CN114821182A - Rice growth stage image recognition method - Google Patents

Abstract

The invention discloses an image recognition method for a rice growth stage, which comprises the following steps: step 1, collecting rice pictures on the spot and on the network as a data set; step 2, dividing the data set into a training set and a testing set, and performing preprocessing and data enhancement; step 3, constructing an ODRL-Swin transducer model; and step 4, setting configuration parameters of the ODRL-Swin transducer model as the optimal configuration parameters through training, and step 5, outputting a final rice prediction identification result by the ODRL-Swin transducer model. The method can improve the accuracy of rice image recognition.

Description

A method for image recognition of rice growth stages

technical field

The invention relates to the field of crop image recognition methods, in particular to an image recognition method for rice growth stages.

Background technique

In order to better cultivate rice, it is necessary to understand the various growth stages of rice, and take corresponding planting measures at different growth stages. In current smart agriculture, images are generally collected at various growth stages of rice, and the collected images are identified to know what growth stage the rice is currently in, and to take corresponding planting measures based on the identification results. Therefore, how to know the current growth state of rice by identifying the images of rice growth stages is an important factor affecting the intelligent management and planting of rice.

With the continuous development of computer technology, more and more technologies can be combined with other fields. In agriculture, accurate, rapid and convenient identification technology can reduce the cost of agricultural workers and also have a positive impact on the yield of crops. Most of the existing technologies use convolutional networks as models for detection and identification of agricultural crops. Convolutional networks are easy to use, but the detection of models based on convolutional networks has the disadvantage of low accuracy. For the Transformer-based model, its performance is higher than that of traditional convolutional networks, but it requires a large number of datasets as support to train the model. In agriculture, there are often very few datasets, and most of them need to be collected and labeled manually. The production of a large number of datasets requires a lot of manpower and material resources, which is characterized by high cost. Therefore, there is currently a lack of a recognition model that can achieve high accuracy on small-scale datasets.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image recognition method of rice growth stage, so as to solve the problems of low accuracy of rice identification in the prior art and the need for a large number of data sets for training.

In order to achieve the above object, the technical scheme adopted in the present invention is:

An image recognition method for rice growth stage, comprising the following steps:

Step 1. Obtain multiple image data of rice in different growth stages as a data set;

Step 2: Divide the data set obtained in Step 1 into a training set and a test set, and preprocess the data in the training set and the test set respectively, and then perform data enhancement respectively;

Step 3. Based on the Swin Transformer model, the Swin Transformer module consists of block division, linear embedding, and several Swin Transformer Blocks. Optimized dense relative localization is added to the SwinTransformer Block in the Swin Transformer model, and the loss is The Dense relative localization loss is added to the function to construct the ODRL-Swin Transformer model;

In the original Swin Transformer module, the feature map is divided into non-overlapping windows according to the size of the window. The elements in the window are called blocks, and the dense relative positioning is optimized by learning the relative position between the blocks. can integrate more spatial information; in the ODRL-SwinTransformer model, the optimized dense relative positioning is performed by densely sampling the blocks in the original Swin Transformer window, and randomly selecting two blocks in the window during sampling, here we call the selected block The two blocks are embedding pairs, and then the geometric relative position distance of the embedding pair is calculated and the relative position distance of the embedding pair is predicted using MLP to collect spatial information, and the spatial relative position loss is added to the loss function of the Swin Transformer. further guides the computation of the relative positions of the embedding pairs;

Step 4. Use the training set of Step 2 to train the ODRL-Swin Transformer model constructed in Step 3, and adjust the parameters of the ODRL-Swin Transformer model in combination with the training results and the test set of Step 2 until the parameters of the ODRL-Swin Transformer model are the best. optimal configuration parameters;

Step 5: Input the image data of the rice growth stage to be identified into the ODRL-Swin Transformer model under the optimal configuration parameters obtained in step 4, and the ODRL-Swin Transformer model outputs the image prediction and recognition result of the rice growth stage.

Further, the preprocessing in step 2 includes preprocessing including removing duplicate pictures and deleting damaged pictures on multiple image data obtained at different growth stages of rice, deleting unmatched information in the annotation file, and sorting the data set according to the following steps. The ratio of 7:2:1 divides the dataset into training, test, and validation sets.

Further, the data enhancement in step 3 includes Mosaic data augmentation, random flipping, scaling, and random cropping.

Further, in step 4, the detection accuracy of Swin Transformer is greatly improved compared with the traditional convolutional network, but the detection accuracy of Swin Transformer is not high on small-scale data sets, and optimized dense relative localization and dense relative position are optimized. The loss (Dense relative localizationloss) enables the Swin Transformer to perform well on small-scale datasets as well.

Further, during training in step 5, input the training set data into the ODRL-Swin Transformer model to obtain the output results, and perform error calculation between the output results and the test set, and then adjust the configuration of the ODRL-Swin Transformer model based on the error calculation results. parameters until the error calculation results meet expectations, at which time the configuration parameters of the ODRL-SwinTransformer model are the optimal configuration parameters.

