CN109145770B - Automatic wheat spider counting method based on combination of multi-scale feature fusion network and positioning model - Google Patents

Abstract

The invention relates to a method for automatically counting wheat spiders based on combination of a multi-scale feature fusion network and a positioning model, which overcomes the defect of high error rate of image detection aiming at small targets compared with the prior art. The invention comprises the following steps: establishing a training sample; constructing a wheat spider detection counting model; acquiring an image to be counted; and obtaining the number of the wheat spiders. The invention realizes the direct identification and counting of the wheat spiders under the field natural environment.

Description

An automatic counting method of wheat spiders based on the combination of multi-scale feature fusion network and localization model

technical field

The invention relates to the technical field of image recognition, in particular to a wheat spider automatic counting method based on the combination of a multi-scale feature fusion network and a positioning model.

Background technique

Wheat is one of the main food crops in my country. In the process of wheat production, it is vulnerable to a variety of pests. Wheat spider is one of them. It sucks the juice of wheat leaves and even dries up, which seriously affects the output of wheat. The detection of pest population is an important means of pest control, which provides a theoretical basis for pest control decision-making. Therefore, identification and counting of wheat spiders in the field is crucial for improving wheat yield.

With the rapid development of computer vision technology and image processing technology, image-based pest automatic identification and counting technology has become a research hotspot in recent years. Although this method is time-saving and labor-saving, and has the advantages of intelligence, it is not suitable for the identification and counting of field wheat spiders. The reasons are: first, the individual wheat spiders are only a few millimeters in size, and it is difficult to detect such small targets using traditional image recognition technology (SVM). It affects the quality of the image; moreover, in practical applications, the collected images are often mixed with other debris, and the background is more complicated.

Therefore, how to realize the detection of small targets such as wheat spiders in a complex environment has become a technical problem that needs to be solved urgently.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the defect of high error rate of image detection for small targets in the prior art, and to provide an automatic counting method of wheat spiders based on the combination of multi-scale feature fusion network and positioning model to solve the above problems.

In order to achieve the above object, technical scheme of the present invention is as follows:

An automatic counting method for wheat spiders based on the combination of multi-scale feature fusion network and localization model, comprising the following steps:

For the establishment of training samples, more than 2000 images of wheat spiders in the natural environment in the field are obtained as training images, and the wheat spiders in the images have been marked to obtain training samples;

Construct the wheat spider detection and counting model;

Construct a positioning model;

Construct a multi-scale feature fusion network, and transform the structure of the multi-scale feature fusion network;

Train the multi-scale feature fusion network, train the features of the candidate regions located by the training samples according to the localization model, and use the output results of each layer as the prediction results;

The acquisition of the images to be counted is to acquire the images of wheat spiders photographed in the field, and preprocess to obtain the images to be counted;

To obtain the number of wheat spiders, input the image to be counted into the wheat spider detection and counting model to obtain the number of wheat spiders in the image.

The described construction positioning model includes the following steps:

Set the color space conversion module, the color space conversion module is used to convert the RGB color space to the YcbCr color space, and divide it into R={r ₁ , r ₂ ,...r _n } divided regions;

Calculate the similarity of color information, use L1 normalization to obtain 25 histograms of each color channel of the image, and calculate the similarity of the color space. The calculation formula is as follows:

表示第i个通道、第k个直方图向量，i＝1,2,3，k＝0,1....,25，n表示直方图个数，r_i表示分割区域R＝{r₁,r₂,...r_n}第i个区域；

表示第j个通道、第k个直方图向量，j＝1,2,3，m表示m个直方图；Among them, f _color (r _i , r _j ) represents the color space similarity _{between the segmented regions ri and r j ;}

Indicates the i-th channel and the k-th histogram vector, i=1, 2, 3, k=0, 1...., 25, n represents the number of histograms, ri _i represents the segmented area R={r ₁ ,r ₂ ,...r _n } i-th area;

