A non-contact automatic identification method and system for sick chickens in cages

Abstract

The present invention relates to a non-contact automatic identification method for sick chickens in cages, comprising the steps of: collecting audio data near the chicken coop using an acoustic sensor, then converting the audio data into a spectrogram using a STFT; inputting the spectrogram into a trained first target detection network to locate the call area in the spectrogram; cutting out each call area from the spectrogram one by one, then inputting the data into a trained convolutional neural network for voiceprint classification, outputting the call type corresponding to each call area, and issuing a sick chicken alarm if the call area contains a chicken cough. The present invention also relates to a non-contact automatic identification system for sick chickens in cages. The advantages of the present invention include targeted feature extraction of each call area from the spectrogram, inputting all extracted call areas into a convolutional neural network for individual identification, and comprehensively analyzing the call types corresponding to all call areas in the spectrogram, thereby significantly improving the accuracy of identifying abnormal calls from sick chickens.

Description

Technical Field

[0001] The present invention relates to the field of non-contact detection of chickens in cages, and in particular to a non-contact automatic identification method and system for sick chickens in cages. Background Art

[0002] Chicken farms house numerous cages stacked high in density, with each cage housing a large number of birds. This makes it inevitable that birds may become sick. Manually inspecting sick birds is a tedious and meticulous task, consuming both time and effort. Furthermore, because sick birds are not readily apparent within the cages, timely and comprehensive detection of any illnesses is difficult.

[0003] The existing publication number is CN112562739A, "A Sound Data Analysis System for Respiratory Diseases in Large-Scale Broiler Chickens." The disclosed solution is to collect sound data, analyze it through a strengthened neural network, and detect sick chickens. In actual applications, there are many types of sounds collected in the chicken house, including normal chicken calls, abnormal calls of sick chickens, and environmental noise. The collected sounds are directly input into the neural network, and the neural network outputs the sound of the category with the highest matching degree. When the collected audio contains both normal chicken calls and chicken coughs, because the chicken coughs are short and weak, and there is no clear standard for the voiceprint characteristics of chicken coughs, when the neural network recognizes them, the matching degree of the chicken coughs will be significantly lower than that of the normal chicken calls. Therefore, if the collected sound data is directly submitted to the neural network for recognition, it is easy to ignore the chicken coughs and only output the type of call with the highest matching degree - normal chicken calls, resulting in a large recognition error. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a non-contact automatic identification method for sick chickens in cages, which specifically cuts and extracts the calling areas one by one from the spectrum graph, and then inputs all the obtained calling areas into a convolutional neural network for one by one identification, comprehensively analyzes the calling types corresponding to all the calling areas in the spectrum graph, and greatly improves the recognition accuracy of abnormal calls of sick chickens.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] Technical Solution 1

[0007] A non-contact automatic identification method for sick chickens in cages includes a voiceprint detection step: collecting audio data near the chicken cage using an acoustic sensor, and then converting the audio data into a spectrogram using a STFT; inputting the spectrogram into a trained first target detection network to locate the calling area in the spectrogram; cutting out each calling area from the spectrogram one by one, and then inputting the trained convolutional neural network for voiceprint classification, outputting the call type corresponding to each call area, and issuing a sick chicken alarm if the call type contains a chicken cough.

[0008] More preferably, the size of the cropped call area image is matched to the original size of the spectrogram and then input into the trained convolutional neural network.

[0009] Preferably, the method also includes an image detection step: collecting images of the upper part and the bottom part of the chicken cage through a camera, and then inputting them into a trained second target detection network; the second target detection network recognizes the chicken head and comb from the upper image and the bottom image, outputs an identification image marked with the position of the chicken head and comb frame, and determines whether the comb is atrophied based on the comb feature value. If so, a sick chicken alarm is issued; at the same time, the second target detection network further identifies the comb, and if it is a comb with white spots, an alarm is issued; the bottom image is input into a trained third target detection network, and an image containing the chicken body is identified from the bottom image, and a bottom image marked with the position of the chicken body is output; the historical bottom image of the chicken cage is extracted, and if there are chickens in the historical bottom image, the chicken body position is compared to determine whether the chicken body has moved. If not, a sick chicken alarm is issued. If so, it is determined to be a healthy chicken; if there is no chicken body in the historical bottom image, it is determined to be a healthy chicken.

[0010] More preferably, the specific process of judging whether the comb is shrunk based on the comb characteristic value is as follows: analyzing the recognition image, binding the comb and the chicken head one by one: judging whether there are more than one comb positions intersecting with the same chicken head position, if not, binding the chicken head and the comb; if so, calculating the intersection value of each comb and the chicken head one by one, and then calculating the ratio of the intersection value to the comb area; if the ratio is less than a preset intersection threshold, judging that the comb and the chicken head have no binding relationship; if the ratio is greater than the preset intersection threshold, taking the comb with the highest ratio and binding it to the chicken head; calculating the comb characteristic value: the comb characteristic value is the ratio of the comb size to the chicken head size, if the ratio is less than the preset characteristic threshold, judging that the comb is shrunk, and issuing a sick chicken alarm.

