Tea garden pest hazard identification method and equipment based on convolutional neural network

Abstract

The invention discloses a tea garden pest hazard identification method and equipment based on a convolutional neural network, and relates to the field of image identification, and the method comprises the steps: collecting first image data including tea trees based on an image collection device preset in a tea garden; based on the first recognition result of the convolutional neural network, adjusting the shooting range of the image acquisition device by controlling a holder and a focal length corresponding to the image acquisition device, and acquiring second image data through the adjusted image acquisition device; and preprocessing the second image data, inputting the preprocessed second image data into a pre-trained convolutional neural network, and outputting a second recognition result. Different image acquisition devices provide different visual angles and image information, and scene understanding is enhanced, so that the recognition difficulty caused by visual angle limitation is reduced, the recognition accuracy is improved, and the probability of false alarms is reduced.

Description

Technical Field

[0001] The present application relates to the field of image recognition, and specifically to a method and device for identifying tea garden pest damage based on a convolutional neural network. Background Art

[0002] With the development of technology, people can now use automated image recognition to assist agricultural production. For example, they can use image recognition to identify pests during the growth of crops.

[0003] Taking tea as an example, during the growth process of the tea tree, images of the tea tree are collected and automatically identified and analyzed to determine whether there is any pest damage in the tea tree.

[0004] However, in the traditional image recognition process, which relies on the recognition accuracy of the algorithm and the clarity of the image, it is easy to have missed detections and false detections. Summary of the invention

[0005] In order to solve the above problems, this application proposes a tea garden pest damage identification method based on convolutional neural network, including:

[0006] Based on an image acquisition device pre-set in the tea garden, first image data including tea trees is acquired;

[0007] Preprocessing the first image data, and inputting the preprocessed first image data into a pre-trained convolutional neural network;

[0008] Based on the first recognition result of the convolutional neural network, the shooting range of the image acquisition device is adjusted by controlling the pan / tilt and focal length corresponding to the image acquisition device, and the second image data is acquired by the adjusted image acquisition device;

[0009] The second image data is preprocessed, and the preprocessed second image data is input into a pre-trained convolutional neural network to output a second recognition result.

[0010] On the other hand, the present application also proposes a tea garden pest damage identification device based on a convolutional neural network, comprising:

[0011] at least one processor; and,

[0012] a memory communicatively connected to the at least one processor; wherein,

[0013] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute: the tea garden pest damage identification method based on convolutional neural network described in the above example.

[0014] On the other hand, the present application also proposes a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured as the tea garden pest damage identification method based on a convolutional neural network as described in the above example.

[0015] The tea garden pest damage identification method based on convolutional neural network proposed in this application can bring the following beneficial effects:

[0016] For pests, which are usually small targets, multiple rounds of multi-position image recognition, different image acquisition devices provide different perspectives and image information to enhance scene understanding, thereby helping to reduce the difficulty of recognition due to perspective limitations, improve recognition accuracy, and reduce the limitations of a single image acquisition device, reducing the probability of false alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 Schematic diagram of the process of identifying tea garden pest damage based on convolutional neural network in the embodiment of the present application;

[0019] Figure 2 This is a schematic diagram of a tea garden pest damage identification device based on a convolutional neural network in an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0021] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, the embodiment of the present application provides a method for identifying tea garden pest damage based on a convolutional neural network, comprising:

[0023] S101: Based on an image acquisition device pre-installed in a tea garden, first image data including tea trees is acquired.

[0024] The image acquisition device is usually a camera installed in the tea garden, and the number of the cameras is usually multiple. The camera is installed through a corresponding support rod so that its height can be higher than the height of the tea tree, and a pan / tilt is installed on the camera to enable the camera to rotate. Of course, the camera can also have some preliminary data processing capabilities (such as automatic zoom, image preprocessing, etc.), so that the first image data collected is clearer and more accurate.

[0025] Since tea trees in a tea garden are usually arranged in rows, in order to make the image acquisition device more reasonably arranged and to be able to more comprehensively acquire images of tea trees in the entire tea garden, the arrangement position of the image acquisition device in the tea garden can be planned and arranged in advance. In order to facilitate data processing, the tea trees included in the first image data are usually arranged in rows rather than including a single tea tree.

[0026] Specifically, an image acquisition device is set for each row of tea trees in the tea garden.

