Neural network-based henhouse inspection method and robot

Abstract

The invention relates to the technical field of dead chicken recognition, in particular to a chicken house inspection method and robot based on a neural network. The invention discloses a chicken house inspection method based on a neural network and a robot. The method comprises the following steps: S1, acquiring primary image information of a chicken according to a first preset path; s2, judging whether the chicken is an abnormal chicken or not according to the primary image information and a neural network model, and retaining comprehensive image information of the abnormal chicken; and S3, uploading the reserved comprehensive image information of the abnormal chicken to a mobile phone APP of a breeder, and processing the abnormal chicken by the breeder. In use, due to the use of the image acquisition module, the interior of the chicken coop can be automatically monitored in real time, the phenomenon that live chickens in the chicken coop are infected due to dead chickens is prevented, the working pressure of breeders is greatly relieved, and the maintenance cost and the breeding cost of the live chickens in the chicken coop are reduced.

Description

Technical Field

[0001] This invention relates to the field of diseased and dead chicken identification technology, and in particular to a method and robot for inspecting chicken coops based on neural networks. Background Technology

[0002] When raising broiler chickens on a large scale, it is especially important to pay attention to whether there are dead chickens. If dead chickens are found, they should be removed from the chicken house immediately to prevent the dead chickens from causing a group effect and resulting in large-scale disease and economic losses. At present, the main method is to manually inspect the chicken houses to check for dead chickens. However, the chicken house environment is poor, which can easily cause infection to the inspectors and damage their health. In addition, some omissions may occur, resulting in poor inspection results. Some people use robots to inspect the chicken houses, using the difference in image information between dead and live chickens to distinguish between them. However, this inspection method is too simple and cannot fundamentally identify the condition of the chickens. Summary of the Invention

[0003] To overcome the above-mentioned technical problems, the present invention provides a method for inspecting chicken coops based on neural networks.

[0004] Technical solution: A method and robot for inspecting chicken coops based on neural networks, comprising:

[0005] S1: Obtain one image of the chicken according to the first preset path;

[0006] S2: Based on the image information and neural network model, determine whether the chicken is an abnormal chicken, and retain the comprehensive image information of the abnormal chicken;

[0007] S3: Upload the comprehensive image information of the abnormal chickens to the zookeeper's mobile APP, and the zookeeper will then handle the abnormal chickens.

[0008] Preferably, the image information includes: a photo of the chicken, the chicken's location information, the chicken coop's number information, and the number of layers, etc.

[0009] Preferably, the step of determining whether the chicken is an abnormal chicken based on the image information and the neural network model, and retaining the image information of the abnormal chicken, includes:

[0010] Initialize the neural network model;

[0011] Acquire historical image data of chickens in the chicken coop and preprocess the historical image data of chickens;

[0012] The training and testing sets of historical image data are constructed based on the preprocessed historical image data. The neural network model is then trained using the training and testing sets of historical image data to obtain the trained neural network model.

[0013] Preferably, the preprocessing of historical image data of chickens includes:

[0014] The historical image data of chickens is subjected to image denoising processing to obtain historical denoised image data of chickens; the image denoising using Gaussian filtering includes:

[0015] Where σ is the standard deviation; x and y are the coordinates of the convolution parameters;

[0016] Data augmentation processing is performed on the historical denoised image data of chickens to obtain historical augmented image data of chickens;

[0017] The data augmentation process includes at least one of the following: image rotation, image scaling, image cropping, color adjustment, random slicing, and center cropping.

