A cordyceps locating system based on binocular stereo and deep learning technology

Abstract

The application discloses a kind of based on binocular stereo and deep learning technology's worm grass location system, it is related to worm grass collection technical field, including: hardware structure and software system, wherein hardware structure includes mountain crutch, support crossbeam, CCD camera, controller, battery, camera support, controller support frame, battery support frame and tablet computer;Software system includes image acquisition module, training learning module, identification module and interactive module;Image acquisition module is based on CCD camera and collects worm grass image, and the worm grass image collected is uploaded to training learning module inside.The image after identification is sent to tablet computer in wireless communication mode, and display software is deployed on tablet computer, and when worm grass is found, tablet computer will issue prompt sound, while worm grass is framed on picture, compared with traditional worm grass collection method, with the advantages of fast identification efficiency, high accuracy, can effectively reduce the labor intensity of worm grass collection personnel, improve collection efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Cordyceps collection, in particular to a Cordyceps locating system based on binocular stereo and deep learning technology. BACKGROUND

[0002] Cordyceps is a kind of fungus of Clavicipitaceae, also known as "Cordyceps sinensis", which is a special product of China. It is a strange thing that the worm and the grass grow together. In winter, it is a worm, and in summer, it grows into grass from the worm. The worm is the larva of the Cordyceps moth, and the grass is a kind of Cordyceps fungus. In summer, the worm lays eggs on the leaves of the grass, and the leaves fall to the ground. After about a month, the larvae hatch and then drill into the moist and soft soil layer. The spores of the Cordyceps fungus in the soil layer invade the healthy and well-developed larvae. After the larvae are invaded by the spores, they drill into the shallow layer of the ground. The spores grow in the body of the larvae, and the internal organs of the larvae disappear slowly. The body of the larvae becomes a shell full of mycelium, which is buried in the soil layer. After a winter, the mycelium starts to grow in the spring of the next year and grows into a small grass on the ground in summer. In this way, the shell of the larvae and the small grass together form a complete "Cordyceps sinensis". Cordyceps is mainly produced in the Qinghai-Tibet Plateau in Sichuan and Qinghai of China. As a kind of traditional Chinese medicine, Cordyceps is favored by people because of its high medicinal value and medical effect. However, due to the harsh growing environment, it is difficult to collect Cordyceps, which leads to a high price of Cordyceps in the market. In recent years, with the improvement of people's living standards, the demand for Cordyceps is increasing. In the prior art, the collection of wild Cordyceps is mainly carried out by manual identification and excavation. Since the ground part of Cordyceps is not easy to distinguish from the environment and weeds, manual collection is usually carried out by lying or squatting on the ground, which not only has high labor intensity and low efficiency, but also causes large-scale damage to the vegetation. To solve the above problems, a Cordyceps locating system based on binocular stereo and deep learning technology is provided. SUMMARY

[0003] To solve the above technical problems, a Cordyceps locating system based on binocular stereo and deep learning technology is provided. The technical scheme solves the problems of high labor intensity, low efficiency and large-scale damage to vegetation in manual collection.

[0004] To achieve the above purposes, the technical scheme adopted by the present application is as follows: a Cordyceps locating system based on binocular stereo and deep learning technology, comprising: a hardware structure and a software system, wherein the hardware structure comprises a mountaineering walking stick, a supporting cross beam, a CCD camera, a controller, a battery, a camera support, a controller support, a battery support and a tablet computer;

[0005] The software system comprises an image acquisition module, a training and learning module, an identification module and an interaction module.

[0006] The image acquisition module acquires the image of the caterpillar fungus based on a CCD camera, and uploads the acquired image of the caterpillar fungus to the training learning module;

[0007] The training learning module trains the caterpillar fungus image data based on a deep learning framework, learns and optimizes the CA attention mechanism, adds the CA attention mechanism in front of SPPF in the BackBone of the yolov5s model, and constructs a caterpillar fungus recognition model;

[0008] The recognition module inputs the collected image data based on the constructed caterpillar fungus recognition model, reconstructs the caterpillar fungus point cloud based on the deep learning model, performs three-dimensional positioning, extracts the position of the caterpillar fungus from the color image, and judges and analyzes the position of the caterpillar fungus.

