CN112036437A - Rice seedling detection model based on improved YOLOV3 network and method thereof - Google Patents

Abstract

The invention discloses a rice seedling detection model based on an improved YOLOV3 network and a method thereof, which solves the disadvantages of poor adaptability and robustness of traditional image processing methods for rice seedling detection and poor image processing algorithm effect; The main points of the scheme include a feature extraction module that performs multi-scale feature extraction on the input rice seedling image, and a multi-scale prediction module that predicts the position of rice seedlings according to the feature map extracted from the multi-scale features; the multi-scale prediction module includes a multi-scale prediction module. A multi-scale fusion feature building module for merging and constructing the extracted feature maps, a multi-scale seedling position prediction module for predicting the position of seedlings according to the feature maps constructed by fusion, the rice seedling detection model and method based on the improved YOLOV3 network of the present invention , to solve the problem of poor robustness and accuracy of rice seedling detection in complex paddy field environments with high and low weed densities, different lighting conditions, and lack of seedlings.

Description

Rice seedling detection model and method based on improved YOLOV3 network

technical field

The invention relates to the technical field of artificial intelligence, in particular to a rice seedling detection model based on an improved YOLOV3 network and a method thereof.

Background technique

Rice is one of the three major food crops in the world. Its cultivation area and total output are second only to wheat, and it is the staple food for more than half of the world's population. At present, the weeding methods used in paddy fields are mainly spraying chemical agents and artificial weeding. The use of chemical agents can easily cause problems such as excessive pesticide residues, environmental pollution and damage to the ecological chain. Manual weeding is time-consuming, laborious, inefficient and costly. Therefore, the demand for mechanized weeding of rice is more and more urgent. Rice mechanized weeding technology is of great significance for reducing the use of chemical agents, reducing ecological environment pollution and improving food security.

Rice seedling detection is one of the key links in mechanized weeding of rice, and plays an important guiding role in precision weeding and fertilization in precision agriculture. The position information of the rice seedlings can be obtained by detecting the rice seedlings, which can guide the action of the weeding agency to remove the weeds between the rows before the rice is closed.

There are two main types of methods for crop and weed detection. The first type is based on image processing. The crop is segmented from the background, and then the centerline of the crop line is detected. The second category is deep learning based methods for crop and weed detection. Due to the complex paddy field environment and many factors affecting the detection of rice seedlings, traditional image processing methods have poor adaptability and robustness to the detection of rice seedlings. Especially when the paddy field image does not change significantly in R, G and B pixel intensities, when there are many weeds or when the size is similar to the crops, the traditional image processing algorithm does not work well.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a rice seedling detection model and method based on the improved YOLOV3 network, which solves the problem of poor robustness and accuracy of rice seedling detection under various conditions in a complex paddy field environment.

The above-mentioned technical purpose of the present invention is achieved through the following technical solutions:

A rice seedling detection model based on an improved YOLOV3 network, including a feature extraction module that performs multi-scale feature extraction on an input rice seedling image to obtain a multi-scale rice seedling feature map, and performs a multi-scale seedling feature map. Predicted multi-scale prediction module;

The multi-scale prediction module includes a multi-scale fusion feature construction module that performs fusion construction processing on multi-scale seedling feature maps to obtain multi-scale fusion feature maps, and a multi-scale fusion feature construction module that predicts the position of the seedlings according to the fusion features corresponding to the fusion feature maps. Multi-scale seedling location prediction module.

Preferably, the feature extraction module includes a 52-scale feature extraction sub-module for sequentially performing magnification sampling and extraction processing on the input rice seedling images to obtain a 52-scale seedling feature map, a 26-scale seedling feature map, and a 13-scale seedling feature map, respectively. 26-scale feature extraction sub-module, 13-scale feature extraction sub-module.

Preferably, the multi-scale fusion feature construction module includes a 52-scale fusion feature construction sub-module, a 26-scale fusion feature construction sub-module, and a 13-scale fusion feature construction sub-module for fusing the seedling feature maps of different scales.