[0011] More preferably, the comb characteristic value is multidimensional data, including the ratio of comb size to head size and any combination of the following items: head size, comb size, comb length-to-width ratio, head length-to-width ratio; after the multidimensional data is standardized by standard deviation, it is input into a trained LSTM neural network, and the LSTM neural network analyzes the data to identify whether the comb is atrophied. If so, a sick chicken alarm is issued.

[0012] More preferably, the bottom image is input into a trained third target detection network, an image containing the chicken body is identified from the bottom image, and the position of the chicken body is located. The chicken body position comparison is specifically as follows: extracting the position information of the chicken body in the current bottom image, extracting the historical bottom image of the chicken in the cage, if the chicken body position is located in the historical bottom image, then extracting the position information of the chicken body in the historical bottom image, calculating the intersection value IoU of the current position information and the historical position information, and judging whether there has been a chicken at the position for a long time based on the intersection value IoU. If the intersection value IoU is greater than the preset intersection threshold, the overlap is high and it is judged to be a sick chicken, otherwise, it is judged to be a healthy chicken. If the chicken body position is not located in the historical bottom image, then the chicken in the current identified image is judged to be a healthy chicken.

[0013] The present invention also aims to provide a non-contact automatic identification system for sick chickens in cages.

[0014] To achieve the above object, the present invention adopts the following technical solutions:

[0015] Technical Solution 2

[0016] A non-contact automatic identification system for sick chickens in cages comprises an acoustic sensor, a memory, and a processor. The acoustic sensor collects audio data near the chicken coop, and the memory stores instructions suitable for being loaded by the processor and executing voiceprint detection steps: collecting audio data near the chicken coop using the acoustic sensor, and then converting the audio data into a spectrogram using a STFT; inputting the spectrogram into a trained first target detection network, and locating the calling area in the spectrogram; cutting out each calling area from the spectrogram one by one, and then inputting the trained convolutional neural network for voiceprint classification, outputting the call type corresponding to each call area, and issuing a sick chicken alarm if the call type contains a chicken cough.

[0017] More preferably, the size of the cropped call area image is matched to the original size of the spectrogram and then input into the trained convolutional neural network.

[0018] More preferably, the system further comprises a camera, which captures images of the upper portion and the bottom portion of the chicken cage. The instructions are loaded by the processor and further execute an image detection step: the captured upper and lower images are input into a trained second target detection network; the second target detection network identifies the chicken head and comb from the upper and lower images, outputs an identification image marked with the positions of the chicken head and comb borders, and determines whether the comb is atrophied based on the comb feature value. If so, a sick chicken alarm is issued; at the same time, the second target detection network further identifies the comb, and if it is a comb with white spots, an alarm is issued;

[0019] The bottom image is input into a trained third object detection network, an image containing a chicken body is identified from the bottom image, and a bottom image with the chicken body position marked is output; historical bottom images of the chicken cage are extracted, and if a chicken exists in the historical bottom image, the chicken body position is compared to determine whether the chicken body has moved. If not, a sick chicken alarm is issued; if so, the chicken is determined to be healthy; if no chicken body is present in the historical bottom image, the chicken is determined to be healthy.

[0020] Furthermore, the specific process of judging whether the comb is shrunk based on the comb characteristic value is as follows: analyzing the recognition image, binding the comb and the chicken head one by one: judging whether there are more than one comb positions intersecting with the same chicken head position, if not, binding the chicken head and the comb; if so, calculating the intersection value of each comb and the chicken head one by one, and then calculating the ratio of the intersection value to the comb area; if the ratio is less than a preset intersection threshold, judging that the comb and the chicken head have no binding relationship; if the ratio is greater than the preset intersection threshold, taking the comb with the highest ratio and binding it to the chicken head; calculating the comb characteristic value: the comb characteristic value is the ratio of the comb size to the chicken head size, if the ratio is less than the preset characteristic threshold, judging that the comb is shrunk, and issuing a sick chicken alarm.

[0021] The present invention has the following beneficial effects:

[0022] 1. The present invention provides a non-contact automatic identification method for sick chickens in cages. First, a target detection network is used to locate all the call areas containing calls in the spectrum graph. Then, a convolutional neural network is used to identify the call type corresponding to each call area one by one, effectively preventing missed detections and significantly improving recognition accuracy.

[0023] 2. The present invention provides a non-contact automatic identification method for sick chickens in cages, which successfully transforms the original spectrum map recognition with a large amount of data into the spectrum map recognition of a small area with calls. This not only greatly saves target detection network resources, but also further improves the recognition efficiency and accuracy of convolutional neural networks.

[0024] 3. The present invention provides a non-contact automatic identification method for sick chickens in cages. The method uses voiceprint recognition to identify sick chickens that make abnormal calls. At the same time, it also uses image recognition to detect sick chickens with abnormal behaviors and abnormal combs. Sick chickens can be identified from multiple angles, greatly improving the recognition rate of sick chickens.

