CN113177486B - Identification method of Odonata insects based on region proposal network - Google Patents

Abstract

The invention discloses a dragonfly order insect identification method based on a regional suggested network, which comprises the following steps: s1, cleaning and sorting images of insects in the order of dragonfly to obtain a data set of the images of the insects in the order of dragonfly; s2, enhancing the data set of the dragonfly order insect image to obtain an enhanced data set; s3, dividing the enhanced data set to obtain a training set, a verification set and a test set of the dragonfly order insect images; s4, constructing a deep convolutional network model based on the regional suggestion network; s5, training a deep convolution network model by using the training set and the verification set to obtain a trained network model; and S6, inputting the test set into the trained network model, and outputting to obtain a classification result of the test set. The method can enable the identification processing to be simple and rapid, save a large amount of labor cost, and solve the problem of difficult identification caused by complex background of the dragonfly-order insect pictures in natural environment.

Description

Recognition method of Odonata insects based on region proposal network

technical field

The invention relates to the field of identification, in particular to a method for identifying Odonata insects based on a region suggestion network.

Background technique

The existing Odonata automatic identification algorithm uses the features designed by humans as the classification basis, and uses the traditional identification method to construct the identification frame, which can only identify the specimen pictures of several species of dragonflies, and the identification rate is low, which is very difficult for those with complex backgrounds in the natural environment. Dragonfly pictures do not have the ability to identify;

Existing insect automatic identification algorithms are often implemented in two steps. The first step is detection. That is, the detection algorithm is firstly implemented for the target to be identified. This process relies on a large amount of manual annotation information, which is time-consuming and labor-intensive, making the initial cost very high. The second step is to identify. There are two ways. One is to create a data set based on the detection results of the detection algorithm in the first step, so as to train the corresponding classifier for classification. The effect of this method will be limited by the performance of the detection algorithm. , perform manual box selection and segmentation on the data set, train the classifier on the premise that the training image contains the complete target to be recognized, and then use the detection algorithm to segment the image during testing, which is also time-consuming and labor-intensive, increasing the algorithm labor costs.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to overcome the defects in the prior art, and provide a method for identifying Odonata insects based on a regional suggestion network, which can make the identification process simple and fast, save a lot of labor costs, and solve the problem of Odonata in the natural environment. The complex background of insect pictures makes it difficult to identify.

The Odonata insect identification method based on the regional suggestion network of the present invention comprises the following steps:

S1. Collect images of Odonata insects, clean and organize the images of Odonata insects, and obtain a dataset of Odonata insect images;

S2. Enhance the dataset of dragonfly images to obtain an enhanced dataset;

S3. Divide the enhanced data set according to a set ratio to obtain a training set, a verification set and a test set of the Odonata images;

S4. Build a deep convolutional network model based on the region proposal network;

S5. set training parameters, and use the training set and the verification set to train the deep convolutional network model to obtain the trained network model;

S6. Input the test set into the trained network model, and output the classification result of the test set.

Further, the Odonata insect image is an Odonata insect image in a natural environment.

Further, in step S2, the enhancement processing is performed on the data set of Odonata images, which specifically includes: random horizontal flipping of the data set and random center cropping of the data set.

Further, the step S4 specifically includes:

S41. Build a regional proposal network;

S42. Use ResNet50 as the feature extraction network of the deep convolutional network model, and use the region proposal network as the feature screening network of the deep convolutional network model;

S43. Determine the loss function of the deep convolutional network model.

Further, the loss function L of the deep convolutional network model is determined according to the following formula:

L=L ₁ +μL ₂ ;

Among them, L ₁ is the cross entropy loss function; L ₂ is the improved Focal Loss; μ is the setting coefficient;

y_i为当前样本的真实结果，

为所选子区域的预测结果；said

y _i is the real result of the current sample,

is the prediction result of the selected sub-region;

为当前样本的预测概率，α与γ分别为控制参数；said

is the predicted probability of the current sample, α and γ are the control parameters;

id为样本编号，classes为样本的总标签数。said

id is the sample number, and classes is the total number of labels of the sample.

Further, in step S5, set training parameters, and use the training set and the verification set to train the deep convolutional network model, specifically including:

S51. Use the pre-trained weight model on the ImageNet dataset as the initialization weight model of ResNet50;

S52. Set the input size of the Odonata images in the training set and the validation set, and set the preset size of the sub-region;

S53. Use Batch Normalization to achieve regularization, and use stochastic gradient descent to optimize the deep convolutional network model.

The beneficial effects of the present invention are as follows: a method for identifying Odonata insects based on a region suggestion network disclosed in the present invention, by using ResNet50 as a feature extraction network, no manual feature design and extraction are required, and a convolutional neural network is designed as a region suggestion The network is used for feature screening, which enhances the model's ability to extract effective features from complex backgrounds, and solves the problem of difficult identification caused by complex backgrounds in Odonata pictures in natural environments.

