Intelligent control method of vegetable planting machine based on neural network

Abstract

The present invention relates to the technical field of intelligent control of planters, and in particular to an intelligent control method for vegetable planters based on neural networks. The method obtains a single vegetable image at an abnormal position according to abnormal sensor information. A variational autoencoder network is trained using a single vegetable image as a training image, and the training image is divided into an occluded part image and an unoccluded part image. Different reconstruction models are selected according to the occlusion degree category of the occluded part, and the training of the variational autoencoder network is completed through continuous iterative updates. A reconstructed vegetable image of a single vegetable image is reconstructed according to the variational autoencoder network. The parameters of the vegetable planter are adjusted by the offset information between the reconstructed vegetable image and the single vegetable image. The embodiment of the present invention uses image information to analyze the growth of vegetables, and uses the offset information between the reconstructed vegetable image of normal growth and the real image to control the parameters of the vegetable planter, so as to obtain vegetables with uniform growth and high quality.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of planters, and in particular to an intelligent control method for vegetable planters based on a neural network. Background Art

[0002] Smart vegetable planting machines are a form of facility agriculture. They simplify many steps of traditional planting methods and meet the needs of some modern urban residents to grow vegetables independently. In addition, modern smart planting machines are equipped with multiple electronic devices to complete functions such as automatic water replenishment, liquid replenishment, light replenishment, and atomization, and are connected to user mobile phones, computers and other electronic devices to achieve intelligent data display and intelligent automatic control. The sensors set in the smart planting machine obtain various status information of vegetables, and transmit the sensor information feedback to the background terminal to perform intelligent functions such as growth status monitoring, soil environment and light index collection and analysis.

[0003] The intelligent vegetable planting machine can adjust the light intensity by controlling the light cycle and light quality ratio, and use the Internet of Things to obtain growth data and environmental data during the vegetable cultivation process, and realize automatic irrigation, fertilization and spraying, light compensation, etc. through data analysis. However, the relevant control is based on the changes in the data itself, without considering the growth of the vegetables themselves, which will also affect the final vegetable harvest. Only analyzing the data cannot guarantee the uniform quality of the finished vegetable products. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide an intelligent control method for a vegetable planter based on a neural network. The technical solution adopted is as follows:

[0005] The present invention proposes an intelligent control method for a vegetable planting machine based on a neural network, the method comprising:

[0006] Acquire sensor information of the vegetable planting machine; determine the position of abnormal vegetables according to the sensor information; collect a single vegetable image of each vegetable at the position of the abnormal vegetable;

[0007] The single vegetable image is input into a pre-trained variational autoencoder network to obtain a reconstructed vegetable image; the variational autoencoder network uses the single vegetable image as a training image, and divides the training image into an occluded partial image and an unoccluded partial image according to the symmetry of the vegetable; the occlusion degree category of the training image is determined according to the occluded partial image; the occlusion degree category includes complete occlusion and partial occlusion; if the occluded partial image is the partial occlusion, the latent variable probability distribution model of the occluded partial image is used as a reconstruction model, and the model difference between the reconstruction model and the latent variable probability distribution model of the unoccluded partial image is obtained, and the reconstruction model is reconstructed according to the occluded partial image. The reconstruction model is updated based on the occlusion area of ​​the image and the model difference; if the occluded part is completely occluded, the latent variable probability distribution model of the unoccluded part is used as the reconstruction model, the occlusion area of ​​the occluded part image is obtained according to the reconstructed vegetable image, and the reconstruction model is updated based on the model difference and the occlusion area between the latent variable probability distribution model of the occluded part image at the corresponding position of the reconstructed vegetable image and the reconstruction model; the variational autoencoder network uses the KL divergence between the reconstruction model and the standard Gaussian distribution as the first loss function, and the difference between the reconstructed vegetable image and the training image as the second loss function;

[0008] The center point offset information of the center point of the vegetable area is obtained according to the coordinate information of the center point of the reconstructed vegetable area in the reconstructed vegetable image and the center point of the vegetable area in the single vegetable image; and the parameters of the vegetable planting machine are controlled according to the center point offset information.

[0009] Further, determining the location of abnormal vegetables according to the sensor information includes:

[0010] Acquire the sensor information containing multiple dimensional information; acquire the difference index between the sensor information of each dimension and the standard sensor information of the corresponding dimension; if the difference index is greater than a preset difference threshold, the corresponding sensor information is considered to be abnormal sensor information, and the corresponding abnormal vegetable position is obtained according to the abnormal sensor information.