Preferably, the 52-scale feature extraction sub-module performs feature extraction on the input seedling feature map of the 52-scale feature extraction sub-module to obtain a 52-scale seedling feature map including the 52-scale seedling features; the 26-scale feature extraction sub-module is for The input seedling feature map of the 26-scale feature extraction sub-module performs feature extraction to obtain a 26-scale seedling feature map containing the 26-scale seedling features; the 13-scale feature extraction sub-module pair is the input seedling feature map of the 13-scale feature extraction sub-module Feature extraction is performed to obtain a 13-scale seedling feature map containing 13-scale seedling features.

A detection method for a rice seedling detection model based on an improved YOLOV3 network, comprising the following steps:

Obtain the input rice seedling image;

The images of rice seedlings are down-sampled at different magnifications respectively, and the 52-scale seedling feature map, the 26-scale seedling feature map, and the 13-scale seedling feature map are obtained through the feature extraction module;

Through the multi-scale fusion feature building module, feature fusion is performed on the obtained 13-scale seedling feature map, 26-scale seedling feature map, and 52-scale seedling feature map in turn to obtain a 13-scale seedling fusion feature map, a 26-scale seedling fusion feature map, and a 52-scale seedling feature map. fusion feature map;

Through the multi-scale seedling position prediction module, the obtained 13-scale seedling fusion feature map, 26-scale seedling fusion feature map, and 52-scale seedling fusion feature map are successively predicted to obtain 13-scale seedling prediction results, 26-scale seedling prediction results, and 52-scale seedling prediction results. forecast result.

Preferably, the specific steps of feature extraction performed by the feature extraction module are as follows:

The input convolution layer of the 52-scale feature extraction sub-module is obtained through 10 convolution layers and 3 residual layers for the input rice seedling image, and the 52-scale seedling features are extracted through the 52-scale feature extraction sub-module;

The output layer of the 52-scale feature extraction sub-module is convolved with a 3*3/2*512 convolution kernel to obtain the input convolution layer of the 26-scale feature extraction sub-module, and the 26-scale feature extraction sub-module is extracted to obtain 26-scale seedlings feature;

The output layer of the 26-scale feature extraction sub-module is convolved with a convolution kernel of 3*3/2*512 to obtain the input convolution layer of the 13-scale feature extraction sub-module, and the 13-scale seedlings are extracted by the 13-scale feature extraction sub-module. feature.

Preferably, the multi-scale fusion feature building module performs the following specific steps for feature fusion construction:

The output layer of the 26-scale feature extraction sub-module is downsampled to obtain a 13*13*512 feature map, and tensor splicing is performed with the output layer of the 13-scale feature extraction sub-module to obtain a 13-scale seedling fusion feature map;

After downsampling the output layer of the 52-scale feature extraction sub-module, a 26*26*256 feature map is obtained, and the output layer of the 13-scale feature extraction sub-module is upsampled to obtain a 26*26*512 feature map. The 26*26*256 feature map, the 26*26*512 feature map obtained by upsampling, and the output layer of the 26-scale feature extraction sub-module are spliced by tensor, and the 26-scale seedling fusion feature map is obtained;

The 104-scale output seedling feature map obtained by feature extraction from the input rice seedling image is down-sampled to obtain a 52*52*128 feature map, and the output layer of the 26-scale feature extraction sub-module is upsampled to obtain a 52*52*512 feature map. , the 52*52*128 feature map obtained by downsampling, the 52*52*512 feature map obtained by upsampling and the output layer of the 52-scale feature extraction sub-module are tensor-spliced to obtain a 52-scale seedling fusion feature map.