Represents the jth channel and the kth histogram vector, j=1, 2, 3, m represents m histograms;

Calculate the similarity of edge information, calculate the Gaussian differential with variance σ=1 for 8 different directions of each color channel, obtain 10 histograms for each color of each channel and use L1 normalization to calculate the similarity of edge information, Its calculation formula is as follows:

表示第i个通道、第k个直方图向量，

表示第j个通道、第k个直方图向量，其中i＝1,2,3，j＝1,2,3，k＝0,1....,10；n表示直方图个数，m表示m个直方图，r_i表示分割区域R＝{r₁,r₂,...r_n}第i个区域；Among them, f _edage (r _i , r _j ) represents the edge information similarity _{between the segmented regions ri and r j ,}

represents the i-th channel, the k-th histogram vector,

Represents the jth channel and the kth histogram vector, where i=1,2,3, j=1,2,3, k=0,1....,10; n represents the number of histograms, m represents m histograms, and ri represents the _i -th region of the segmented region R={r ₁ , _{r 2} , ... rn };

Calculate the similarity of the size of the area, and its calculation formula is as follows:

where f _area (r _i , r _j ) represents the _{similarity of the size of the segmented areas ri and r j ,} area() represents the area area; area(img) represents the image area, and _ri represents the segmented area R={r ₁ ,r ₂ ,...rn } i- _th area, r _j represents the division area R={r ₁ ,r ₂ ,... rn } j- _th area;

The color information similarity, edge information similarity, and area size similarity are fused, and the calculation formula is as follows:

f(r _i ,r _j )=w ₁ f _color (r _i ,r _j )+w ₂ f _edage (r _i ,r _j )+w ₃ f _area (r _i ,r _j ),

Among them, f(r _i , r _j ) represents the similarity of the segmentation regions ri and r _{j after} fusion, w ₁ , w ₂ , and w ₃ represent the weights of information similarity, edge information similarity, and region size similarity, respectively , ri represents the _ith area of the divided area R={r ₁ , r ₂ ,...rn }, _{r j} represents the _jth area of the divided area R={r ₁ , r ₂ ,... rn } .

The described construction of a multi-scale feature fusion network includes the following steps:

Set up an n-layer multi-scale neural network, and perform deconvolution operations from the top layer to generate a deconvolution layer;

Set the input of the first layer as the training sample, output the feature map of the first layer, the feature map of the first layer as the input of the second layer, output the feature map of the second layer, the feature map of the second layer as the input of the third layer, … Until the n-1th layer feature map is used as the input of the nth layer;

The first layer feature map, the second layer feature map...nth layer feature map and the corresponding first layer deconvolution layer, second layer deconvolution layer, ... nth layer deconvolution layer through 1*1 volume The kernel connection is accumulated to generate a multi-scale feature fusion network.

The described training multi-scale feature fusion network includes the following steps:

Input the training samples into the positioning model, and the positioning model locates the candidate regions of the training samples;

The candidate regions of the training samples are respectively input into the first layer of the multi-scale neural network, and the first layer of the multi-scale neural network outputs the first layer feature map;

The feature map of the first layer is input to the second layer of the multi-scale neural network, and the first layer of the multi-scale neural network outputs the feature map of the second layer, until the feature map of the n-1 layer is input to the nth layer of the multi-scale neural network;

Perform a deconvolution operation on the feature map of the nth layer, generate the nth layer convolutional layer;

The first layer feature map, the second layer feature map, ... until the nth layer feature map and the first layer deconvolution layer, the second layer deconvolution layer, ... until the nth layer deconvolution layer through 1*1 volume Accumulation nuclear connection;