[0011] Furthermore, dividing the training image into an occluded portion image and an unoccluded portion image according to the symmetry of the vegetable comprises:

[0012] The variational autoencoder network includes a semantic segmentation subnetwork, which inputs the training image into the semantic segmentation subnetwork and outputs the occluded portion image, the unoccluded portion image and the occluded region on the occluded portion image; if the occluded portion image is completely occluded, the occluded region information does not exist in the semantic segmentation result of the semantic segmentation subnetwork.

[0013] Further, updating the reconstructed model according to the occluded area of ​​the occluded partial image and the model difference comprises:

[0014] The ratio of the reconstructed model to the latent variable probability distribution model of the unblocked portion of the image is used as the model difference, and a model adjustment parameter is obtained according to the model difference and the occlusion area. The model adjustment parameter includes:

[0015] α=e s+β

[0016] Wherein, α is the model adjustment parameter, s is the occlusion area, and β is the model difference;

[0017] Updating the reconstructed model according to the model adjustment parameters includes:

[0018]

[0019] Among them, z1 is the reconstruction model, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, and α is the model adjustment parameter.

[0020] Furthermore, the using the KL divergence between the reconstructed model and the standard Gaussian distribution as the first loss function includes:

[0021]

[0022] Among them, L1 is the first loss function, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, α is the model adjustment parameter, and KL[] is the KL divergence calculation function.

[0023] Furthermore, taking the difference between the reconstructed vegetable image and the training image as the second loss function comprises:

[0024]

[0025] Wherein, L2 is the second loss function, x is the training image, is the reconstructed vegetable image, and ||||2 is the Euclidean distance calculation function.

[0026] Furthermore, the obtaining of the center point offset information of the center point of the vegetable region according to the coordinate information of the center point of the reconstructed vegetable region in the reconstructed vegetable image and the center point of the vegetable region in the single vegetable image also includes:

[0027] Obtain the reconstructed vegetable edge gradient amplitude in the reconstructed vegetable image and the single vegetable edge gradient amplitude in the single vegetable image; obtain the deviation rate between the reconstructed vegetable image and the single vegetable image according to the overlap rate formula, and the deviation rate formula includes:

[0028]

[0029] in, is the deviation rate, (a x ,a y ) is the coordinate of the center point of the reconstructed vegetable area, (b x ,b y ) is the coordinate of the center point of the vegetable area, n is the number of reconstructed vegetable edge pixels, m is the number of single vegetable edge pixels, g i is the reconstructed vegetable edge gradient amplitude of the i-th reconstructed vegetable edge pixel point, g j is the single vegetable edge gradient amplitude of the j-th single vegetable edge pixel point;

[0030] If the deviation rate is greater than a preset deviation rate threshold, the center point offset information is obtained.

[0031] Further, the acquiring the center point offset information includes:

[0032] The offset information includes an offset direction and an offset distance; the direction from the center point of the reconstructed vegetable area to the center point of the vegetable area is used as the offset direction, and the distance from the center point of the reconstructed vegetable area to the center point of the vegetable area is used as the offset distance.

[0033] Further, the parameters of the vegetable planting machine controlled according to the center point offset information include:

[0034] The opposite direction of the offset direction is used as the light compensation direction of the vegetable planter; the offset distance and the light adjustment amplitude are fitted according to historical data, and the light adjustment amplitude of the vegetable planter is controlled according to the fitting relationship.

[0035] The present invention has the following beneficial effects:

[0036] 1. The embodiment of the present invention obtains the image information of the vegetables at the abnormal data through computer vision, and reflects the growth information of the vegetables based on the image information. Because the light received by the blocked position of the vegetables is weak, the single vegetable image is reconstructed using the variational autoencoder network to obtain the reconstructed vegetable image of normal growth. According to the overlapping analysis of the reconstructed vegetable image and the single vegetable image, the parameters required to adjust the vegetable planter are obtained, and the vegetable planter is intelligently controlled according to the sensor information and the vegetable growth information.

[0037] 2. In the process of training the variational autoencoder network, the embodiment of the present invention divides the training image into an occluded part and an unoccluded part. The occlusion degree category is obtained according to the morphological characteristics of the occluded part, and the variational autoencoder network performs different reconstruction methods according to different occlusion degree categories. The generalization ability of the network is improved, the network reconstruction ability is stronger, and the scope of application is wider. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A flow chart of a neural network-based intelligent control method for a vegetable planter provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a vegetable planter intelligent control method based on a neural network proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0041] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0042] The specific scheme of the intelligent control method of a vegetable planter based on neural network provided by the present invention is described in detail below with reference to the accompanying drawings.