Preferably, the specific prediction steps of the multi-scale seedling position prediction model are as follows:

The 13-scale seedling fusion feature constructed by the 13-scale fusion feature construction sub-module is obtained by convolution of a set of convolution kernels, a 3*3*512 convolution kernel and a 1*1*18 convolution kernel to obtain a 13-scale fusion feature. The prediction result is 13*13*18;

The 26-scale seedling fusion feature constructed by the 26-scale fusion feature construction sub-module is obtained by convolution of a set of convolution kernels, a 3*3*512 convolution kernel and a 1*1*18 convolution kernel to obtain a 26-scale fusion feature. The prediction result is 26*26*18;

The 52-scale seedling fusion feature constructed by the 52-scale fusion feature construction sub-module is obtained by convolution of a set of convolution kernels, a 3*3*256 convolution kernel and a 1*1*18 convolution kernel to obtain a 52-scale fusion feature. The predicted result is 52*52*18.

To sum up, the present invention has the following beneficial effects:

The rice seedling detection model based on the improved Yolov3 network can automatically extract the characteristics of rice seedlings. In the original feature extraction network, the model uses the 52-scale feature extraction sub-module to replace the original residual unit for the 52-scale feature map that is close to the rice seedling size. In the wide network, sampling features of different scales are added when constructing fusion features of various scales. The fusion features increase the combination of low-level location information and high-level semantic information, and the resulting fusion feature levels of rice seedlings at various scales are more abundant. , the improved rice seedling detection model has better accuracy and robustness for the high and low weed density distribution in complex paddy field environment, different light conditions, and the absence of rice seedlings.

Description of drawings

Fig. 1 is the structural schematic block diagram of the detection model;

Fig. 2 is the network structure diagram of the detection model;

Fig. 3 is the schematic flow chart of the detection method of the present invention;

Figure 4 is a test p-r curve diagram based on the improved Yolov3 and Yolov3 rice seedling detection models.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings.

According to one or more embodiments, a rice seedling detection model based on an improved YOLOV3 network is disclosed, as shown in FIG. 1 , which includes performing multi-scale rice seedling feature extraction on an input rice seedling image to obtain multi-scale rice seedling features The feature extraction module of the map, and the multi-scale prediction module for predicting the position of rice seedlings according to the multi-scale seedling feature map extracted by the feature extraction module.

As shown in Figure 1, the feature extraction module includes a 52-scale feature extraction sub-module, a 26-scale feature extraction sub-module, and a 13-scale feature extraction sub-module that sequentially perform magnification sampling and extraction processing on the input rice seedling image. For the sake of clarity, the input image of rice seedlings is an image of 416*416 scale. When sampling and extraction, down-sampling is performed by 8 times, 16 times and 32 times respectively.

The multi-scale prediction module includes a multi-scale fusion feature construction module that fuses and constructs the seedling feature maps of different scales extracted by the feature extraction module to obtain a multi-scale fusion feature map. A multi-scale seedling position prediction module for seedling position prediction. Among them, the multi-scale fusion feature building module specifically includes a 52-scale fusion feature building sub-module, a 26-scale fusion feature building sub-module, and a 13-scale fusion feature building sub-module. The multi-scale seedling position prediction module specifically includes a 52-scale position predictor. module, 26-scale position prediction sub-module, 13-scale position prediction sub-module.

The input seedling feature map extracted by the 52-scale feature extraction sub-module is specifically: the input convolutional layer of the 52-scale feature extraction sub-module is obtained after 10 convolution layers and 3 residual layers for the input rice seedling image as the 52-scale input convolution layer. The input seedling feature map of the feature extraction submodule. The 52-scale feature extraction sub-module performs feature extraction on its input seedling feature map to obtain a 52-scale seedling feature map containing 52-scale seedling features.

The input seedling feature map extracted by the 26-scale feature extraction sub-module is specifically: the output layer of the 52-scale feature extraction sub-module, that is, the output 52-scale seedling feature map, is convolved with a 3*3/2*512 convolution kernel. The input convolution layer of the 26-scale feature extraction sub-module is obtained, and the input seedling feature map of the 26-scale feature extraction sub-module is obtained. The 26-scale feature extraction sub-module performs feature extraction on its input seedling feature map to obtain a 26-scale seedling feature map containing 26-scale seedling features.