The feature map of the first layer and the deconvolution layer of the first layer are connected by a 1*1 convolution kernel to extract the features of the first layer, and then the prediction result of the first layer is generated; the feature map of the second layer and the deconvolution layer of the second layer are passed through After the 1*1 convolution kernel is connected, the second layer features are extracted, and the second layer prediction result is generated; ... until the nth layer feature map and the nth layer deconvolution layer are connected by 1*1 convolution kernel, the nth layer is extracted. After the feature, the nth layer prediction result is generated;

Perform regression processing on the first-layer prediction results, the second-layer prediction results, ... until the n-th layer results to generate the final prediction results. The regression function is as follows:

Among them, C(λ) is the final prediction result, λ is the training parameter, n is the number of network layers, y ^(j) is the real category, p _λ (x ^(j) ) is the prediction result of the jth layer; x ^{( j)} represents the feature vector of the jth layer;

The final score prediction category and the coordinates of the category in the graph are obtained by C(λ).

The acquisition of the number of wheat spiders comprises the following steps:

Input the image to be counted into the positioning model, and the positioning model locates the candidate area of the image to be counted;

Input the candidate area of the image to be counted into the multi-scale neural network to obtain the predicted classification of the spider in the image, and count the number of spiders to obtain the number of spiders in the image.

beneficial effect

Compared with the prior art, the automatic counting method of wheat spiders based on the combination of a multi-scale feature fusion network and a positioning model of the present invention realizes the direct identification and counting of the wheat spiders in the field natural environment.

The invention eliminates the influence of illumination on the detection count through preprocessing, and simplifies the complex environment; then locates the candidate area of the suspected wheat spider through the localization model method; uses the multi-scale feature fusion network to extract the feature of the candidate area, and then uses the multi-scale feature fusion network to extract features. The prediction result regression finally determines the wheat spider area. The positioning (determination) of the candidate area greatly reduces the feature extraction time and feature dimension, and enhances the real-time performance of counting. Robustness and accuracy of detection counts.

Description of drawings

Fig. 1 is the method sequence diagram of the present invention;

Fig. 2a is the detection result diagram that adopts traditional SVM technology to training sample in the prior art;

Fig. 2b is the detection result diagram of utilizing the method of the present invention;

FIG. 3 is a schematic diagram of the structure of a multi-scale feature fusion network in the present invention.

Detailed ways

In order to have a further understanding and understanding of the structural features of the present invention and the effects achieved, the preferred embodiments and accompanying drawings are used in conjunction with detailed descriptions, and the descriptions are as follows:

As shown in Figure 1, a method for automatic counting of wheat spiders based on the combination of a multi-scale feature fusion network and a positioning model according to the present invention includes the following steps:

The first step is the establishment of training samples. Obtain more than 2000 images of wheat spiders in the natural environment of the field as training images, and the wheat spiders in the images have been marked to obtain training samples.

The second step is to construct the wheat spider detection and counting model. Construct a localization model and a multi-scale feature fusion network, use the localization model to extract the candidate area of the training sample, (locate the candidate area of the wheat spider), and then extract the features of the candidate area through the multi-scale fusion network and classify it. Spider output coordinate value, if it is not a wheat spider candidate area, cancel.

First, a positioning model is constructed. In order to reduce the time of feature extraction, reduce the dimension of feature vector, and enhance the real-time performance of automatic counting, the localization model is used first to locate the candidate area of wheat spider, and then the feature extraction is carried out according to the candidate area.

The steps are as follows:

(1) Setting a color space conversion module, the color space conversion module is used to convert the RGB color space into the YcbCr color space, and divide it into R={r ₁ , _{r 2} ,...rn } divided regions.

(2) Calculate the similarity of color information. Use L1 normalization to obtain 25 histograms of each color channel of the image, and calculate the similarity of the color space. The calculation formula is as follows:

表示第i个通道、第k个直方图向量，i＝1,2,3，k＝0,1....,25，n表示直方图个数，r_i表示分割区域R＝{r₁,r₂,...r_n}第i个区域；

表示第j个通道、第k个直方图向量，j＝1,2,3，m表示m个直方图。Among them, f _color (r _i , r _j ) represents the color space similarity _{between the segmented regions ri and r j ;}

Indicates the i-th channel and the k-th histogram vector, i=1, 2, 3, k=0, 1...., 25, n represents the number of histograms, ri _i represents the segmented area R={r ₁ ,r ₂ ,...r _n } i-th area;

Represents the jth channel and the kth histogram vector, j=1, 2, 3, m represents m histograms.