[0009] The interactive module performs human-computer interaction based on a tablet computer, serves as a handheld terminal of a caterpillar fungus collector, is deployed with a client software, receives the returned caterpillar fungus image in real time, instructs the caterpillar fungus collector to collect the caterpillar fungus, and emits an indication sound after recognizing the caterpillar fungus.

[0010] Preferably, one end of the support beam is a semicircular structure matched with a trekking stick, and is fixed on the trekking stick through a hexagonal bolt, a CCD camera, a controller and a battery are arranged on the support beam, the battery is used to power the controller, the CCD camera is connected with the controller through a USB or a network cable, and the tablet computer is manually held.

[0011] Preferably, the CA attention mechanism structure firstly performs average pooling operation on the horizontal and vertical directions of the acquired image, integrates the feature information of the horizontal and vertical directions and performs convolution processing, and the specific steps are as follows:

[0012] Given an input layer X with a size of HxWxC, using a pooling kernel with a size of (H, 1) and (1, W), each channel is encoded along the horizontal and vertical directions, and the outputs of the cth channel with a height of H and the cth channel with a width of W are respectively:

[0013]

[0014] In the formula, H and W respectively represent the height and width of the feature layer, and are known, and and represent the output results of the cth channel in the H and W directions, respectively, x c (h,i) and x c (j,w) represent the input of the input feature X along the H and W directions.

[0015] Preferably, the two independent direction feature maps output by the above two groups of formulas are subjected to splicing operation, 1x1 convolution operation and activation operation, and the formula is:

[0016] f=σ(F1[Zh ,Z w ])

[0017] where f denotes the intermediate feature map after encoding the spatial information in horizontal and vertical directions, [.,.] denotes the concatenation operation in spatial dimension, F1 denotes the convolution operation, and δ denotes the Sigmoid function;

[0018] f is split along the spatial dimension, and then two 1x1 convolution operations are used to restore the channel number of the feature map to the same channel number as the input feature X, the attention weight of the feature map is calculated using the Sigmoid function, and the obtained result is:

[0019] g h =σ(F h (f h ))

[0020] g w =σ(F w (f w ))

[0021] where g h and g w denote the attention weights of the convolution and activation after splitting f, f h and f w denote the feature maps along the H direction and the W direction after splitting f, and F h and F w denote the convolution operations.

[0022] Preferably, g h and g w are multiplied with the input feature map weight to weight the output feature map with the coordinate attention weight, and the formula is represented as:

[0023]

[0024] where y c (i,j) and x c (i,j) denote the input and output of the cth channel, respectively, and denote the attention weights along the H direction and the W direction on the cth channel, respectively.

[0025] Preferably, the recognition module recognizes the method by inputting the collected image data into the cordyceps recognition model, representing the position of the cordyceps by a bounding box box(x, y, width, height), x is the value of the upper left point of the rectangular box along the horizontal direction of the image coordinate, y is the value of the upper left point of the rectangular box along the vertical direction of the image coordinate, width and height are the length and height of the rectangular box along the horizontal direction and vertical direction of the image coordinate respectively, and the bounding box box(x, y, width, height) is the rectangular box output after algorithm recognition.

[0026] Preferably, the depth map of the cordyceps is obtained based on the image acquisition module, a rectangular box box(x, y, width, height) is drawn, the region of interest of the depth map is extracted, the region of interest is extracted based on the comprehensive rectangular box by using the image processing algorithm, and then the pixel points of the region of interest are reconstructed by using the formula of monocular vision single point reconstruction, and the point cloud after reconstruction is the point cloud of the cordyceps, that is, the position of the cordyceps is found, and the formula is:

[0027] x ccs =(u-cx)*z ccs / f x

[0028] y ccs =(v-cy)*z ccs / f y

[0029] z ccs =fb / d,d=u l -u r

[0030] In the formula, (x ccs , y ccs , z ccs ) represents the three-dimensional point in the camera coordinate system after reconstruction, (u, v) is the image coordinate of the pixel in the depth map, f is the focal length of the camera, (f x , f y ) is the component of the focal length of the camera, b is the baseline distance of the camera, d is the parallax of the camera, u l , u r are the projection points of the reconstructed point in the left camera and the right camera respectively.