The input seedling feature map extracted by the 13-scale feature extraction sub-module is specifically: the output layer of the 26-scale feature extraction sub-module, that is, the output 26-scale seedling feature map, is convolved with a 3*3/2*512 convolution kernel. The input convolution layer of the 13-scale feature extraction sub-module is obtained, and the input seedling feature map of the 13-scale feature extraction sub-module is obtained. The 13-scale seedling feature map containing the 13-scale seedling features is extracted by the 13-scale feature extraction sub-module. The feature extraction modules of the three scales are extracted sequentially.

The feature map constructed by the fusion of the 13-scale feature construction sub-module includes the output layer of the 13-scale feature extraction sub-module, which is the 13-scale seedling feature map extracted, and the output layer of the 26-scale feature extraction sub-module, which is the extracted 26-scale feature map. The 13*13*512 feature map is obtained by down-sampling the seedling feature map. The obtained 13-scale seedling feature map and the 13*13*512 feature map are tensor spliced to obtain the 13-scale seedling fusion feature map.

Similarly, the seedling feature map constructed by the fusion of the 26-scale feature construction sub-module includes the output layer of the 26-scale feature extraction sub-module, and also includes the output layer of the 52-scale feature extraction sub-module, that is, the extracted 52-scale seedling feature map. Perform downsampling to obtain a 26*26*256 feature map and upsample the output layer of the 13-scale feature extraction sub-module to obtain a 26*26*512 feature map. The obtained 26*26*512 feature map and the output layer of the 26-scale feature extraction sub-module are spliced by tensor to obtain the 26-scale seedling fusion feature map.

Similarly, the seedling feature map constructed by the 52-scale fusion feature construction sub-module includes the output layer of the 52-scale feature extraction sub-module, and also includes the 104-scale output seedling feature map obtained by feature extraction from the input rice seedling image. Sampling to obtain a 52*52*128 feature map and upsampling the output layer of the 26-scale feature extraction sub-module to obtain a 52*52*512 feature map. The 52*52*512 feature map and the output layer of the 52-scale feature extraction sub-module are spliced by tensor to obtain the 52-scale seedling fusion feature map.

The 13-scale position prediction sub-module of the multi-scale seedling position prediction module passes the 13-scale seedling fusion features through a set of convolution kernels, and a 3*3*512 convolution kernel and a 1*1*18 convolution kernel. The convolution obtains a prediction result of 13 scales 13*13*18.

The 26-scale position prediction sub-module obtains 26-scale prediction by convolution of 26-scale seedling fusion features through a set of convolution kernels, and a 3*3*512 convolution kernel and a 1*1*18 convolution kernel The result is 26*26*18.

The 52-scale position prediction sub-module obtains a 52-scale prediction by passing the 52-scale seedling fusion feature through a set of convolution kernels, and convolution with a 3*3*256 convolution kernel and a 1*1*18 convolution kernel The result is 52*52*18.

The original Yolov3 network model adopts the method of small-scale feature map upsampling and fusion of the scale feature map when each scale feature is fused and constructed. The present invention uses small-scale feature map upsampling based on the improved Yolov3 model. The scale feature map downsampling and the scale feature map are fused. The fused features increase the combination of low-level location information and high-level semantic information, and the resulting fusion feature levels of rice seedlings at various scales are more abundant.