(3) Calculate edge information similarity. Calculate the Gaussian differential with variance σ=1 for 8 different directions of each color channel, obtain 10 histograms for each color channel and use the L1 norm to normalize, and calculate the edge information similarity. The calculation formula is as follows:

表示第i个通道、第k个直方图向量，

表示第j个通道、第k个直方图向量，其中i＝1,2,3，j＝1,2,3，k＝0,1....,10；n表示直方图个数，m表示m个直方图，r_i表示分割区域R＝{r₁,r₂,...r_n}第i个区域。Among them, f _edage (r _i , r _j ) represents the edge information similarity _{between the segmented regions ri and r j ,}

represents the i-th channel, the k-th histogram vector,

Represents the jth channel and the kth histogram vector, where i=1,2,3, j=1,2,3, k=0,1....,10; n represents the number of histograms, m represents m histograms, and ri represents the _ith region of the divided region R={r ₁ , _{r 2} , . . . rn }.

(4) Calculate the similarity of the size of the area, and the calculation formula is as follows:

where f _area (r _i , r _j ) represents the _{similarity of the size of the segmented areas ri and r j ,} area() represents the area area; area(img) represents the image area, and _ri represents the segmented area R={r ₁ ,r ₂ ,...rn } i- _th region, r _j represents the division region R={r ₁ ,r ₂ ,... rn } j- _th region.

(5) Fusion of color information similarity, edge information similarity, and area size similarity, and the calculation formula is as follows:

f(r _i ,r _j )=w ₁ f _color (r _i ,r _j )+w ₂ f _edage (r _i ,r _j )+w ₃ f _area (r _i ,r _j ),

Among them, f(r _i , r _j ) represents the similarity of the segmentation regions ri and r _{j after} fusion, w ₁ , w ₂ , and w ₃ represent the weights of information similarity, edge information similarity, and region size similarity, respectively , ri represents the _ith area of the divided area R={r ₁ , r ₂ ,...rn }, _{r j} represents the _jth area of the divided area R={r ₁ , r ₂ ,... rn } .

After the color information similarity, edge information similarity, and region size similarity are merged, the n regions finally generated by r _i and r _j are continuously combined, which are the candidate regions of wheat spiders.

Secondly, construct a multi-scale feature fusion network, and transform the structure of the multi-scale feature fusion network. In order to better extract the features of each scale and various forms of the wheat spider, a multi-scale feature fusion network is designed to accurately distinguish the exact region of the wheat spider candidate region.

As shown in Fig. 3, the multi-scale feature fusion network structure constructs a multi-scale feature network with margins by exploiting the feature maps of the inherent multi-scale and conical hierarchy, where a top-down architecture with lateral connections is developed. , for building high-level semantic feature maps at all scales, relying on a way to combine low-resolution but semantically strong features with high-resolution semantically weak features through top-down paths and lateral connections, so that Obtaining high-resolution and strong semantic features is conducive to the detection of small targets such as wheat spiders. The steps are as follows:

(1) Set up an n-layer multi-scale neural network, and perform deconvolution operations from the top layer to generate deconvolution layers.

(2) Set the input of the first layer as the training sample, output the feature map of the first layer, the feature map of the first layer as the input of the second layer, output the feature map of the second layer, and the feature map of the second layer as the feature map of the third layer Input, ... up to the n-1th layer feature map as the input of the nth layer.