[0027] The setup process includes:

[0028] In each row of tea trees, an initial row of tea trees is selected, and in the initial row of tea trees, a corresponding initial position is selected. Generally speaking, for smaller tea gardens, a single row of tea trees with the highest height is generally selected as the initial row of tea trees. Since smaller tea gardens do not need to set up image acquisition devices in too many rows, cameras may only be set up in one or two rows. At this time, if other positions are used, it is easy to fail to capture the image of the top tea tree. For larger tea gardens, settings can be made based on actual conditions. For example, if the position of the highest row of tea trees is not convenient for setting up an image acquisition device, you can also choose to set it at other positions such as the center height as the initial row of tea trees.

[0029] Of course, if the terrain in a tea garden is undulating, the tea garden can be divided into multiple areas so that the terrain of the tea trees in each area is relatively uniform. For example, a single hilltop area or a flat area can be divided into independent areas, and an image acquisition device can be independently set in each area. At this time, the most central, or the highest or the single row of tea trees at the center of the height can be selected as the initial row of tea trees. The initial position can be set randomly.

[0030] When the image acquisition device is set up, it is usually set inside the tea tree and supported by a vertical rod so that the camera is higher than the tea tree, thereby not affecting people's progress in the aisle.

[0031] In the initial row of tea trees, an image acquisition device is set at the initial position, and traversal is performed in the initial row of tea trees starting from the initial position, that is, moving along the setting direction of the initial row of tea trees (either clockwise or counterclockwise). Whenever a new position appears, the distance between the new position and the position of the last set image acquisition device reaches a preset distance, or the corner corresponding to the new position is greater than a preset angle, that is to say, the image acquisition device is set at the new position until the initial row of tea trees is traversed.

[0032] When the distance between the new image acquisition device and the previous image acquisition device is too far, it may affect the accuracy of image acquisition. Therefore, a preset distance is set. The preset distance is set according to the shooting accuracy of the image acquisition device, and is usually the maximum shooting distance under the premise of ensuring that the clarity of the image captured by the image acquisition device meets expectations.

[0033] Since tea trees are usually arranged in rows, when the corners of a single row of tea trees are too large, it often means that the current terrain is relatively close to the edge. Therefore, a new image acquisition device is set up. By adjusting its shooting angle, tea trees in various ranges can be more comprehensively captured.

[0034] After the image acquisition device is set in the initial row of tea trees, it can be set in other rows of tea trees in the same way. Of course, in order to save costs, it is not necessary to set the image acquisition device in each row of tea trees.

[0035] Starting from the initial row of tea trees, traverse all single rows of tea trees. Whenever a new single row of tea trees appears, the row number interval between the new single row of tea trees and the single row of tea trees of the last image acquisition device that has been set reaches a first preset number, or the height difference between the new single row of tea trees and the single row of tea trees of the last image acquisition device that has been set (the height difference can be any one of the maximum height difference, the minimum height difference or the average height difference) is greater than the preset difference, then in the new single row of tea trees, the image acquisition device is set (the same as the setting in the initial row of tea trees) until all rows of tea trees are traversed.

[0036] Generally speaking, the distance between each row of tea trees in each tea garden is relatively fixed, so when the row number interval is far apart, an image acquisition device can be set in the row. Of course, if the distance between each row of tea trees is large, the row number interval can also be changed to the distance interval.

[0037] In actual situations, the tea garden may be planted on a hilltop, so there will be a height difference between different rows. When the height difference is high, incomplete images may occur easily. Therefore, it is determined whether to set up an image acquisition device based on the height difference.

[0038] Further, after completing the above settings, in the single row of tea trees, for each image acquisition device, the initial state of the corresponding pan / tilt and focal length is set. The shooting directions of the image acquisition devices of the single row of tea trees are consistent, for example, they are all set clockwise along the single row of tea trees.

[0039] In addition, to ensure the acquisition range, within the shooting range of each image acquisition device in a single row of tea trees, at least one other image acquisition device belonging to the same single row of tea trees as the image acquisition device and at least one other image acquisition device belonging to a different single row of tea trees as the image acquisition device can be captured. This ensures the size of the initial acquisition range of each image acquisition device and the setting density of the image acquisition devices. Of course, this is a more preferred setting, and in fact, it is also possible to set this condition based on demand, thereby reducing the setting density of the image acquisition devices and reducing costs.