[0043] See also Figure 1 , which shows a flow chart of a vegetable planting machine intelligent control method based on a neural network provided by an embodiment of the present invention, the method comprising:

[0044] Step S1: Acquire sensor information of the vegetable planting machine; determine the position of abnormal vegetables based on the sensor information; and collect a single vegetable image of each vegetable at the position of the abnormal vegetable.

[0045] There are a large number of sensors in the vegetable planter, and the distribution of distributed sensors can obtain sensor information at various locations in the vegetable planter. Because sensors contain multiple categories, sensor information is data containing multiple dimensions, each dimension representing a data category, such as water demand, light intensity, nutrient solution ratio, pH value, carbon dioxide concentration, soil moisture, etc.

[0046] The sensor information of each dimension corresponds to a standard sensor information, that is, the data displayed at the current moment during the normal growth process of vegetables. For example, the water demand will gradually increase as the vegetables continue to grow and develop. Therefore, the difference index between the sensor information of each dimension and the standard sensor information of the corresponding dimension is obtained. If the difference index is greater than the difference threshold, the corresponding sensor information is considered to be abnormal sensor information, and the position corresponding to the abnormal sensor information is the abnormal vegetable position.

[0047] In the embodiment of the present invention, when analyzing the difference index of the sensor information in each dimension, in order to facilitate the analysis, the difference information of each dimension is normalized and the difference index is set to 0.2.

[0048] After the abnormal vegetable position is determined, the vegetable image at the abnormal position can be obtained. The embodiment of the present invention aims to analyze the growth of vegetables, so after the vegetable image is obtained, it needs to be segmented and cropped to obtain a single vegetable image of each vegetable.

[0049] It should be noted that the segmentation of a single vegetable image can be achieved using a variety of technical means well known to those skilled in the art, such as semantic segmentation neural networks, etc., which will not be elaborated here.

[0050] Step S2: Input the single vegetable image into a pre-trained variational autoencoder network to obtain a reconstructed vegetable image.

[0051] If a vegetable is blocked by adjacent vegetables or other objects due to its growth, the vegetable area in the single vegetable image is an incomplete vegetable shape. Therefore, in order to analyze the growth of the vegetable at this time, it is necessary to reconstruct a complete growth and development reconstructed vegetable image based on the single vegetable image.

[0052] The variational autoencoder network describes the observation of the latent space in a probabilistic way and can be effectively used in the field of data generation. The traditional variational autoencoder network is an encoder-decoder structure, where the encoder is used to extract the features of the input data and the decoder is used to reconstruct the reconstructed image based on the features. It should be noted that a single vegetable image contains a variety of image features, such as image edge features, pixel features, color features, texture features, etc. The variational autoencoder network uses the latent variable probability distribution model corresponding to each feature as the latent variable feature output by the encoder. Because there are multiple features in the image, the encoder outputs a multi-dimensional latent variable probability distribution model, in which each dimension corresponds to an image feature. For the sake of ease of description, only the latent variable probability distribution model under one dimension will be used in the subsequent description.

[0053] In the process of training the variational autoencoder network, a single vegetable image with occlusion is used as a training image. In order to enable the trained variational autoencoder network to reconstruct the image of the vegetable growing normally, it is necessary to perform a corresponding parameter update process according to the occlusion degree of the single vegetable image during the training process.

[0054] It is known from prior knowledge that vegetables of excellent quality in vegetable planting machines have a certain degree of symmetry, that is, the left and right sides of the vegetables grow in the same way and will not grow in a deformed manner. Therefore, according to the symmetry of the vegetables, the vegetable image can be segmented with the root direction as the axis of symmetry, and the training image can be divided into an occluded part image and an unoccluded part image. It should be noted that the symmetry here is not the symmetry of the vegetable area in an image, that is, the occluded part image and the unoccluded part image in the segmentation result are not the same. The specific segmentation operation includes: inputting the training image into the semantic segmentation subnetwork in the variational autoencoder network, and outputting the occluded part image, the unoccluded part image and the occluded area on the occluded part image. It should be noted that the data feedback cycle of the vegetable planting machine is short, so there will be no situation where one vegetable completely blocks the adjacent vegetables, resulting in too little image information collected or the inability to collect vegetable image information.