[0018] Preferably, the step of determining whether the chicken is an abnormal chicken based on the image information and the neural network model, and retaining the image information of the abnormal chicken, includes:

[0019] S21: Input the chicken photo from the acquired image information into the neural network model, and determine whether the chicken is an abnormal chicken based on the output of the neural network model;

[0020] If the output result is an abnormal chicken, then save the image information for that instance;

[0021] If the output result is not an abnormal chicken, then delete the image information for that instance;

[0022] S22: Statistically analyze the retained image information to form an abnormal chicken image dataset;

[0023] S23: Plan a second preset path based on the image information in the abnormal chicken image dataset;

[0024] S24: Obtain secondary image information of the chicken according to the second preset path, input the chicken photo in the obtained secondary image information into the neural network model, and determine whether the chicken is an abnormal chicken based on the output result of the neural network model;

[0025] If the output result is an abnormal chicken, then save the secondary image information;

[0026] If the output result is not an abnormal chicken, then delete the secondary image information.

[0027] Preferably, the step of planning a second preset path based on primary image information in the abnormal chicken image dataset includes:

[0028] The location information of the chicken in the primary image information of the abnormal chicken image dataset is obtained, and a second preset path is planned based on the location information of the primary image information of the abnormal chicken image dataset. The second preset path is the shortest path through the chicken positions contained in the abnormal chicken image data. The step of controlling the robot to obtain secondary image information of the chicken according to the second preset path includes: the secondary image information is the recording of the chicken in the primary image information.

[0029] Preferably, if the output result is an abnormal chicken, then saving the secondary image information includes:

[0030] The similarity between the chicken photo in the secondary image information and the chicken photo in the primary image information is compared. If the similarity is less than a first preset value, the secondary image information is deleted.

[0031] If the similarity is greater than the first preset value, a stimulus signal is applied to the chicken, and three images of the chicken are acquired. The similarity of the chicken photos in the three images is compared with the chicken photos in the first and second images. If the similarity is less than the second preset value, the three images are deleted. If the similarity is greater than the second preset value, the chicken is determined to be an abnormal chicken, and the first, second, and third images of the chicken are retained as the comprehensive image information of the abnormal chicken.

[0032] Preferably, both the first preset value and the second preset value are 95%.

[0033] Preferably, the stimulation signal includes one of: irradiation by light or contact.

[0034] Preferably, a neural network-based chicken coop inspection robot includes a mobile module, a main housing mounted on the mobile module, a lifting module mounted on the main housing, an image acquisition module mounted on the lifting module, a stimulus application module mounted on the image acquisition module, and a positioning module, a control terminal, and a data processing module disposed inside the main housing.

[0035] The beneficial effects of this invention are as follows: By using an image acquisition module, this invention can automatically monitor the inside of the chicken cage in real time, preventing infection of live chickens in the chicken house due to dead chickens. This greatly reduces the workload of the breeders and lowers the maintenance and breeding costs of live chickens in the chicken house. This invention adopts a progressive screening method, replanning the path after the initial screening, saving equipment working time. It uses similarity to solve the problem of difficulty in distinguishing between non-standing chickens and dead chickens in the neural network model. The final comprehensive image information is sent to the breeder's APP, which makes it convenient for the breeder to deal with dead chickens in the chicken house in a timely manner. Attached Figure Description

[0036] Figure 1 A flowchart for chicken coop inspection methods;

[0037] Figure 2 This is a flowchart illustrating a specific implementation of S2 of the present invention;

[0038] Figure 3 This is a three-dimensional structural diagram of the present invention.

[0039] In the attached diagram, 1: moving module, 2: main housing, 3: lifting module, 4: image acquisition module, and 5: stimulus application module. Detailed Implementation

[0040] The present invention will be further described below with reference to the embodiments shown in the accompanying drawings.

[0041] Example 1: A method for inspecting chicken coops based on neural networks, such as... Figures 1-2 As shown, it includes:

[0042] S1: Obtain one image of the chicken according to the first preset path;

[0043] The image information includes: a photo of the chicken, the chicken's location information, the chicken coop number and the number of layers, etc. Each image information is the first image information of the chicken obtained when the patrol begins.

[0044] The first preset path is the shortest path to obtain image information of all chickens in the chicken coop.