[0040] Based on this, when collecting the first image data, if it is only collected through a fixed camera position, it is easy to have incomplete image collection, or only the surface of the tea tree can be collected, and it is difficult to collect the situation in other directions.

[0041] Therefore, in the actual first image data acquisition process, every time a preset time acquisition point (for example, a fixed time every day, or a fixed time every other day) is reached, image acquisition is performed based on the image acquisition device.

[0042] During the image acquisition process, the image acquisition device is controlled by the pan-tilt head to acquire images at multiple angles, and the acquired image data are spliced ​​to obtain first image data including the tea tree, thereby enabling multi-directional and multi-angle image data acquisition of the tea tree.

[0043] Among them, for the image acquisition device set with a corner greater than the preset angle, due to the terrain, there is often only a large number of tea trees on one side, and the other side can only take pictures of the top of the tea trees. Therefore, the required pan-tilt rotation angle range can be set to [0, a / 2] or [-a / 2, 0], where a is the actual angle of the corner. The specific setting direction is set based on the actual situation, and is mainly used to collect image data of the bottom of other rows of tea trees on the side with higher terrain. For the remaining image acquisition devices, it can be defaulted to a corner close to 180 degrees, so the pan-tilt rotation angle range is [0°, 90°] or [-90°, 0°].

[0044] S102: Preprocess the first image data, and input the preprocessed first image data into a pre-trained convolutional neural network.

[0045] The preprocessing of image data may include: image enhancement, grayscale, normalization, adaptive histogram equalization, etc. The architecture of convolutional neural network may include: input layer, residual block, attention module, fully connected layer, output layer.

[0046] Among them, the input image size can be unified in the input layer. In order to ensure that the details can be captured, the size of the input image can be set to a higher resolution, which can retain the small-scale features of tea leaves and pests.

[0047] Residual blocks are connected to the input layer. The residual blocks contain multiple convolutional layers, and the convolutional layers are connected by residual connections. The convolutional layers are connected by pooling layers. In order to cope with the different backgrounds and possible blur of tea leaves, the residual connection can ensure that the flow of information is not lost. In this way, the initial features can be retained even after multiple convolution operations.

[0048] The convolutional layers include convolution kernels of various scales and use depth-wise separable convolution. The convolutional layers include Inception modules, which perform parallel calculations on convolution kernels of various scales.

[0049] In order to capture features of different scales, the convolution kernel of the convolution layer can be set to multiple sizes, such as 3x3, 5x5, and 7x7. Small-scale convolution kernels help capture detailed features (such as the shape and texture of pests), while large-scale convolution kernels help capture features in a larger range (such as the overall position of pests on leaves).

[0050] Using the Inception module in the convolutional layer, convolution kernels of different sizes can be calculated in parallel in the same layer to extract multi-scale features. This supports capturing pests of different sizes and their locations on the leaves.

[0051] For efficient feature extraction, Depthwise Separable Convolutions are used to reduce the amount of calculation while retaining important features, which is particularly effective for capturing details.

[0052] Convolutional layers are used in conjunction with pooling layers (for example, the Max Pooling layer) to help reduce the size of the input image and extract more stable features. Pooling operations help increase the model's robustness to position changes.

[0053] The attention module is connected to the residual block, and the attention module is a self-attention mechanism. For example, by setting the SE module (Squeeze-and-Excitation Block), by introducing the Squeeze-and-Excitation module, the model can learn which channels are more important for feature recognition of pests at different levels. The SE module emphasizes the area with pests by adaptively adjusting the weight of the channel, thereby improving the detection accuracy.

[0054] The fully connected layer is a classification layer. Through multiple fully connected layers, the network can perform the final pest classification based on the extracted features. The Softmax activation function can be used to map the output to the probability of each pest category.

[0055] Output layer: It can be a multi-classification Softmax layer, which is used to identify the types of pests on tea leaves, or output the category of "no pests", or the probability of the presence of pests, etc.

[0056] Furthermore, the training process of the convolutional neural network includes:

[0057] Collect training samples and determine positive training samples; wherein the training samples include image data of tea trees in different shooting conditions, and the shooting conditions include: shooting time, shooting angle, and weather. For example, the shooting state label of a training sample may include: 8:00am, front shooting, and windy. Positive training samples are training samples with pests.