According to one or more embodiments, a detection method for a rice seedling detection model based on an improved YOLOV3 network is disclosed, as shown in Figures 2 and 3, including the following steps:

Obtain the input rice seedling image;

The images of rice seedlings were down-sampled at different magnifications, and the feature maps of 52-scale seedlings, 26-scale seedling feature maps, and 13-scale seedling feature maps were obtained through the feature extraction module;

Through the multi-scale fusion feature building module, feature fusion is performed on the obtained 13-scale seedling feature map, 26-scale seedling feature map, and 52-scale seedling feature map in turn to obtain a 13-scale seedling fusion feature map, a 26-scale seedling fusion feature map, and a 52-scale seedling feature map. fusion feature map;

Through the multi-scale seedling position prediction module, the obtained 13-scale seedling fusion feature map, 26-scale seedling fusion feature map, and 52-scale seedling fusion feature map are successively predicted to obtain 13-scale seedling prediction results, 26-scale seedling prediction results, and 52-scale seedling prediction results. forecast result.

Specifically, the specific steps for the feature extraction module to perform feature extraction are as follows:

The input convolution layer of the 52-scale feature extraction sub-module is obtained through 10 convolution layers and 3 residual layers for the input rice seedling image, and the 52-scale seedling features are extracted through the 52-scale feature extraction sub-module;

The output layer of the 52-scale feature extraction sub-module is convolved with a 3*3/2*512 convolution kernel to obtain the input convolution layer of the 26-scale feature extraction sub-module, and the 26-scale feature extraction sub-module is extracted to obtain 26-scale seedlings feature;

The output layer of the 26-scale feature extraction sub-module is convolved with a convolution kernel of 3*3/2*512 to obtain the input convolution layer of the 13-scale feature extraction sub-module, and the 13-scale seedlings are extracted by the 13-scale feature extraction sub-module. feature.

Specifically, the specific steps of the multi-scale fusion feature construction module for feature fusion construction are as follows:

The output layer of the 26-scale feature extraction sub-module is downsampled to obtain a 13*13*512 feature map, and tensor splicing is performed with the output layer of the 13-scale feature extraction sub-module to obtain a 13-scale seedling fusion feature map;

After downsampling the output layer of the 52-scale feature extraction sub-module, a 26*26*256 feature map is obtained, and the output layer of the 13-scale feature extraction sub-module is upsampled to obtain a 26*26*512 feature map. The 26*26*256 feature map, the 26*26*512 feature map obtained by upsampling, and the output layer of the 26-scale feature extraction sub-module are spliced by tensor, and the 26-scale seedling fusion feature map is obtained;

The 104-scale output seedling feature map obtained by feature extraction from the input rice seedling image is down-sampled to obtain a 52*52*128 feature map, and the output layer of the 26-scale feature extraction sub-module is upsampled to obtain a 52*52*512 feature map. , the 52*52*128 feature map obtained by downsampling, the 52*52*512 feature map obtained by upsampling and the output layer of the 52-scale feature extraction sub-module are tensor-spliced to obtain a 52-scale seedling fusion feature map.

The 52-scale feature extraction sub-module is the Inception module, which uses a three-way convolution to widen the network: the first convolution uses a 1*1*64 convolution kernel to reduce the number of channels of the input seedling feature map, and then uses 3*3*128 and 3*3*64 two convolution kernels are used to extract features by convolution; in the second way, the 1*1*64 convolution kernel is used to reduce the number of channels of the input seedling feature map, and then the 3*3*128 convolution kernel is used for convolution. Product extraction features; the third channel reduces the number of channels of the input seedling feature map through 1*1*64 convolution. The 52-scale seedling feature map obtained after the three-way convolution processing of the Inception module, the 52*52*128 feature map and the 52*52*512 feature map obtained by up/down sampling are tensor spliced to complete the 52-scale feature fusion. Among them, the number of Inception modules is preferably set to 4, instead of the original residual unit to widen the network, more features can be extracted, and parameters and training time can be effectively reduced.