The invention detects the rice pictures to be recognized based on the ODRL-Swin Transformer model, so as to realize the detection of different stages of rice. The ODRL-Swin Transformer model used is added at the Swin Transformer Block in the Swin Transformer, and the ODRL-Swin Transformer model is added by the ODRL-Swin Transformer model. Output the final rice growth stage prediction and recognition result.

The invention can ensure the accuracy of recognizing rice images, and realize the accurate recognition of rice through small-scale data sets. The invention can efficiently and accurately identify which growth stage the rice is in, so that farmers can provide the most reasonable planting measures for the rice. Increase rice yield.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention.

FIG. 2 is a structural diagram of the ODRL-Swin Transformer model of the present invention.

FIG. 3 is a structural diagram of the Optimized Dense Relative Localization part of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, an image recognition method of a rice growth stage of the present invention includes the following steps:

(1) Prepare the dataset:

Collect the picture data of different growth stages of rice in the field and online to form a data set;

(2) Processing datasets:

First, the specific labels and numbers of images of various rice growth stages are obtained from the data set obtained in step 1, and duplicate and abnormal data are eliminated. The data in the test set and the test set are preprocessed. The preprocessing includes acquiring multiple image data of rice in different growth stages to remove duplicate images and delete damaged images, delete the mismatched information in the annotation files, and press the data set to 7 The :2:1 ratio divides the dataset into training, test, and validation sets.

(3) Data enhancement:

Data augmentation is performed on the data in the training set and the test set respectively. Data augmentation includes Mosaic data augmentation, random flip (RandomFlip), scaling (Resize), and random cropping (RandomCrop). After the data is enhanced, pad the data to avoid feature loss and retain the features of the rice dataset.

(4) Build the ODRL-Swin Transformer model based on Swin Transformer:

The optimized dense relative localization is added to the SwinTransformer Block in the Swin Transformer model, and the Dense relative localization loss is added to the loss function to construct the ODRL-Swin Transformer model.

The feature map obtained after block division and linear embedding is divided into non-overlapping windows according to the size of the window in the original Swin Transformer Block. The elements in the window are called blocks, and the optimized dense relative positioning makes the relative position between the blocks to be learned. Swin Transformer can fuse more spatial information without additional annotation information. In our ODRL-Swin Transformer model, the dense relative positioning is optimized by densely sampling the blocks in the original SwinTransformer window, and randomly selecting two blocks in the window during sampling, where the selected two blocks are called embedding pairs, and then There are two ways to collect spatial information by calculating the geometric relative position distance of the embedding pair and using MLP to predict the relative position distance of the embedding pair. The implementation methods are as follows:

Given an image x, the Swin Transformer Block in the Swin Transformer model divides the feature map obtained by block division and linear embedding with a window size of 7*7, and divides the input feature map into H*W windows of the same size , where H is the number of rows of the divided window, and W is the number of columns of the divided window. One of the windows can be expressed as G _x ={e _i,j } _{1≤i≤H,1≤j≤W} , where e _i,j ∈R ^D , e _i,j represents the embedding block, and D is the partitioned window space dimension. i represents the embedded block in the i-th row, and j represents the j-th column of the embedded block.

For each G _x , pairs of embeddings are randomly sampled in the optimized dense relative localization module, and for each sampled pair ( _ei,j ,e _p,j ), a 2D normalized translation offset (t _u , t _v ) ^T , calculated as follows:

(t_u,t_v)^T∈[0,1]².

(t _u ,t _v ) ^T ∈[0,1] ² .

in

sign() is a sign function. The function returns an integer variable indicating the sign of the parameter. If the return value is greater than 0, sign returns 1; if it is equal to 0, it returns 0; if it is less than 0, it returns -1. The sign of the number parameter determines the return value of the sign function. Immediately return 1 for positive input, -1 for negative input, and 0 for other input. α is the segment point of the piecewise function, and the default value is 4. p means that the other embedding block in the embedding pair is the pth row, and h means that the other embedding block in the embedding pair is the hth row.

The selected embedded vectors e _i,j and e _p,h are connected to the operation, and the vector after the connection operation is input into the multi-layer perceptron MLP that optimizes the dense relative positioning module. The multi-layer perceptron MLP has two The hidden layer and two output neurons are used to predict the relative distance between position (i, j) and position (p, h) on the grid, d _u represents the distance on the predicted abscissa, and d _v represents the predicted ordinate on the distance. Its structure is shown in Figure 3, and the calculation formula is as follows:

(d _u ,d _v ) ^T =f(e _i,j ,e _p,j ) ^T

B represents the number of pictures processed simultaneously when the model is calculated in parallel, and the dense relative localization loss proposed by the present invention is:

被添加到Swin Transformer的标准交叉熵损失

中。最后总的损失为：

在本发明中初始时使用λ＝0.5。正则化损失的引入能让SwinTransformer在不使用额外的人工标注的情况下学习空间信息。实现不依靠大量的数据集就能有效学习图像内的空间关系。

Standard cross-entropy loss added to Swin Transformer

middle. The final total loss is:

In the present invention, λ=0.5 is used initially. The introduction of regularization loss enables SwinTransformer to learn spatial information without using additional human annotations. This enables efficient learning of spatial relationships within images without relying on large datasets.