[0058] For each shooting state of the positive training sample, the first probability of the tea tree being identified as a pest by the pre-trained model (the model is an untrained model that can be directly obtained from the corresponding platform or system and has basic recognition capabilities) is counted. This first probability can be obtained through statistics when annotating the sample, or it can be annotated in advance through expert experience combined with actual conditions. Generally speaking, the probability of the occurrence of pests and diseases is different for tea trees in different seasons, different locations, and different ages, but it may not be reflected in the image. For example, an image of a tea tree taken from the top can identify pests, but the front shot cannot identify pests. At this time, for the positive training samples with pests, only the probability of the pests being identified due to the shooting state of the image is collected.

[0059] At this time, the second probability of the training sample being identified as a pest is calculated based on the first probability and the shooting state corresponding to each training sample. For example, the shooting state labels of a training sample may include: 8:00am, front shooting, windy, and after statistics, the first probabilities of each shooting state being identified as a pest are: a, b, c, respectively, and the second probability d = 1-(1-a)*(1-b)*(1-c).

[0060] At this time, the hit probability of the target event is obtained according to the second probability (the second probability is negatively correlated with the hit probability, for example, the second probability is inversely calculated to obtain the hit probability. Of course, the corresponding coefficient can be set based on the needs), and whether each training sample hits the target event is determined according to the hit probability. The higher the second probability, the lower the probability of hitting the target event. At this time, if the target event is hit, the corresponding training sample is data enhanced.

[0061] Data augmentation methods can include random rotation, cropping, flipping, etc. In order to cope with the deformation of tea leaves in different environments and changes in lighting, data augmentation methods can be used during training to increase the generalization ability of the model.

[0062] For the situation where there are pests but they are not identified due to the shooting status, by setting the target event, the positive training samples of the shooting status with a lower probability of identifying the pests (second probability) are set with a corresponding higher hit probability, so that the positive training samples can be more understood and identified by the model, increasing their probability of being identified.

[0063] Furthermore, corresponding data enhancement methods are set in advance for different shooting conditions. For example, if the shooting conditions are: 8:00am, front shooting, windy, then "8:00am" may cause recognition problems due to lighting. At this time, the data enhancement method can be to change the image contrast, color, etc., and for "windy", the recognition problem may be caused by wind, so the data enhancement method is rotation.

[0064] Based on this, for the corresponding training sample, the corresponding shooting state is determined; according to the shooting state, the corresponding data enhancement method is selected to perform data enhancement on the corresponding training sample. Of course, the data enhancement method can be one or more, and some shooting states do not have a strict corresponding data enhancement method, and a default or random data enhancement method can be set for them.

[0065] S103: Based on the first recognition result of the convolutional neural network, the shooting range of the image acquisition device is adjusted by controlling the gimbal and focal length corresponding to the image acquisition device, and the second image data is acquired by the adjusted image acquisition device.

[0066] Specifically, for each image acquisition device, a first recognition result thereof in a convolutional neural network is determined, and the first recognition result includes the probability of whether a pest is recognized, and image coordinate information of the pest in the image.

[0067] At this time, for all first recognition results, first position coordinate information in which the probability of pest damage exceeds the first preset probability and second position coordinate information in which the probability of pest damage exceeds the second preset probability and is lower than the first preset probability is determined. The first position coordinate information and the second position coordinate information here refer to the position coordinates in reality.

[0068] For a single image acquisition device, the camera is calibrated in advance to obtain the corresponding internal parameters (including focal length, principal point (optical center), distortion coefficient, etc.) and external parameters (including the position and orientation of the camera).

[0069] At this time, for all first recognition results, first image coordinate information in which the probability of pest damage exceeds the first preset probability and second image coordinate information in which the probability of pest damage exceeds the second preset probability and is lower than the first preset probability is determined. The first image coordinate information and the second image coordinate information here refer to the position coordinates in the image.

[0070] The first image coordinate information and the second image coordinate information are respectively converted into first position coordinate information and second position coordinate information in reality through perspective transformation, for example, through projection equation.

[0071] At this time, after obtaining the actual position coordinate information, it is considered that there is a high probability of pests in the first position coordinate information. If the same position coordinate information is determined to belong to the second position coordinate information (that is, there is a certain probability of pests) in more than a second preset number (for example, 2 or 3) of image acquisition devices, then the position coordinate information is changed to the first position coordinate information, and it is considered that there is a high probability of pests.