[0031] Preferably, the tablet computer serves as a carrier of the interactive module, transmits the collected cordyceps picture to the client software on the tablet computer end through a network transmission link, and visually displays the analysis result to show the position of the cordyceps.

[0032] Compared with the prior art, the beneficial effects of the present application are:

[0033] The present application proposes a system combining binocular stereo vision technology, image fusion technology, deep learning technology and embedded technology, which recognizes and records the position of Cordyceps through deep learning, realizes the rapid identification and positioning of Cordyceps, and sends the identified image to a tablet computer in a wireless communication mode, and the tablet computer is deployed with display software, and when Cordyceps is found, the tablet computer will issue a prompt sound, and the Cordyceps is framed on the picture, compared with the traditional Cordyceps collection method, the method has the advantages of fast identification efficiency and high accuracy, and can effectively reduce the labor intensity of Cordyceps collection personnel and improve the collection efficiency;

[0034] The present application solves the problems of high labor intensity, low efficiency and high loss of Cordyceps when workers collect Cordyceps, thereby improving the collection efficiency and accuracy of Cordyceps, realizing the position indication and prompting function, improving the collection efficiency, reducing the labor intensity, reducing the eye fatigue of workers, and the mountain crutch can also be used as the support of workers to reduce the fatigue degree of workers climbing the mountain. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a structure schematic view of the Cordyceps location system of the present application;

[0036] Figure 2 It is a CA attention mechanism structure diagram of the present application;

[0037] Figure 3 It is a Cordyceps identification and positioning flow chart of the present application;

[0038] Figure 4 It is a Cordyceps reconstruction and positioning point cloud of the present application;

[0039] Figure 5 It is a working principle diagram of the Cordyceps location system of the present application.

[0040] In the figure: 1, mountain crutch; 2, support beam; 3, CCD camera; 4, controller; 5, battery; 6, camera support; 7, controller support frame; 8, battery support frame; 9, tablet computer. DETAILED DESCRIPTION

[0041] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0042] REFERENCE Figures 1 to 5 As shown in the figure, a Cordyceps location system based on binocular stereo and deep learning technology comprises a hardware structure and a software system, wherein the hardware structure comprises a mountain crutch 1, a support beam 2, a CCD camera 3, a controller 4, a battery 5, a camera support 6, a controller support frame 7, a battery support frame 8 and a tablet computer 9;

[0043] The software system comprises an image acquisition module, a training learning module, an identification module and an interaction module.

[0044] The image acquisition module acquires the image of the caterpillar fungus based on the CCD camera 3, and uploads the acquired image of the caterpillar fungus to the training learning module.

[0045] The training learning module trains the image data of the caterpillar fungus based on a deep learning framework, learns and optimizes the CA attention mechanism, adds the CA attention mechanism in front of SPPF in the BackBone of the yolov5s model, and constructs a caterpillar fungus identification model.

[0046] The identification module inputs the acquired image data based on the constructed caterpillar fungus identification model, reconstructs the point cloud of the caterpillar fungus based on the deep learning model, performs three-dimensional positioning, extracts the position of the caterpillar fungus from the color image, and judges and analyzes the position of the caterpillar fungus.

[0047] The interaction module performs human-computer interaction based on the tablet computer 9, serves as a handheld terminal of the caterpillar fungus collection worker, is deployed with a client software, receives the returned image of the caterpillar fungus in real time, instructs the caterpillar fungus collection worker to collect the caterpillar fungus, and emits an indication sound after the caterpillar fungus is identified.

[0048] The application utilizes a deep learning framework, can fully excavate deep features and rules contained in the image data of the caterpillar fungus, has powerful nonlinear modeling capability of deep learning, can automatically learn unique appearance features of the caterpillar fungus, such as the texture of the joint part of the worm body and the grass body, the overall shape contour and the difference from the surrounding environment, greatly improves the recognition accuracy and efficiency compared with traditional recognition methods, utilizes the constructed caterpillar fungus identification model, can not only identify the caterpillar fungus in the image, but also reconstructs the point cloud of the caterpillar fungus and performs three-dimensional positioning based on the powerful function of deep learning, which means that the position information of the caterpillar fungus in the actual space can be accurately obtained, and for the caterpillar fungus growing in the mountainous area with ups and downs, the deep grass and different terrain environments, the caterpillar fungus collection worker can more accurately find the specific position through three-dimensional positioning, and the positioning deviation problem that may occur when only relying on two-dimensional image judgment is avoided.