[0045] The steps to obtain a trained neural network model of a chicken include:

[0046] Initialize the neural network model;

[0047] Acquire historical image data of chickens in the chicken coop and preprocess the historical image data of chickens;

[0048] Preprocessing of historical image data of chickens includes:

[0049] The historical image data of chickens is subjected to image denoising processing to obtain historical denoised image data of chickens; the image denoising using Gaussian filtering includes:

[0050] Where σ is the standard deviation; x and y are the coordinates of the convolution parameters;

[0051] Data augmentation processing is performed on the historical denoised image data of chickens to obtain historical augmented image data of chickens;

[0052] The data augmentation process includes at least one of the following: image rotation, image scaling, image cropping, color adjustment, random slicing, and center cropping.

[0053] The training and testing sets of historical image data are constructed based on the preprocessed historical image data. The neural network model is then trained using the training and testing sets of historical image data to obtain the trained neural network model.

[0054] S2: Based on the image information and neural network model, determine whether the chicken is an abnormal chicken, and retain the comprehensive image information of the abnormal chicken;

[0055] S21: Input the chicken photo from the acquired image information into the neural network model, and determine whether the chicken is an abnormal chicken based on the output of the neural network model;

[0056] If the output result is an abnormal chicken, then save the image information for that instance;

[0057] If the output result is not an abnormal chicken, then delete the image information for that instance;

[0058] Specifically, in the above process, "abnormal chickens" refers to chickens that are not standing or dead, not just dead chickens. The neural network model cannot completely distinguish between chickens that are not standing and dead chickens. If the chicken photos are side or back views, some chickens that are not standing will still be classified as dead. Therefore, the image information of dead chickens identified by the neural network is set as abnormal chickens. When the output result is abnormal chickens, the image information is saved to facilitate the screening of abnormal chickens in subsequent processes, thereby improving the accuracy of screening dead chickens. When the output result is not abnormal chickens, the chicken is a normal chicken, and the image information of the normal chicken is not needed. Therefore, the image information of the normal chicken needs to be deleted.

[0059] S22: Statistically analyze the retained image information to form an abnormal chicken image dataset;

[0060] Specifically, the image information of the abnormal chickens (first-order image information) is integrated, and only the abnormal chickens are further screened in the subsequent processing to reduce the amount of data processing. If all chickens are still screened, it will lead to an excessive amount of information to be processed, especially when the number of dead chickens is small.

[0061] S23: Plan a second preset path based on the primary image information in the abnormal chicken image dataset; obtain the location information in the primary image information in the abnormal chicken image dataset, and plan a second preset path based on the location information in the primary image information in the abnormal chicken image dataset. The second preset path is the shortest path that passes through the chicken locations contained in the abnormal chicken image data. For example, the first preset path needs to pass through the first passage of the chicken coop. However, after the primary image information collection is completed, no abnormal chicken is found in the first passage. Therefore, it is not necessary to pass through the first passage during the secondary image information collection.

[0062] S24: Obtain secondary image information of the chicken according to the second preset path. The secondary image information is a second image record of the chicken in the saved primary image information, avoiding the need to collect image information according to the first preset path and reducing the acquisition time of secondary image information.

[0063] The chicken photos from the acquired secondary image information are input into a neural network model, and the output of the neural network model is used to determine whether the chicken is an abnormal chicken.

[0064] If the output result is not an abnormal chicken, then delete the secondary image information and the primary image information.

[0065] If the output result is an abnormal chicken, then save the secondary image information;

[0066] Specifically, when taking a second picture of a chicken in a primary image, the existing neural network model is used first to identify the chicken in the secondary image without using other identification methods. If the chicken changes from a non-standing state to a standing state during the two pictures, it proves that the chicken in the primary image is a non-standing chicken and not a dead chicken. After the output result of the secondary image is no longer an abnormal chicken, both the primary and secondary image information of the chicken are deleted.

[0067] If the output of the secondary image information is still an abnormal chicken, then the probability that the chicken is dead is relatively high. Compare the similarity between the chicken photo in the secondary image information and the chicken photo in the primary image information.