Further, the specific prediction steps of the multi-scale seedling position prediction model are as follows:

The 13-scale seedling fusion feature of the 13-scale seedling fusion feature map is subjected to the convolution of a set of convolution kernels, a 3*3*512 convolution kernel and a 1*1*18 convolution kernel to obtain a 13-scale prediction result. 13*13*18;

The 26-scale seedling fusion feature of the 26-scale seedling fusion feature map is subjected to the convolution of a set of convolution kernels, a 3*3*512 convolution kernel and a 1*1*18 convolution kernel to obtain a 26-scale prediction result. 26*26*18;

The 52-scale seedling fusion feature of the 52-scale seedling fusion feature map is subjected to the convolution of a set of convolution kernels, a 3*3*256 convolution kernel and a 1*1*18 convolution kernel to obtain a 52-scale prediction result. 52*52*18.

In order to express clearly, an example is given. The rice seedling detection model based on the improved Yolov3 and Yolov3 is tested using the rice seedling test set, and p-r curves are drawn respectively after the test, as shown in Figure 4(a), (b), a) The picture shows the p-r curve of the improved Yolov3, and (b) the picture shows the p-r curve of the Yolov3. When the confidence threshold is 0.5, the accuracy of the rice seedling detection model based on the improved Yolov3 and Yolov3 networks is 0.82 and 0.48 respectively, so based on The improved Yolov3 rice seedling detection algorithm improves the accuracy by about 34%.

This specific embodiment is only an explanation of the present invention, and it does not limit the present invention. Those skilled in the art can make modifications without creative contribution to the present embodiment as needed after reading this specification, but as long as the rights of the present invention are used All claims are protected by patent law.

Claims (8)

1. A rice seedling detection model based on an improved YOLOV3 network is characterized in that: the system comprises a feature extraction module for performing multi-scale feature extraction on an input rice seedling image to obtain a multi-scale seedling feature map, and a multi-scale prediction module for predicting the position of a rice seedling according to the multi-scale seedling feature map;

the multi-scale prediction module comprises a multi-scale fusion characteristic construction module for performing fusion construction processing on the multi-scale seedling characteristic diagram to obtain a multi-scale fusion characteristic diagram, and a multi-scale seedling position prediction module for predicting the seedling position according to the corresponding fusion characteristic in the fusion characteristic diagram.

2. The improved YOLOV3 network-based rice seedling detection model of claim 1, wherein: the characteristic extraction module comprises a 52-scale characteristic extraction submodule, a 26-scale characteristic extraction submodule and a 13-scale characteristic extraction submodule which are used for sequentially carrying out multiplying factor sampling extraction processing on the input rice seedling image so as to respectively obtain a 52-scale seedling characteristic diagram, a 26-scale seedling characteristic diagram and a 13-scale seedling characteristic diagram.

3. The improved YOLOV3 network-based rice seedling detection model of claim 2, wherein: the multi-scale fusion feature construction module comprises a 52-scale fusion feature construction submodule for performing fusion processing on seedling feature maps with different scales, a 26-scale fusion feature construction submodule and a 13-scale fusion feature construction submodule.

4. The rice seedling detection model based on the improved YOLOV3 network as claimed in claim 3, wherein: the 52-scale feature extraction sub-module performs feature extraction on the input seedling feature map of the 52-scale feature extraction sub-module to obtain a 52-scale seedling feature map containing 52-scale seedling features; the 26-scale feature extraction sub-module performs feature extraction on the input seedling feature map of the 26-scale feature extraction sub-module to obtain a 26-scale seedling feature map containing 26-scale seedling features; and the 13-scale feature extraction sub-module performs feature extraction on the input seedling feature map of the 13-scale feature extraction sub-module to obtain a 13-scale seedling feature map containing 13-scale seedling features.