[0049] One end of the support beam 2 is a semicircular structure and is transitionally matched with the trekking stick 1, is fixed on the trekking stick 1 through a hexagonal bolt, the support beam 2 is provided with a CCD camera 3, a controller 4 and a battery 5, the battery 5 is used for power supply of the controller 4, the CCD camera 3 is connected with the controller 4 through a USB or a network cable, and the tablet computer 9 is manually held.

[0050] The half-round structure can well fit the shape of the trekking stick 1, reduces the gap between the two, avoids loosening and unstable shaking during use, and ensures reliable connection of the whole device when walking and climbing in complex terrain in the wild.

[0051] The CA attention mechanism structure firstly performs average pooling operation on the acquired image in horizontal and vertical directions, integrates the feature information in the horizontal and vertical directions, and performs convolution processing, and the specific steps are as follows:

[0052] Given an input layer X with a size of HxWxC, using a pooling kernel with a size of (H, 1) and (1, W), each channel is encoded along the horizontal and vertical directions, and the outputs of the cth channel with a height of H and the cth channel with a width of W are respectively:

[0053]

[0054] In the formula, H and W respectively represent the height and width of the feature layer, and are known, and respectively represent the output results of the cth channel in the H and W directions, x c (h,i) and x c (j,w) represent the input of the input feature X along the H direction and the W direction.

[0055] By respectively encoding each channel along the horizontal (H direction) and vertical (W direction), the application can focus and capture key information in the input feature layer from different dimensions, such as in image data, the input layer X is regarded as a feature representation related to the image, H and W correspond to the height and width of the image respectively, and C is the number of channels such as RGB channels. The horizontal pooling can extract the horizontal distribution features of the object in the image, such as the shape and texture continuity features of the main body of the insect herb along the horizontal direction in the insect herb image; the vertical pooling can capture the key features of the object in the vertical direction, such as the relative position and height features of the insect herb to the surrounding vegetation in the vertical direction. This multi-directional feature extraction helps to more accurately identify and analyze the object in the subsequent process, and excavates more discriminative features, laying a foundation for accurate classification and positioning tasks.

[0056] The two independent direction feature maps output by the above two groups of formulas are subjected to splicing operation, 1x1 convolution operation and activation operation, and the formula is:

[0057] f=σ(F1[Z h ,Z w ])

[0058] In the formula, f represents the intermediate feature map after encoding the spatial information in the horizontal and vertical directions, [.,.] represents the splicing operation of the spatial dimension, F1 represents the convolution operation, and δ represents the Sigmoid function.

[0059] The f is split along the spatial dimension, and then two 1x1 convolution operations are used to restore the channel number of the feature map to the same channel number as the input feature X. The attention weight of the feature map is calculated using the Sigmoid function, and the result is:

[0060] g h =σ(F h (f h ))

[0061] g w =σ(F w (f w ))

[0062] In the formula, g h and g w represent the attention weights of the convolution and activation after splitting f, f h and f w represent the feature maps along the H direction and the W direction after splitting f, and F h and F w represent the convolution operations.

[0063] The g h and g w are multiplied by the input feature map weight to output the feature map with coordinate attention weight, and the formula is:

[0064]

[0065] In the formula, y c (i,j) and x c (i,j) represent the input and output of the cth channel, respectively, and represent the attention weights along the H direction and the W direction on the cth channel, respectively.

[0066] The recognition module recognition method inputs the collected image data into the Cordyceps recognition model, and uses the bounding box box(x,y,width,height) to represent the position of Cordyceps, x is the value of the upper left point of the rectangular frame along the horizontal direction of the image coordinates, y is the value of the upper left point of the rectangular frame along the vertical direction of the image coordinates, width and height are the length and height of the rectangular frame along the horizontal direction and vertical direction of the image coordinates, respectively, and the bounding box box(x,y,width,height) is the rectangular frame output by the algorithm after recognition.

[0067] The application uses a bounding box to mark the location of Cordyceps, which can very intuitively show the specific area where Cordyceps is located on the image. For Cordyceps collection workers or related monitoring personnel, without a complex interpretation process, they can clearly see the approximate orientation of Cordyceps in the picture through the rectangular box on the image at a glance, and know where Cordyceps exists, so as to facilitate quick positioning and take corresponding collection or subsequent analysis actions, so that the position information of Cordyceps is clear at a glance, and the efficiency of information transmission is greatly improved.