[0072] For each first position coordinate information (including the first position information converted from the second position information), determine the third preset number (for example, 3) of image acquisition devices closest to it, and adjust the shooting range of the third preset number of image acquisition devices by controlling the pan / tilt and focal length corresponding to the third preset number of image acquisition devices, so as to further shoot the position where the pests may be identified. At this time, multiple image acquisition devices are used to capture images of the same position again, so as to prevent false alarms.

[0073] The second image data is acquired by the adjusted image acquisition device, and in the second image data, the image area of ​​the first position coordinate information accounts for a higher proportion than a preset proportion. For example, by adjusting the focal length, the first position coordinate information is focused on, thereby reducing the interference of surrounding noise on the model.

[0074] S104: Preprocess the second image data, input the preprocessed second image data into a pre-trained convolutional neural network, and output a second recognition result.

[0075] Similar to the first image data, the second recognition result is output through the convolutional neural network. If the probability of the pest appearing is still considered to exceed the first preset probability in the second recognition result, it is considered that there are pests at the location and reported. When there are multiple second recognition results, it can be determined by proportion. If more than a certain proportion of the second recognition results are considered to have a probability of the pest appearing that exceeds the first preset probability, it is considered that there are pests at the location and reported.

[0076] For pests, which are usually small targets, multiple rounds of multi-position image recognition, different image acquisition devices provide different perspectives and image information to enhance scene understanding, thereby helping to reduce the difficulty of recognition due to perspective limitations, improve recognition accuracy, and reduce the limitations of a single image acquisition device, reducing the probability of false alarms.

[0077] like Figure 2 As shown, the embodiment of the present application also proposes a tea garden pest damage identification device based on a convolutional neural network, comprising:

[0078] at least one processor; and,

[0079] a memory communicatively connected to the at least one processor; wherein,

[0080] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute: the tea garden pest damage identification method based on convolutional neural network described in any of the above embodiments.

[0081] The embodiment of the present application also proposes a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured as the tea garden pest damage identification method based on convolutional neural network described in any of the above embodiments.

[0082] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0083] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0084] The above is only the embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the present application.

Claims

1. A method for identifying tea garden pests based on convolutional neural networks, characterized in that: include: Based on an image acquisition device pre-set in the tea garden, first image data including tea trees is acquired; Preprocessing the first image data, and inputting the preprocessed first image data into a pre-trained convolutional neural network; Based on the first recognition result of the convolutional neural network, the shooting range of the image acquisition device is adjusted by controlling the pan / tilt and focal length corresponding to the image acquisition device, and the second image data is acquired by the adjusted image acquisition device; The second image data is preprocessed, and the preprocessed second image data is input into a pre-trained convolutional neural network to output a second recognition result.

2. The method according to claim 1, characterized in that: Before acquiring the first image data including the tea trees based on the image acquisition device pre-set in the tea garden, the method further includes: For each row of tea trees in the tea garden, set up the image acquisition device; The setup process includes: In each row of tea trees, an initial row of tea trees is selected, and in the initial row of tea trees, a corresponding initial position is selected; In the initial row of tea trees, an image acquisition device is set at the initial position, and traversing the initial row of tea trees is performed with the initial position as the starting point. Whenever a new position appears, the distance between the new position and the position of the last set image acquisition device reaches a preset distance, or the corner corresponding to the new position is greater than a preset angle, an image acquisition device is set at the new position until the initial row of tea trees is completely traversed; Taking the initial row of tea trees as the starting point, traversing through all single rows of tea trees, whenever a new single row of tea trees appears, the row number interval between the new single row of tea trees and the single row of tea trees in the last image acquisition device that has been set reaches a first preset number, or the height difference between the new single row of tea trees and the single row of tea trees in the last image acquisition device that has been set is greater than a preset difference, then the image acquisition device is set in the new single row of tea trees until all rows of tea trees are traversed.

3. The method according to claim 2, characterized in that Before acquiring the first image data including the tea trees based on the image acquisition device pre-set in the tea garden, the method further includes: In a single row of tea trees, for each image acquisition device, set the initial state of its corresponding pan / tilt and focal length; Among them, the shooting direction of the image acquisition device of a single row of tea trees is consistent; Within the shooting range of each image acquisition device in a single row of tea trees, at least one other image acquisition device belonging to the same single row of tea trees as the image acquisition device and at least one other image acquisition device belonging to a different single row of tea trees as the image acquisition device can be captured.