5. A detection method of a rice seedling detection model based on an improved YOLOV3 network is characterized by comprising the following steps:

acquiring an input rice seedling image;

respectively carrying out down-sampling with different multiplying powers on the rice seedling images, and obtaining a 52-scale seedling characteristic diagram, a 26-scale seedling characteristic diagram and a 13-scale seedling characteristic diagram through a characteristic extraction module;

sequentially performing feature fusion on the obtained 13-scale seedling feature map, 26-scale seedling feature map and 52-scale seedling feature map through a multi-scale fusion feature construction module to obtain a 13-scale seedling fusion feature map, a 26-scale seedling fusion feature map and a 52-scale seedling fusion feature map;

and predicting the obtained 13-scale seedling fusion characteristic graph, 26-scale seedling fusion characteristic graph and 52-scale seedling fusion characteristic graph in sequence through a multi-scale seedling position prediction module to obtain a 13-scale seedling prediction result, a 26-scale seedling prediction result and a 52-scale seedling prediction result.

6. The improved YOLOV3 network-based rice seedling detection method as claimed in claim 5, wherein the feature extraction module performs the following steps:

the input convolution layer of the 52-scale feature extraction submodule is obtained by passing the input rice seedling image through 10 convolution layers and 3 residual error layers, and the 52-scale seedling features are obtained by extraction of the 52-scale feature extraction submodule;

convolving the output layer of the 52-scale feature extraction submodule by a convolution kernel of 3 × 3/2 × 512 to obtain an input convolution layer of the 26-scale feature extraction submodule, and extracting by the 26-scale feature extraction submodule to obtain 26-scale seedling features;

and (3) convolving the output layer of the 26-scale feature extraction submodule by a convolution kernel of 3 × 3/2 × 512 to obtain an input convolution layer of the 13-scale feature extraction submodule, and extracting by the 13-scale feature extraction submodule to obtain 13-scale seedling features.

7. The improved YOLOV3 network-based rice seedling detection method as claimed in claim 6, wherein the multi-scale fusion feature construction module performs the specific steps of feature fusion construction as follows:

carrying out down-sampling on the output layer of the 26-scale feature extraction sub-module to obtain a 13 x 512 feature map, and carrying out tensor splicing on the feature map and the output layer of the 13-scale feature extraction sub-module to obtain a 13-scale seedling fusion feature map;

carrying out down-sampling on an output layer of the 52-scale feature extraction sub-module to obtain 26 × 256 feature maps, carrying out up-sampling on an output layer of the 13-scale feature extraction sub-module to obtain 26 × 512 feature maps, and carrying out tensor splicing on the 26 × 256 feature maps obtained by down-sampling, the 26 × 512 feature maps obtained by up-sampling and the output layer of the 26-scale feature extraction sub-module to obtain 26-scale seedling fusion feature maps;

and carrying out down-sampling on 104-scale output seedling feature maps obtained by carrying out feature extraction on the input rice seedling pictures to obtain 52 x 128 feature maps, carrying out up-sampling on output layers of the 26-scale feature extraction sub-modules to obtain 52 x 512 feature maps, and carrying out tensor splicing on the 52 x 128 feature maps obtained by down-sampling, the 52 x 512 feature maps obtained by up-sampling and the output layers of the 52-scale feature extraction sub-modules to obtain 52-scale seedling fusion feature maps.

8. The improved YOLOV3 network-based rice seedling testing method as claimed in claim 7, wherein the multi-scale seedling position prediction model comprises the following steps:

the 13-scale seedling fusion features constructed by the 13-scale fusion feature construction submodule are subjected to convolution of a set of convolution kernels, 13 × 512 convolution kernels and 1 × 18 convolution kernel to obtain a 13-scale prediction result 13 × 18;

the 26-scale seedling fusion features constructed by the 26-scale fusion feature construction submodule are subjected to convolution of a set of convolution kernels, 13 × 512 convolution kernels and 1 × 18 convolution kernel to obtain a 26-scale prediction result 26 × 18;

the 52-scale seedling fusion features constructed by the 52-scale fusion feature construction submodule are subjected to convolution of a set of convolution kernels, 13 × 256 convolution kernel and 1 × 18 convolution kernel to obtain a 52-scale prediction result 52 × 18.

Images