[0068] Based on the image acquisition module, the depth map of Cordyceps is obtained, a rectangular box (x, y, width, height) is drawn, the region of interest of the depth map is extracted, based on the comprehensive rectangular box, an image processing algorithm is used to extract, and then a formula for single-point reconstruction of binocular vision is used to reconstruct the pixel points of the region of interest. The point cloud after reconstruction is the point cloud of Cordyceps, that is, the location of Cordyceps is found. The formula is:

[0069] x ccs =(u-cx)*zccs / f x

[0070] y ccs =(v-cy)*z ccs / f y

[0071] z ccs =fb / d,d=u l -u r

[0072] In the formula, (x ccs , y ccs , z ccs ) represents the three-dimensional point in the camera coordinate system after reconstruction, (u, v) is the image coordinate of the pixel in the depth map, f is the focal length of the camera, (f x , f y ) is the component of the focal length of the camera, b is the baseline distance of the camera, d is the parallax of the camera, u l , u r are the projection points of the reconstructed point in the left camera and the right camera respectively.

[0073] By using the single-point reconstruction formula of binocular vision, the application can convert the two-dimensional information of Cordyceps in the depth map into three-dimensional point cloud information in the camera coordinate system, realizing the leap from the plane to the three-dimensional space. This enables us to accurately know the position of Cordyceps in the actual three-dimensional space. Compared with positioning relying only on two-dimensional images, it can more accurately reflect the real orientation of Cordyceps, whether it is in the mountainous environment with ups and downs or in the different levels of depth position in the grass, it can accurately capture, effectively avoiding the positioning error caused by the lack of depth perception of two-dimensional images, and providing very accurate position guidance for Cordyceps collection subsequent work.

[0074] The tablet 9 is used as a carrier of the interaction module, and the collected cordyceps picture is transmitted to the client software on the tablet 9 end through a network transmission link, and the analysis result is directly displayed to show the position of the cordyceps.

[0075] The cordyceps picture collected by the network transmission link is transmitted to the client software on the tablet 9 end in time, so that the collector can obtain the latest cordyceps image information in the first time, whether the cordyceps image is collected by the camera or the unmanned aerial vehicle device installed in the surrounding area, and the image can be quickly displayed on the tablet 9 screen, which means that the collector can know the distribution of the cordyceps in each collection area in real time without waiting, reduces the time consumed for searching the cordyceps, and makes the collection work more time-efficient and targeted; the direct display of the cordyceps position can assist the collector to make positioning decisions quickly, in the vast cordyceps collection area, it is difficult to quickly and accurately find the cordyceps only by naked eyes in the face of complex terrain and various vegetation, and the cordyceps position information displayed on the tablet 9 can guide the collector to quickly move towards the target direction and accurately locate the cordyceps, so that blind search is avoided, the collection path is optimized, and the efficiency and accuracy of the cordyceps collection are further improved.

[0076] The working principle of the cordyceps positioning system is to use a mountain walking stick to position and collect cordyceps images, use a mainstream deep learning framework to train a cordyceps recognition model, identify and locate the cordyceps position, obtain a depth map cordyceps ROI based on a rectangular frame, reconstruct a cordyceps point cloud, and realize three-dimensional positioning.

