning and fine-grained feature extraction according to claim 1, Features: In step (3b), the fine-grained feature extraction model includes a transform coding module for enhancing residuals and a feature selection module; The calculation expression of the MSA part of the transform coding module of the enhanced residual is: Among them, AugMSA (e l ) is the transform coding module of the enhanced residual of the lth layer, MSA(e l ) is the MSA part in the transform coding module of the enhanced residual of the lth layer, A li (·) represents the i-th enhanced residual connection of the l-th layer transform coding module, θ li represents its calculation parameters; T is the number of enhanced residual connections, l is the lth layer of the transform coding module of the enhanced residual; L represents the number of layers of the transform coding module of the enhanced residual; e l represents the input of the transform coding module of the enhanced residual of layer l; For the transform coding module of the L-th layer enhanced residual, the output of the L-1th layer is expressed as The attention weights calculated for the previous layers are written as: Among them, atten i is the attention weight of the i-th layer, is the attention weight in the j-th multi-head attention in the i-th layer, N represents the number of image blocks, K represents the number of multi-head attention, and j represents the j-th multi-head attention; Recursively apply matrix multiplication to the original attention weights in all layers to get the final attention weights atten final , the calculation process is as follows: Select the index corresponding to the maximum value from the K attention heads representing different feature spaces [D 1 ,D 2 ,…,D K ], these indices will guide the model from e L-1 The corresponding tags are extracted from the , and finally the selected tags and the original global tags are concatenated as the final input sequence e final , expressed as: Wherein, D represents the index.

Images