In the present invention, a relative position module optimized by OptimizedDense Relative Localization is added to the Swin Transformer Block in the Swin Transformer model, and the Swin Transformer Block is modified into an ODRL-Swin Transformer Block. The structure diagram of the modified ODRL-Swin Transformer network is shown in Figure 2. The Swin Transformer can be made to learn spatial information without using additional human annotations by densely sampling multiple embedding pairs for each image and asking the network to predict their relative positions.

(5) Train the ODRL-Swin Transformer model and set the optimal configuration parameters of the model:

The output result after training the ODRL-Swin Transformer model is the prediction and recognition result of the growth stage of rice, and the classification error and regression error are calculated between the model output and the test set. According to the verification and test results, the trained ODRL-Swin Transformer model is used. The configuration parameters of the ODRL-Swin Transformer model are adjusted to the optimal configuration parameters, and the data set of the images of the rice growth stage to be recognized is input into the ODRL-Swin Transformer model whose parameters are adjusted to the optimal configuration parameters, and the ODRL-Swin Transformer model outputs the final rice prediction and recognition results.

By recognizing the images of different stages of rice, these rice image data are used as the data set to be recognized and input into the final ODRL-Swin Transformer model, and the ODRL-Swin Transformer model outputs the detection results of rice, and the input image belongs to Which growth stage of rice can be accurately identified and detected.

The embodiments of the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Various modifications and improvements made should fall within the protection scope of the present invention, and the technical content claimed in the present invention has been fully recorded in the claims.

Claims (5)

1. A rice growth stage image recognition method is characterized by comprising the following steps:

step 1, acquiring a plurality of image data of rice at different growth stages as a data set;

step 2, dividing the data set obtained in the step 1 into a training set and a testing set, respectively preprocessing the data in the training set and the testing set, and then respectively enhancing the data;

step 3, based on the Swin Transformer model, the Swin Transformer module is composed of Block division, linear embedding and a plurality of Swin Transformer blocks; adding optimized dense relative positioning on a SwinTransformer Block in a SwinTransformer model, and adding dense relative position loss in a loss function, thereby constructing an ODRL-Swin Transformer model;

an original Swin Transformer module divides a characteristic graph into non-overlapping windows according to the size of the windows, elements in the windows are called blocks, and optimization dense relative positioning enables the Swin Transformer to fuse more spatial information without additional marking information through learning relative positions between the blocks; in an ODRL-Swin transform model, optimizing dense relative positioning by densely sampling blocks in an original Swin transform window, randomly selecting two blocks in the window during sampling, wherein the two blocks are called as an embedding pair, then calculating the geometric relative position distance of the embedding pair and predicting the relative position distance of the embedding pair by using MLP (Multi-level Linear programming) to realize the collection of spatial information, and further guiding the calculation of the relative position of the embedding pair by adding spatial relative position loss on a loss function of the Swin transform;

step 4, training the ODRL-Swin transducer model constructed in the step 3 by using the training set in the step 2, and adjusting parameters of the ODRL-Swin transducer model by combining the training result and the test set in the step 2 until the parameters of the ODRL-Swin transducer model are optimal configuration parameters;

and 5, inputting the rice growth stage image data to be identified into the ODRL-Swin Transformer model under the optimal configuration parameters obtained in the step 4, and outputting the rice growth stage image prediction identification result through the ODRL-Swin Transformer model.

2. The method for identifying the images of the rice in the growing stage as claimed in claim 1, wherein the preprocessing in the step 2 includes the operations of removing the duplicate images and the damaged images of the obtained image data of the rice in different growing stages, deleting the unmatched information in the annotation file, and dividing the data set into a training set, a testing set and a verification set according to a ratio of 7:2: 1.

3. The method for recognizing the rice growth stage image as claimed in claim 1, wherein the data enhancement in the step 3 comprises Mosaic data expansion, random inversion, scaling and random cropping.

4. The rice growth stage image recognition method as claimed in claim 1, wherein in step 4, Swin Transformer detection accuracy is greatly improved compared with that of a conventional convolution network, but Swin Transformer detection accuracy is not high on a small-scale data set, and optimization of dense relative positioning and dense relative position loss enables Swin Transformer to perform better on the small-scale data set as well.

5. The method as claimed in claim 1, wherein during the training in step 5, the data of the training set is input into the ODRL-Swin Transformer model to obtain an output result, the output result and the test set are subjected to error calculation, and then the configuration parameters of the ODRL-Swin Transformer model are adjusted based on the error calculation result until the error calculation result is in accordance with the expectation, at which time the configuration parameters of the ODRL-Swin Transformer model are the optimal configuration parameters.

Images