[0077] The basic principle, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A worm location system based on binocular stereo and deep learning technology, characterized in that, Comprise: Hardware structure and software system, wherein the hardware structure comprises a mountain climbing stick (1), a support beam (2), a CCD camera (3), a controller (4), a battery (5), a camera support (6), a controller support (7), a battery support (8) and a tablet computer (9); The software system comprises an image acquisition module, a training learning module, an identification module and an interaction module; The image acquisition module acquires the image of the caterpillar fungus based on the CCD camera (3), and uploads the acquired image of the caterpillar fungus to the training learning module; The training learning module trains the image data of the caterpillar fungus based on a deep learning framework, learns and optimizes the CA attention mechanism, adds the CA attention mechanism in front of SPPF in the BackBone of the yolov5s model, and constructs a caterpillar fungus identification model; The identification module inputs the collected image data based on the constructed caterpillar fungus identification model, reconstructs the caterpillar fungus point cloud based on the deep learning model, performs three-dimensional positioning, extracts the position of the caterpillar fungus from the color image, and judges and analyzes the position of the caterpillar fungus; The interaction module performs human-computer interaction based on the tablet computer (9), serves as a handheld terminal of a caterpillar fungus collector, is deployed with a client software, receives the returned caterpillar fungus image in real time, instructs the caterpillar fungus collector to collect the caterpillar fungus, and emits an instruction sound after the caterpillar fungus is identified; The CA attention mechanism structure firstly performs average pooling operation on the horizontal and vertical directions of the acquired image, integrates the feature information of the horizontal and vertical directions, and performs convolution processing, and the specific steps are as follows: Given an input layer X with a size of HxWxC, use a pooling kernel with a size of (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions, and the outputs of the c channels with a height of H and the c channels with a width of W are respectively: , , where H and W represent the height and width of the feature layer, respectively, and and represent the output results of the c-th channel in the H and W directions, respectively, and represent the inputs of the input feature X along the H and W directions, respectively. The two independent direction feature maps output by the above two groups of formulas are spliced, and 1x1 convolution operation and activation operation are performed, and the formula is: In the formula, f represents an intermediate feature map after encoding spatial information in the horizontal direction and the vertical direction, represents a splicing operation in the spatial dimension, and F1 represents a convolution operation, represents a Sigmoid function; Split f along the spatial dimension, then use two 1x1 convolution operations to restore the number of channels of the feature map to the same number of channels as the input feature X, calculate the attention weight of the feature map using the Sigmoid function, and the result is: , In the formula, and denote the attention weight of the f-splitting post-convolution and activation, and denote the feature map along the H direction and the W direction respectively after f-splitting, and denote the convolution operation; Based on the image acquisition module, the depth map of the worm grass is obtained, and a rectangle is drawn The region of interest of the depth map is extracted, a comprehensive rectangular frame is adopted, an image processing algorithm is used for extraction, then the pixel points of the region of interest are reconstructed by using the formula of binocular vision single point reconstruction, and the point cloud after reconstruction is the point cloud of the worm grass, that is, the worm grass is located, and the formula is: , , , in which formula (1) represents a three-dimensional point in the camera coordinate system after reconstruction, is the image coordinate of the pixel in the depth map, f is the focal length of the camera, (f ) is the component of the camera focal length, b is the baseline distance of the camera, and d is the parallax of the camera, are the projection points of the reconstructed point in the left camera and the right camera, respectively.​ 2. The Cordyceps locating system based on binocular stereo and deep learning technology according to claim 1, characterized in that, One end of the support beam (2) is a semicircular structure and is transitionally matched with the mountain climbing stick (1), is fixed on the mountain climbing stick (1) through an inner hexagonal bolt, the support beam (2) is provided with the CCD camera (3), the controller (4) and the battery (5), the battery (5) is used for power supply for the controller (4), the CCD camera (3) is connected with the controller (4) through a USB or a network cable, and the tablet computer (9) is manually held. 3.The wormgrass locating system based on binocular stereo and deep learning technology according to claim 1, characterized in that, will be described below with reference to the accompanying drawings. and The input feature map is multiplied by the weight to output the feature map with coordinate attention weight, and the formula is represented as: wherein, and denote the input and output of the c-th channel, respectively, and denote the attention weights in the H and W direction on the c-th channel, respectively.

4. The system according to claim 1, wherein the system is characterized by, The recognition module recognition method inputs the collected image data into a cordyceps recognition model, uses a bounding box to represent the position of the cordyceps, wherein x is the value of the left upper point of the rectangular box along the horizontal direction of the image coordinates, and y is the value of the left upper point of the rectangular box along the vertical direction of the image coordinates, respectively the length and the height of the rectangular box along the horizontal direction and the vertical direction of the image coordinates, and the bounding box is the rectangular box output after the algorithm recognition.

5. The system according to claim 1, wherein the system is characterized by, The tablet computer (9) is used as a carrier of the interaction module, transmits the collected caterpillar fungus picture to the client software on the tablet computer (9) through a network transmission link, and visually displays the analysis result to show the position of the caterpillar fungus.

Images