(3) Pass the feature map of the first layer, the feature map of the second layer ... the feature map of the nth layer and the corresponding first layer deconvolution layer, the second layer deconvolution layer, ... the nth layer deconvolution layer through 1 *1 The convolution kernel is connected to generate a multi-scale feature fusion network.

Here, a feature map is generated by downsampling, each training image is used as input, and a multi-scale neural network is used to extract features, and each layer of the multi-scale neural network will generate a feature map through downsampling;

Then the last layer of deconvolution generates a feature map of the size of the previous layer and iterates in turn until the size of the second layer is generated. Since each layer of the multi-scale network will be downsampled, the feature map will become smaller and smaller, and the wheat spider in the feature image will be smaller or even several pixels in size, which has a great impact on the detection count of the wheat spider. In order to avoid this problem, the deconvolution operation is used for each layer of the pyramid image, and the feature map is enlarged to the size of the previous layer by upsampling, which can not only effectively extract the pest features, but also ensure the size of the wheat spider in the picture;

The feature maps of each layer generated by deconvolution are connected through a 1*1 convolution kernel to generate a multi-scale feature fusion network.

Finally, a multi-scale feature fusion network is trained. According to the positioning model, the candidate regions located by the training samples are used as features for training, and the output results of each layer are used as the prediction results. The specific steps are as follows:

(1) Input the training samples into the positioning model, and the positioning model locates the candidate regions of the training samples.

(2) The candidate regions of the training samples are respectively input into the first layer of the multi-scale neural network, and the first layer of the multi-scale neural network outputs the first layer feature map.

(3) The first layer feature map is input to the second layer of the multi-scale neural network, the first layer of the multi-scale neural network outputs the second layer feature map, until the n-1 layer feature map is input to the nth layer of the multi-scale neural network .

(4) Perform the deconvolution operation on the feature map of the nth layer to generate the nth layer of deconvolution layer, perform the deconvolution operation on the n-1th layer, and generate the n-1th deconvolution layer, thus until the th 1 deconvolution layer.

(5) The first layer feature map, the second layer feature map, ... until the nth layer feature map and the first layer deconvolution layer, the second layer deconvolution layer, ... until the nth layer deconvolution layer passes 1 *1 convolution kernel connection.

(6) After the first layer feature map and the first layer deconvolution layer are connected by a 1*1 convolution kernel, the first layer feature is extracted, and the first layer prediction result is generated; the second layer feature map and the second layer are reversed After the layer is connected by a 1*1 convolution kernel, the second layer features are extracted, and the second layer prediction result is generated; ... until the nth layer feature map and the nth layer deconvolution layer are connected by a 1*1 convolution kernel and then extracted After the nth layer features, the nth layer prediction result is generated.

(7) Perform regression processing on the first-layer prediction results, the second-layer prediction results, ... until the n-th layer results, and generate the final prediction results. The regression function is as follows:

Among them, C(λ) is the final prediction result, λ is the training parameter, n is the number of network layers, y ^(j) is the real category, p _λ (x ^(j) ) is the prediction result of the jth layer; x ^{( j)} represents the feature vector of the jth layer;

(8) Obtain the final score prediction category and the coordinates of the category in the picture through C(λ), and the coordinates are the position of the wheat spider in the picture.

The third step is to acquire the image to be counted. Acquire images of wheat spiders taken in the field, and perform preprocessing to obtain images to be counted.

The fourth step is to obtain the number of wheat spiders. Input the image to be counted into the wheat spider detection and counting model to obtain the number of wheat spiders in the image. The specific steps are as follows:

(1) Input the image to be counted into the positioning model, and the positioning model locates the candidate region of the image to be counted;

(2) Input the candidate region of the image to be counted into the multi-scale neural network to obtain the predicted classification of the spider in the image, and count the number of spiders to obtain the number of spiders in the image.