[0055] In an embodiment of the present invention, in the semantic segmentation network, the pixels of the unobstructed image are marked as 0, the obstructed image is marked as 1, and the obstructed area is marked as 2. Through training with a large amount of data, the semantic segmentation network can segment three types of images.

[0056] The unobstructed image contains more complete information than the obstructed image. The semantic segmentation network can determine the degree of occlusion based on the morphological features of the obstructed image. If the obstructed image is partially obstructed, the semantic segmentation network can identify and mark the obstructed area based on the large number of edge features retained in the obstructed image. If the obstructed image is completely obstructed, the obstructed area cannot be identified because the edge information of the obstructed image is incomplete, that is, there is no obstructed area mark.

[0057] Training can be performed based on the regional information of each category in the training image, including:

[0058] If the occluded part of the image is partially occluded, it means that a large number of image features are still retained in the occluded part of the image. Therefore, the latent variable probability distribution model of the occluded part of the image can be used as the reconstruction model to reconstruct the image. Because the unoccluded part of the image has complete image features, the latent variable probability distribution model of the unoccluded part of the image can be used as a reference to update the reconstruction model. Obtain the model difference between the reconstruction model and the latent variable probability distribution model of the unoccluded part of the image, and update the reconstruction model according to the occluded area of ​​the occluded part of the image and the model difference, specifically including:

[0059] The ratio of the reconstructed model to the latent variable probability distribution model of the unoccluded part of the image is taken as the model difference. The model adjustment parameters are obtained according to the model difference and the occlusion area. The model adjustment parameters include:

[0060] α=e s+β

[0061] Among them, α is the model adjustment parameter, s is the occlusion area, and β is the model difference;

[0062] Updating and reconstructing the model based on model adjustment parameters includes:

[0063]

[0064] Among them, z1 is the reconstruction model, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, and α is the model adjustment parameter.

[0065] If the occluded part is completely occluded, it means that the image features of the occluded part are seriously missing and cannot be used to reconstruct the occluded area. Therefore, the latent variable probability distribution model of the unoccluded part is used as the reconstruction model to reconstruct the image and obtain the reconstructed vegetable image. The occluded area of ​​the occluded part image can be obtained according to the difference between the reconstructed vegetable image and the training image, and the reconstruction model is updated according to the model difference and occlusion area between the latent variable probability distribution model of the occluded part image at the corresponding position of the reconstructed vegetable image and the reconstruction model. It should be noted that the essence of the different training methods of the variational autoencoder network for complete occlusion and partial occlusion lies in the selection of the reconstruction model. Therefore, the model adjustment parameters of the reconstruction model are in the same form, and the model adjustment parameters when the occluded part is completely occluded are not repeated here.

[0066] By setting the network loss function, the variational autoencoder network is controlled to perform continuous iterative updates, with the KL divergence between the reconstructed model and the standard Gaussian distribution as the first loss function, and the difference between the reconstructed vegetable image and the training image as the second loss function, specifically including:

[0067]

[0068] Among them, L1 is the first loss function, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, α is the model adjustment parameter, and KL[] is the KL divergence calculation function.

[0069]

[0070] Among them, L2 is the second loss function, x is the training image, To reconstruct the vegetable image, ||||2 is the Euclidean distance calculation function.

[0071] The first loss function and the second loss function are combined as the overall network loss Loss of the variational autoencoder network, that is, Loss = L1 + L2. Through continuous iterative fitting of the network, the network training is completed when the overall network loss reaches the minimum.

[0072] Step S3: Obtain the center point offset information of the center point of the vegetable area according to the coordinate information of the center point of the reconstructed vegetable area in the reconstructed vegetable image and the center point of the vegetable area in the single vegetable image; and control the parameters of the vegetable planting machine according to the center point offset information.

[0073] The reconstructed vegetable image is an image of a normally growing vegetable. If the corresponding vegetable is blocked, the blocked part will be affected in development due to the lack of light intensity, while the unblocked part is normally affected by light, which will cause the vegetable to grow deformed and uneven in shape. That is, the unblocked part develops better and has a larger shape; the blocked part is affected in development and has a smaller shape.