4. The method according to claim 3, characterized in that: Based on the first recognition result of the convolutional neural network, the shooting range of the image acquisition device is adjusted by controlling the pan / tilt and focal length corresponding to the image acquisition device, and the second image data is acquired by the adjusted image acquisition device, specifically including: For each image acquisition device, determining a first recognition result thereof in the convolutional neural network; For all first recognition results, determining first position coordinate information in which the probability of pest damage exceeding a first preset probability exists, and second position coordinate information in which the probability of pest damage exceeding a second preset probability and lower than the first preset probability exists; If the same position coordinate information is determined to belong to the second position coordinate information in image acquisition devices exceeding a second preset number, the position coordinate information is changed to the first position coordinate information; For each of the first position coordinate information, determine a third preset number of image acquisition devices closest to it, and adjust the shooting range of the third preset number of image acquisition devices by controlling the pan / tilts and focal lengths corresponding to the third preset number of image acquisition devices; The second image data is acquired by the adjusted image acquisition device, and in the second image data, the proportion of the image area of ​​the first position coordinate information is higher than the preset proportion.

5. The method according to claim 4, characterized in that For all first recognition results, determining first position coordinate information in which the probability of pest damage exceeding a first preset probability and second position coordinate information in which the probability of pest damage exceeding a second preset probability and lower than the first preset probability specifically includes: For each image acquisition device, perform camera calibration to obtain the corresponding intrinsic and extrinsic parameters; For all first recognition results, determining first image coordinate information in which the probability of pest damage being present exceeds a first preset probability, and second image coordinate information in which the probability of pest damage being present exceeds a second preset probability and is lower than the first preset probability; The first image coordinate information and the second image coordinate information are respectively converted into first position coordinate information and second position coordinate information in reality through perspective transformation.

6. The method according to claim 2, characterized in that Based on the image acquisition device pre-set in the tea garden, first image data including tea trees is acquired, specifically including: Whenever a preset time acquisition point is reached, image acquisition is performed based on the image acquisition device; During the image acquisition process, the image acquisition device is controlled by the pan / tilt platform to acquire images at multiple angles, and the acquired image data are spliced ​​to obtain first image data including the tea tree; Among them, for the image acquisition device set with a corner greater than the preset angle, the pan-tilt rotation angle range is [0, a / 2] or [-a / 2, 0], and for the remaining image acquisition devices, the pan-tilt rotation angle range is [0°, 90°] or [-90°, 0°], where a is the actual angle of the corner.

7. The method according to claim 1, characterized in that The architecture of the convolutional neural network includes: Input layer, residual block, attention module, fully connected layer, output layer; The residual block is connected to the input layer, the residual block includes a plurality of convolutional layers, the plurality of convolutional layers are residually connected, and a pooling layer is connected after the convolutional layer; The convolution layer includes convolution kernels of various scales and uses depth-separable convolution; the convolution layer includes an Inception module, through which the convolution kernels of various scales are calculated in parallel; The attention module is connected to the residual block, and the attention module is a self-attention mechanism.

8. The method according to claim 1, characterized in that The training process of the convolutional neural network includes: Collect training samples and determine positive training samples therein; wherein the training samples include image data of tea trees under different shooting conditions, and the shooting conditions include: shooting time, shooting angle, and weather; and the positive training samples are training samples with pests; For each shooting state of the forward training sample, a first probability of the tea tree being identified as a pest by the pre-training model is counted; Calculate a second probability that the training sample is identified as a pest based on the first probability and the shooting state corresponding to each training sample; Obtaining a hit probability of a target event according to the second probability, and determining whether each training sample hits the target event according to the hit probability; the second probability is negatively correlated with the hit probability; If the target event is hit, data enhancement is performed on the corresponding training sample.

9. The method according to claim 8, characterized in that If the target event is hit, data enhancement is performed on the corresponding training sample, including: For the corresponding training sample, determine the corresponding shooting state; According to the shooting state, a corresponding data enhancement method is selected to perform data enhancement on the corresponding training sample.

10. A tea garden pest damage identification device based on convolutional neural network, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute: the tea garden pest damage identification method based on convolutional neural network as described in any one of claims 1 to 9.

Images