Description of drawings

Below in conjunction with accompanying drawing and embodiment, the present invention is further described:

Fig. 1 is the method flow schematic diagram of the present invention;

2 is an overall network frame diagram of the method of the present invention;

FIG. 3 is a schematic diagram of an area proposal network of the present invention;

4 is a schematic diagram of a depthwise separable convolution of the present invention;

FIG. 5 is a schematic diagram of pooling and de-pooling of the present invention.

Detailed ways

The present invention is further described below in conjunction with the accompanying drawings, as shown in Figure 1:

The Odonata insect identification method based on the regional suggestion network of the present invention comprises the following steps:

S1. Collect images of Odonata insects, clean and organize the images of Odonata insects, and obtain a dataset of Odonata insect images; among which, the acquisition methods of Odonata insect image data include field shooting, laboratory shooting, and network download, etc. Therefore, the final data set of Odonata images does not have the characteristic that the number of images of various Odonata insects is equal under the given conditions, and the overall data has a long-tailed distribution.

S2. Enhance the dataset of dragonfly images to obtain an enhanced dataset;

S3. Divide the enhanced data set according to a set ratio to obtain a training set, a verification set and a test set of Odonata images; wherein, the set ratio is 4.5:4.5:1.

S4. Construct a deep convolutional network model based on the region proposal network; wherein, by constructing a deep convolutional network model based on the region proposal network, the step of firstly detecting Odonata insects is omitted, thereby avoiding a large amount of data processing. The manual annotation work also enables the deep convolutional network model to directly identify the dragonfly species in the image to be tested;

S5. set training parameters, and use the training set and the verification set to train the deep convolutional network model to obtain the trained network model;

S6. Input the test set into the trained network model, and output the classification result of the test set.

In this embodiment, the Odonata insect image is an Odonata insect image in a natural environment. Among them, individual species in the collected Odonata images contain very few laboratory specimen images.

In this embodiment, in step S2, the enhancement processing is performed on the data set of Odonata images, which specifically includes: random horizontal flipping of the data set and random center cropping of the data set.

In this embodiment, the step S4 specifically includes:

S41. Build a region proposal network; wherein, the region proposal network is abbreviated as RPN (region proposal network). As shown in Figure 3, the yellow imaginary blocks represent average pooling, the yellow borderless blocks represent maximum de-pooling, and the green borderless blocks represent Depthwise separable convolution; for an input image of size H×W×C, after passing through the RPN, a set of h×w×C× with a preset number (k) and size (h×w) is obtained k sub-regions, which are added to the original image as the input of the final classification network.

S42. Use ResNet50 as the feature extraction network of the deep convolutional network model, and use the region proposal network as the feature screening network of the deep convolutional network model; specifically, as shown in Figures 4 and 5, for the input picture F∈R ^{H× W×C} selects M sub-regions of different preset scales, through 3×3 depth convolution plus 1×1 point convolution, followed by 3×3 average pooling, and then 3×3 maximum inversion Pooling (a single 1×1 convolution from here), followed by a set of depthwise separable convolutions (a single 1×1 convolution here), followed by a 3×3 standard convolution , followed by 1×1 convolution, the sampling process uses the ReLu function as the activation function, and finally cascades with two branches to obtain the feature map F _R , and then pre-classifies. Then, according to the pre-classification results, the k regions with the highest confidence are selected from the M regions as the sub-regions screened by the RPN as the output S={R ₁ , R ₂ ,...,R _k }.

S43. Determine the loss function of the deep convolutional network model. Among them, by introducing the loss function and adding hyperparameters to control the proportion of the overall network loss, the problem of uneven number of samples of various types in the data set is solved. Specifically, for the input picture F∈R ^H×W×C and the output S={R ₁ ,R ₂ ,...,R _k }, after using ResNet50 for feature extraction, the obtained feature maps are added for classification, The classification loss is processed using the loss function.