[0074] The offset information reflected by the center point coordinate information in the reconstructed vegetable image and the single vegetable image can reflect the degree of occlusion influence of the occluded vegetable. That is, the more the center point of the reconstructed vegetable area in the reconstructed vegetable image is offset, the greater the occlusion influence of the corresponding vegetable.

[0075] The reconstructed vegetable image and the single vegetable image are overlapped and analyzed to be in the same image coordinate system and aligned. The reconstructed vegetable edge gradient amplitude in the reconstructed vegetable image and the single vegetable edge gradient amplitude in the single vegetable image are obtained. The deviation rate between the reconstructed vegetable image and the single vegetable image is obtained according to the overlap rate formula, and the deviation rate formula includes:

[0076]

[0077] in, is the deviation rate, (a x ,a y ) is the coordinate of the center point of the reconstructed vegetable area, (bx ,b y ) is the coordinate of the center point of the vegetable area, n is the number of reconstructed vegetable edge pixels, m is the number of single vegetable edge pixels, g i is the reconstructed vegetable edge gradient amplitude of the i-th reconstructed vegetable edge pixel, g j is the single vegetable edge gradient amplitude of the jth single vegetable edge pixel point.

[0078] The deviation rate formula reflects the degree of deviation based on the distance between the two center points and the difference in edge information in the image.

[0079] If the deviation rate is greater than the preset deviation rate threshold, it means that the occlusion has had a significant impact on the vegetables, and the center point offset information is obtained. In the embodiment of the present invention, in order to facilitate data analysis, the deviation rate is normalized and the deviation rate threshold is set to 0.2.

[0080] The offset information includes an offset direction and an offset distance. The direction from the center point of the reconstructed vegetable region to the center point of the vegetable region is used as the offset direction, and the distance from the center point of the reconstructed vegetable region to the center point of the vegetable region is used as the offset distance.

[0081] The parameters for controlling the vegetable planter based on the center point offset information include:

[0082] The opposite direction of the offset direction is used as the light compensation direction of the vegetable planter. The offset distance and the light adjustment amplitude are fitted according to historical data, and the light adjustment amplitude of the vegetable planter is controlled according to the fitting relationship. By adjusting the light direction and light intensity of the vegetable planter, the growth of vegetables can be adjusted, so that each vegetable grows evenly and the quality of vegetable production is improved.

[0083] In summary, the embodiment of the present invention obtains a single vegetable image at an abnormal position based on abnormal sensor information. A variational autoencoder network is trained using a single vegetable image as a training image, and the training image is divided into an occluded part image and an unoccluded part image. Different reconstruction models are selected according to the occlusion degree category of the occluded part, and the training of the variational autoencoder network is completed through continuous iterative updates. A reconstructed vegetable image of a single vegetable image is reconstructed according to the variational autoencoder network. The parameters of the vegetable planter are adjusted by the offset information between the reconstructed vegetable image and the single vegetable image. The embodiment of the present invention uses image information to analyze the growth of vegetables, and uses the offset information between the reconstructed vegetable image of normal growth and the real image to control the parameters of the vegetable planter, so as to obtain vegetables with uniform growth and high quality.