As shown in Figure 2a, it is the result of wheat spider detection obtained by using the SVM algorithm. As can be seen from Figure 2a, the small box detects a very large area of the spider, especially in the large box in the middle of Figure 2a, which mistakenly includes multiple relatively concentrated spiders into one large box. . The reason for this mislabeling is that the traditional SVM algorithm does not perform pre-positioning. If the positioning model is used to locate the candidate area first, this phenomenon can be avoided; in Figure 2a, the small-frame detected wheat spider area is very large. The reason is that the traditional SVM algorithm does not use the regression fusion of multiple prediction results, and in Figure 2a, there are still some small boxes that are misidentified.

As shown in Fig. 2b, compared with the traditional SVM algorithm, the present invention can accurately locate the number and specific positions of the spiders, and has higher robustness and accuracy.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions describe only the principles of the present invention. Without departing from the spirit and scope of the present invention, there are various Variations and improvements are intended to fall within the scope of the claimed invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims (2)

1. A wheat spider automatic counting method based on combination of a multi-scale feature fusion network and a positioning model is characterized by comprising the following steps:

11) establishing a training sample, namely acquiring more than 2000 images of the wheat spiders in the field natural environment as training images, and marking the wheat spiders in the images to obtain the training sample;

12) constructing a wheat spider detection counting model;

121) constructing a positioning model; the construction positioning model comprises the following steps:

1211) Setting a color space conversion module, wherein the color space conversion module is used for converting the RGB color space into the YcbCr color space and dividing the RGB color space into R (R) and { R ═ R₁,r₂,...r_n} segmented areas;

1212) calculating the similarity of color information, obtaining 25 histograms of each color channel of the image by normalization using an L1 paradigm, and calculating the similarity of color space according to the following calculation formula:

wherein, f_color(r_i,r_j) Represents a divided region r_iAnd r_jThe degree of similarity in the color space of (c),

denotes the a channel, k histogram vector, a is 1,2,3, k is 0,1_iDenotes a divided region R ═ { R ═ R₁,r₂,...r_nThe ith area;

represents the b-th channel, the k-th histogram vector, b ═ 1,2, 3;

1213) calculating the similarity of the edge information, calculating the Gaussian differential with the variance sigma being 1 for 8 different directions of each color channel, acquiring 10 histograms for each color of each channel, and normalizing by using an L1 paradigm to calculate the similarity of the edge information, wherein the calculation formula is as follows:

wherein f is_edage(r_i,r_j) Indicates a divided region r_iAnd r_jThe degree of similarity of the edge information of (2),

representing the e-th channel, the k-th histogram vector,

represents the f-th channel, the k-th histogram vector, where e is 1,2,3, f is 1,2,3, k is 0,1.., 10; q represents the number of histograms, r _iDenotes a divided region R ═ R { (R)₁,r₂,...r_nThe ith area;

1214) calculating the similarity of the sizes of the regions, wherein the calculation formula is as follows:

wherein f is_area(r_i,r_j) Represents a divided region r_iAnd r_jArea () represents an area of the region; area (img) represents the area of the picture, r_iDenotes a divided region R ═ { R ═ R₁,r₂,...r_nI-th area, r_jDenotes a divided region R ═ { R ═ R₁,r₂,...r_nJth area;

1215) and fusing the color information similarity, the edge information similarity and the region size similarity, wherein the calculation formula is as follows:

f(r_i,r_j)＝w₁f_color(r_i,r_j)+w₂f_edage(r_i,r_j)+w₃f_area(r_i,r_j)，

wherein, f (r)_i,r_j) Represents a scoreCutting region r_iAnd r_jSimilarity after fusion, w₁、w₂、w₃Weights r respectively representing information similarity, edge information similarity, and region size similarity_iDenotes a divided region R ═ { R ═ R₁,r₂,...r_nI-th area, r_jDenotes a divided region R ═ { R ═ R₁,r₂,...r_nJth area;

122) constructing a multi-scale feature fusion network, and transforming a multi-scale feature fusion network structure; the construction of the multi-scale feature fusion network comprises the following steps:

1221) setting n layers of multi-scale neural networks, and performing deconvolution operation from the topmost layer to generate a deconvolution layer;

1222) setting the input of the layer 1 as a training sample, outputting a layer 1 feature map, taking the layer 1 feature map as the input of the layer 2 and outputting a layer 2 feature map, taking the layer 2 feature map as the input of the layer 3, and taking … to the layer n-1 feature map as the input of the layer n;

1223) Connecting the feature map of the layer 1 and the feature map of the layer 2 … of the nth layer with the corresponding deconvolution layer of the layer 1, the deconvolution layer of the second layer and the deconvolution layer of the layer … of the nth layer through 1-1 convolution kernels to generate a multi-scale feature fusion network;

123) training a multi-scale feature fusion network, training by taking a candidate region positioned by a positioning model aiming at a training sample as a feature, and taking an output result of each layer as a prediction result;

the training multi-scale feature fusion network comprises the following steps:

1231) inputting the training sample into a positioning model, and positioning a candidate region of the training sample by the positioning model;

1232) respectively inputting the candidate regions of the training sample into the 1 st layer of the multi-scale neural network, and outputting a 1 st layer characteristic diagram by the 1 st layer of the multi-scale neural network;

1233) inputting the feature map of the layer 1 into the layer 2 of the multi-scale neural network, and outputting the feature map of the layer 2 by the layer 2 of the multi-scale neural network until the feature map of the layer n-1 is input into the nth layer of the multi-scale neural network;

1234) performing deconvolution operation on the n-th layer of feature map to generate an n-th deconvolution layer, and performing deconvolution operation on the n-1-th layer to generate an n-1-th deconvolution layer, so as to obtain a 1-st deconvolution layer;

1235) The characteristic diagrams of the 1 st layer, the 2 nd layer, … to the nth layer are connected with the deconvolution layer of the 1 st layer, the deconvolution layer of the 2 nd layer, … to the nth layer by 1 × 1 convolution kernels;

1236) connecting the feature map of the 1 st layer with the deconvolution layer of the 1 st layer through a 1 x 1 convolution kernel, extracting features of the first layer, and generating a prediction result of the first layer; connecting the feature map of the 2 nd layer with the deconvolution layer of the 2 nd layer through a 1 x 1 convolution kernel, extracting features of a second layer, and generating a prediction result of the second layer; …, generating an nth layer of prediction result until the nth layer of feature graph is connected with the nth layer of deconvolution layer through a 1 x 1 convolution kernel and then the nth layer of feature is extracted;

1237) and performing regression processing on the prediction result of the 1 st layer, the prediction result of the 2 nd layer, … till the result of the nth layer to generate a final prediction result, wherein the regression function is as follows:

wherein C (lambda) is the final prediction result, lambda represents the training parameter, n represents the number of network layers, y represents^(j)Representing true classes, p_λ(x^(j)) Representing the result of the prediction of the j-th layer; x is the number of^(j)A feature vector representing a j-th layer;

1238) obtaining a final score through C (lambda) to predict the coordinates of the category and the position of the category in the graph;

13) acquiring an image to be counted, acquiring a wheat spider image shot in the field, and preprocessing the image to be counted;

14) And (3) obtaining the number of the wheat spiders, namely inputting the image to be counted into a wheat spider detection counting model to obtain the number of the wheat spiders in the image.

2. The method for automatically counting the number of the wheat spiders based on the combination of the multi-scale feature fusion network and the positioning model as claimed in claim 1, wherein the obtaining of the number of the wheat spiders comprises the following steps:

21) inputting the image to be counted into a positioning model, and positioning a candidate area of the image to be counted by the positioning model;

22) inputting the candidate area of the image to be counted into a multi-scale neural network to obtain the prediction classification of the wheat spiders in the image, and counting the number of the wheat spiders to obtain the number of the wheat spiders in the image.

Images