In this embodiment, the loss function L of the deep convolutional network model is determined according to the following formula:

L=L ₁ +μL ₂ ;

Among them, L ₁ is the cross entropy loss function; L ₂ is the improved Focal Loss; μ is the setting coefficient;

y_i为当前样本的真实结果，

为所选子区域的预测结果；said

y _i is the real result of the current sample,

is the prediction result of the selected sub-region;

为当前样本的预测概率，α与γ分别为控制参数，所述参数用来调节所述数据集中易分类样本的损失计算权重，本实施例中，设置参数α为0.5，参数γ为2；said

is the prediction probability of the current sample, α and γ are control parameters, respectively, and the parameters are used to adjust the loss calculation weight of the easy-to-classify samples in the data set. In this embodiment, the parameter α is set to 0.5, and the parameter γ is set to 2;

id为样本编号，classes为样本的总标签数。在本实施例中，在制作所述数据集时，按各类所含样本数的多少降序来设置样本编号(id)，即第一类数量最多，最后一类数量最少。当当前训练样本的标签在样本的总标签数(classes)的前一半时，μ＝0.5，反之μ＝1。said

id is the sample number, and classes is the total number of labels of the sample. In this embodiment, when creating the data set, the sample numbers (id) are set in descending order of the number of samples contained in each category, that is, the first category has the largest number and the last category has the smallest number. When the label of the current training sample is in the first half of the total number of labels (classes) of the sample, μ=0.5, otherwise μ=1.

In this embodiment, in step S5, training parameters are set, and the training set and the verification set are used to train the deep convolutional network model, which specifically includes:

S51. Use the pre-trained weight model on the ImageNet dataset as the initialization weight model of ResNet50;

S52. Set the input size of the Odonata images in the training set and the validation set, and set the preset size of the sub-region;

S53. Use Batch Normalization to achieve regularization, and use stochastic gradient descent to optimize the deep convolutional network model.

Specifically, in the training process, the input size of the image (Input size) is set to 448×448 (pixel value), and the preset sizes of the sub-regions are 48×48, 96×96, 192×192 (pixel value), The value of k is 3, and BN (BatchNormalization) is used to achieve regularization. The initial learning rate is 0.001 and the optimizer is Momentum SGD. The NMS threshold is 0.25, the weight decay is 0.0001, and the iteration epoch is 200.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent substitutions without departing from the spirit and scope of the technical solutions of the present invention should be included in the scope of the claims of the present invention.

Claims (4)

1. an Odonata insect identification method based on regional suggestion network, is characterized in that: comprise the steps: S1. Collect images of Odonata insects, clean and organize the images of Odonata insects, and obtain a dataset of Odonata insect images; S2. Enhance the dataset of dragonfly images to obtain an enhanced dataset; S3. Divide the enhanced data set according to a set ratio to obtain a training set, a verification set and a test set of the Odonata images; S4. Build a deep convolutional network model based on the region proposal network; The step S4 specifically includes: S41. Build a regional proposal network; S42. Use ResNet50 as the feature extraction network of the deep convolutional network model, and use the region proposal network as the feature screening network of the deep convolutional network model; specifically including: selecting M sub-regions of different preset scales for the input picture, Depth convolution plus point convolution, followed by average pooling, followed by maximum unpooling, from here a single convolution, followed by a group of depthwise separable convolutions, here a single convolution, and then a single convolution. The standard convolution is followed by the convolution. The sampling process uses the ReLu function as the activation function. Finally, it is cascaded with two branches to obtain a feature map, and then pre-classification is performed. Then, according to the pre-classification results, the M areas with the highest confidence are selected. The k regions of are used as sub-regions of RPN screening, and output S={R ₁ , R ₂ ,...,R _k }; S43. Determine the loss function of the deep convolutional network model; The loss function L of the deep convolutional network model is determined according to the following formula: L=L ₁ +μL ₂ ; Among them, L ₁ is the cross entropy loss function; L ₂ is the improved Focal Loss; μ is the setting coefficient; said

y _i is the real result of the current sample,

is the prediction result of the selected sub-region; said

is the predicted probability of the current sample, α and γ are the control parameters; said

id is the sample number, set the sample number in descending order of the number of samples contained in each category, and classes is the total number of labels of the sample; S5. set training parameters, and use the training set and the verification set to train the deep convolutional network model to obtain the trained network model; S6. Input the test set into the trained network model, and output the classification result of the test set. 2 . The method for identifying Odonata insects based on a region proposal network according to claim 1 , wherein the Odonata insect images are images of Odonata insects in a natural environment. 3 . 3. The method for identifying Odonata insects based on a region suggestion network according to claim 1, wherein in step S2, the data set of the Odonata insect images is enhanced, specifically comprising: randomly performing a randomization process on the data set. Horizontal flipping and random center cropping of the dataset. 4. the Odonata insect identification method based on the regional suggestion network according to claim 1, is characterized in that: in step S5, set training parameter, and use described training set and verification set to train deep convolutional network model, Specifically include: S51. Use the pre-trained weight model on the ImageNet dataset as the initialization weight model of ResNet50; S52. Set the input size of the Odonata images in the training set and the validation set, and set the preset size of the sub-region; S53. Use Batch Normalization to achieve regularization, and use stochastic gradient descent algorithm to optimize the deep convolutional network model.

Images