[0068] If the similarity is less than the first preset value, the secondary image information is deleted. When the similarity between the two photos is low, it proves that the chicken's state has changed significantly during the two photo shooting periods. However, even after a significant change in state, the chicken can still be identified as a dead chicken by the neural network model. For example, if the chicken turns from its side to its back, the neural network model cannot directly distinguish it as a live chicken. Therefore, similarity comparison is used to improve the distinguishability of image information.

[0069] If the similarity is greater than the first preset value, a stimulus signal is applied to the chicken, and three images of the chicken are acquired. The similarity of the chicken photos in the three images is compared with the chicken photos in the first and second images. If the similarity is less than the second preset value, the three images are deleted. If the similarity is greater than the second preset value, the chicken is determined to be an abnormal chicken, and the first, second, and third images of the chicken are retained as the comprehensive image information of the abnormal chicken.

[0070] Both the first preset value and the second preset value are 95%.

[0071] The stimulus signal includes one of the following: illumination by light or contact.

[0072] Specifically, when the similarity between two photos exceeds 95%, it proves that the chicken's state did not change significantly during the two photo-taking periods. Therefore, it is necessary to apply a stimulus to the chicken to determine whether it is dead. If the chicken is dead, it will not stand up after the stimulus is applied. If the chicken is not standing, it will stand up after the stimulus is applied. By comparing the similarity between the photos of the chicken before and after the stimulus is applied, it is determined whether the chicken is dead again. Finally, the primary, secondary, and tertiary image information of the abnormal chicken are used as the comprehensive image information of the abnormal chicken. The image information of the chicken is formatted as follows: the photo of the chicken is used as the background, the position of the chicken in the photo is marked by image blocks, and the chicken's position information, the chicken cage number information, and the layer information are distributed on one side of the chicken photo.

[0073] S3: Upload the comprehensive image information of the abnormal chickens to the keeper's mobile APP. The keeper will then handle the abnormal chickens. The keeper will determine whether the chicken is dead based on the comprehensive image information. If the chicken is dead, it will be moved out of the chicken house. If the chicken is not dead, it may be sick and will be further treated.

[0074] Example 2: Based on Example 1, such as Figure 3 As shown, a neural network-based chicken coop inspection robot includes a mobile module 1, a main shell 2 mounted on the mobile module 1, a lifting module 3 mounted on the main shell 2, an image acquisition module 4 mounted on the lifting module 3, a stimulus application module 5 mounted on the image acquisition module 4, and a positioning module, a control terminal, and a data processing module disposed inside the main shell 2.

[0075] The robot is activated by a timer or manually, collecting image information of chickens in the coop along a path. The lifting module 3 is used to adjust the position of image acquisition or adjust the height when the robot encounters obstacles, thereby achieving automatic obstacle avoidance and greatly improving the practicality and adaptability of the equipment. The stimulation application module 5 includes a light output device and a robotic arm. When it is necessary to stimulate the chickens, the light output device outputs light of different intensities or colors to stimulate the chickens, and the robotic arm touches the chickens to increase the intensity of the stimulation, causing non-standing chickens to stand up or move. The movement when touching the chickens should be small to avoid affecting other chickens in the same cage.

[0076] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inspecting chicken coops based on neural networks, characterized in that, include: S1: Obtain one image of the chicken according to the first preset path; S2: Based on the image information and neural network model, determine whether the chicken is an abnormal chicken, and retain the comprehensive image information of the abnormal chicken; S3: Upload the comprehensive image information of the abnormal chickens to the zookeeper's mobile APP, and the zookeeper will then handle the abnormal chickens.

2. The method for inspecting chicken coops based on neural networks as described in claim 1, characterized in that, The image information includes: a photo of the chicken, the chicken's location information, the chicken coop's number information, and the number of layers, etc.