[0084] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0085] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A vegetable planting machine intelligent control method based on neural network, characterized in that: The method comprises: Acquire sensor information of the vegetable planting machine; determine the position of abnormal vegetables according to the sensor information; collect a single vegetable image of each vegetable at the position of the abnormal vegetable; The single vegetable image is input into a pre-trained variational autoencoder network to obtain a reconstructed vegetable image; the variational autoencoder network uses the single vegetable image as a training image, and divides the training image into an occluded partial image and an unoccluded partial image according to the symmetry of the vegetable; the occlusion degree category of the training image is determined according to the occluded partial image; the occlusion degree category includes complete occlusion and partial occlusion; if the occluded partial image is the partial occlusion, the latent variable probability distribution model of the occluded partial image is used as a reconstruction model, and the model difference between the reconstruction model and the latent variable probability distribution model of the unoccluded partial image is obtained, and the reconstruction model is reconstructed according to the occluded partial image. The reconstruction model is updated based on the occlusion area of ​​the image and the model difference; if the occluded part is completely occluded, the latent variable probability distribution model of the unoccluded part is used as the reconstruction model, the occlusion area of ​​the occluded part image is obtained according to the reconstructed vegetable image, and the reconstruction model is updated based on the model difference and the occlusion area between the latent variable probability distribution model of the occluded part image at the corresponding position of the reconstructed vegetable image and the reconstruction model; the variational autoencoder network uses the KL divergence between the reconstruction model and the standard Gaussian distribution as the first loss function, and the difference between the reconstructed vegetable image and the training image as the second loss function; Obtaining center point offset information of the center point of the vegetable area according to coordinate information of the center point of the reconstructed vegetable area in the reconstructed vegetable image and the center point of the vegetable area in the single vegetable image; controlling parameters of the vegetable planting machine according to the center point offset information; The obtaining of the center point offset information of the center point of the vegetable region according to the coordinate information of the center point of the reconstructed vegetable region in the reconstructed vegetable image and the center point of the vegetable region in the single vegetable image also includes: Obtain the reconstructed vegetable edge gradient amplitude in the reconstructed vegetable image and the single vegetable edge gradient amplitude in the single vegetable image; obtain the deviation rate between the reconstructed vegetable image and the single vegetable image according to the overlap rate formula, and the deviation rate formula includes: in, is the deviation rate, (a x ,a y ) is the coordinate of the center point of the reconstructed vegetable area, (b x ,b y ) is the coordinate of the center point of the vegetable area, n is the number of reconstructed vegetable edge pixels, m is the number of single vegetable edge pixels, g i is the reconstructed vegetable edge gradient amplitude of the i-th reconstructed vegetable edge pixel point, g j is the single vegetable edge gradient amplitude of the j-th single vegetable edge pixel point; If the deviation rate is greater than a preset deviation rate threshold, obtaining the center point offset information; The acquiring of the center point offset information comprises: The offset information includes an offset direction and an offset distance; the direction from the center point of the reconstructed vegetable area to the center point of the vegetable area is used as the offset direction, and the distance from the center point of the reconstructed vegetable area to the center point of the vegetable area is used as the offset distance; The parameters of the vegetable planting machine controlled according to the center point offset information include: The opposite direction of the offset direction is used as the light compensation direction of the vegetable planter; the offset distance and the light adjustment amplitude are fitted according to historical data, and the light adjustment amplitude of the vegetable planter is controlled according to the fitting relationship.

2. According to claim 1, a neural network-based intelligent control method for vegetable planting machines is characterized in that: Determining the location of abnormal vegetables according to the sensor information includes: Acquire the sensor information containing multiple dimensional information; acquire the difference index between the sensor information of each dimension and the standard sensor information of the corresponding dimension; if the difference index is greater than a preset difference threshold, the corresponding sensor information is considered to be abnormal sensor information, and the corresponding abnormal vegetable position is obtained according to the abnormal sensor information.

3. The method for intelligently controlling a vegetable planting machine based on a neural network according to claim 1, characterized in that: The step of dividing the training image into an obstructed portion image and an unobstructed portion image according to the symmetry of the vegetable comprises: The variational autoencoder network includes a semantic segmentation subnetwork, which inputs the training image into the semantic segmentation subnetwork and outputs the occluded portion image, the unoccluded portion image and the occluded region on the occluded portion image; if the occluded portion image is completely occluded, the occluded region information does not exist in the semantic segmentation result of the semantic segmentation subnetwork.

4. The method for intelligently controlling a vegetable planting machine based on a neural network according to claim 1, characterized in that: The updating of the reconstructed model according to the occlusion area of ​​the occluded part image and the model difference comprises: The ratio of the reconstructed model to the latent variable probability distribution model of the unblocked portion of the image is used as the model difference, and a model adjustment parameter is obtained according to the model difference and the occlusion area. The model adjustment parameter includes: α＝e s+β Wherein, α is the model adjustment parameter, s is the occlusion area, and β is the model difference; Updating the reconstructed model according to the model adjustment parameters includes: Among them, z1 is the reconstruction model, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, and α is the model adjustment parameter.

5. The method for intelligently controlling a vegetable planting machine based on a neural network according to claim 4 is characterized in that: The using the KL divergence between the reconstructed model and the standard Gaussian distribution as the first loss function comprises: Among them, L1 is the first loss function, σ 2 is the variance of the reconstructed model, μ is the mean of the reconstructed model, α is the model adjustment parameter, and KL[] is the KL divergence calculation function.

6. The method for intelligently controlling a vegetable planting machine based on a neural network according to claim 1, characterized in that: The taking the difference between the reconstructed vegetable image and the training image as the second loss function comprises: Wherein, L2 is the second loss function, x is the training image, is the reconstructed vegetable image, and ||||2 is the Euclidean distance calculation function.

Images