3. The method for inspecting chicken coops based on neural networks as described in claim 1, characterized in that, Before determining whether a chicken is abnormal based on primary image information and a neural network model, and before retaining the image information of abnormal chickens, the following steps are included: Initialize the neural network model; Acquire historical image data of chickens in the chicken coop and preprocess the historical image data of chickens; The training and testing sets of historical image data are constructed based on the preprocessed historical image data. The neural network model is then trained using the training and testing sets of historical image data to obtain the trained neural network model.

4. The method for inspecting chicken coops based on neural networks as described in claim 3, characterized in that, Preprocessing of historical image data of chickens includes: The historical image data of chickens is subjected to image denoising processing to obtain historical denoised image data of chickens; Image noise reduction using Gaussian filtering includes: Where σ is the standard deviation, and x and y are the coordinates of the convolution parameters; Data augmentation processing is performed on the historical denoised image data of chickens to obtain historical augmented image data of chickens; The data augmentation process includes at least one of the following: image rotation, image scaling, image cropping, color adjustment, random slicing, and center cropping.

5. The method for inspecting chicken coops based on neural networks as described in claim 1, characterized in that, The step of determining whether a chicken is abnormal based on primary image information and a neural network model, and retaining the image information of abnormal chickens, includes: S21: Input the chicken photo from the acquired image information into the neural network model, and determine whether the chicken is an abnormal chicken based on the output of the neural network model; If the output result is an abnormal chicken, then save the image information for that instance; If the output result is not an abnormal chicken, then delete the image information for that instance; S22: Statistically analyze the retained image information to form an abnormal chicken image dataset; S23: Plan a second preset path based on the image information in the abnormal chicken image dataset; S24: Obtain secondary image information of the chicken according to the second preset path, input the chicken photo in the obtained secondary image information into the neural network model, and determine whether the chicken is an abnormal chicken based on the output result of the neural network model; If the output result is an abnormal chicken, then save the secondary image information; If the output result is not an abnormal chicken, then delete the secondary image information.

6. The method for inspecting chicken coops based on neural networks as described in claim 5, characterized in that, The step of planning a second preset path based on primary image information in the abnormal chicken image dataset includes: The location information of a single image in the abnormal chicken image dataset is obtained, and a second preset path is planned based on the location information of the single image in the abnormal chicken image dataset. The second preset path is the shortest path that passes through the chicken locations contained in the abnormal chicken image data. The step of controlling the robot to obtain secondary image information of the chicken according to the second preset path includes: the secondary image information is the recording of a second image of the chicken in the saved primary image information.

7. The method for inspecting chicken coops based on neural networks as described in claim 5, characterized in that, If the output result is an abnormal chicken, then saving the secondary image information includes: The similarity between the chicken photo in the secondary image information and the chicken photo in the primary image information is compared. If the similarity is less than a first preset value, the secondary image information is deleted. If the similarity is greater than the first preset value, a stimulus signal is applied to the chicken, and three images of the chicken are acquired. The similarity of the chicken photos in the three images is compared with the chicken photos in the first and second images. If the similarity is less than the second preset value, the three images are deleted. If the similarity is greater than the second preset value, the chicken is determined to be an abnormal chicken, and the first, second, and third images of the chicken are retained as the comprehensive image information of the abnormal chicken.

8. The method for inspecting chicken coops based on neural networks as described in claim 7, characterized in that, Both the first preset value and the second preset value are 95%.

9. The method for inspecting a chicken coop based on a neural network as described in claim 7, characterized in that, The stimulus signal includes one of the following: illumination by light or contact.

10. A chicken coop inspection robot based on neural networks, characterized in that, It includes a mobile module (1), a main housing (2) installed on the mobile module (1), a lifting module (3) installed on the main housing (2), an image acquisition module (4) installed on the lifting module (3), a stimulation application module (5) installed on the image acquisition module (4), and a positioning module, a control terminal and a data processing module are installed inside the main housing (2).

Images