[0025] 4. The non-contact automatic identification system for sick chickens in cages of the present invention has the same beneficial effects as the non-contact automatic identification method. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a non-contact automatic identification method for sick chickens in cages according to the present invention;

[0027] Figure 2This is a flowchart of locating the calling area in the present invention;

[0028] Figure 3 This is a schematic diagram of the spectrum of sick chicken calls of the present invention;

[0029] Figure 4 This is a flow chart of the second embodiment of the present invention;

[0030] Figure 5 This is a schematic diagram of the upper part of the chicken cage of the present invention;

[0031] Figure 6 This is a schematic diagram of the bottom image of a chicken cage according to the present invention;

[0032] Figure 7 and Figure 8 The identification image of the present invention is marked with the position of the chicken head and the comb frame;

[0033] Figure 9 To Figure 11 The bottom image of the chicken body appears at the bottom of the chicken cage of the present invention;

[0034] Figure 12 The present invention is a block diagram of a non-contact automatic identification system for sick chickens in cages. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] The main symptom of sick chickens in a chicken farm is abnormal sounds, such as hoarse sounds or mixed breathing sounds, snoring, and strange sounds. In this embodiment, the abnormal sounds are defined as chicken coughing.

[0038] See also Figure 1 and Figure 2 A non-contact automatic identification method for sick chickens in cages includes the following steps: collecting audio data near the chicken cage using an acoustic sensor, and then converting the audio data into a spectrogram using a STFT; inputting the spectrogram into a trained first target detection network to locate the calling area in the spectrogram; cutting out each calling area from the spectrogram one by one, and then inputting the sound into a trained convolutional neural network for voiceprint classification, outputting the call type corresponding to each call area, and issuing a sick chicken alarm if the call type contains a chicken cough.

[0039] like Figure 3As shown, the first object detection network uses a large number of spectrograms of sick chicken calls, normal chicken calls, and non-calling spectrograms (including spectrograms of equipment noise, environmental noise, and silence) as training samples, enabling it to distinguish environmental noise from calls and identify and locate the calling area. The architecture of the first object detection network is primarily lightweight. The network consists of a ShuffleNetV2 backbone network, an FPN dilated encoder, and a nanohead head. The backbone network extracts basic features of the input image, while the FPN dilated encoder fuses features from different layers to generate new features. The nanohead head ultimately predicts the type and location of the object. Specifically: 1) The input size of the ShuffleNetV2 backbone network is resized from 256*256 of the original object detection model Nanodet to 128*128. Using a 128*128 size preserves the original spectrogram information and saves object detection network resources. 2) The features from the ShuffleNetV2 feature extraction process to the dilated encoder FPN are the features of the 3rd, 4th, and 5th layers of ShuffleNetV2. The feature scale of a single call is small, and discarding the first and second layer features with larger features helps to fit the model.

[0040] The convolutional neural network can preferably be a deep convolutional neural network (DCNN). The convolutional neural network uses a large number of sick chicken call spectrograms, normal rooster call spectrograms, and normal hen call spectrograms as training samples, enabling it to identify call types. The call area image from which features are extracted is resized to the original size of the spectrogram. For example, the cropped call area image is enlarged to the original 500*500 spectrogram before being fed into the trained convolutional neural network. The larger-scale call area spectrogram enables more precise classification of voiceprints.

[0041] The above-mentioned embodiment adopts a method for comprehensively analyzing all types of calls in the audio data. First, all the call areas in the spectrum are located through the target detection network, and then the call type corresponding to each call area is identified one by one through the convolutional neural network. Because the convolutional neural network inputs the spectrum with the call area cut out in the process of identifying the call type, it has eliminated most of the noise interference and other normal calls, and only identifies the call area in a targeted manner. Therefore, the present invention can effectively prevent missed detection and greatly improve the recognition accuracy. At the same time, the method of first locating the call area and then identifying the call type successfully transforms the original spectrum recognition with a large amount of data into the spectrum recognition of a small area with calls, which greatly saves the target detection network resources and further improves the recognition efficiency and accuracy of the convolutional neural network.

[0042] Example 2

[0043] The difference between this embodiment and embodiment 1 is that a non-contact sick chicken identification method combining voiceprint recognition with image recognition is provided, which can not only identify sick chickens with abnormal calls, but also identify sick chickens with abnormal behaviors and abnormal combs, thereby achieving a better sick chicken identification rate.

[0044] The main symptom of sick chickens in a chicken farm is abnormal sounds, such as hoarseness or interspersed breathing sounds, snoring, and strange cries. Abnormal sounds are defined as chicken coughs. In addition, sick chickens have other symptoms such as white spots on the comb and long-term immobility. These sick chickens need to be promptly discovered and eliminated during inspections. In addition, low-producing chickens with low egg production rates or chickens that have no eggs also fall into the category of chickens that need to be promptly discovered and eliminated during inspections. Therefore, in the invention, these are also included in the category of sick chickens. The characteristic of chickens that have no eggs - atrophy of the comb - is used to identify them and issue an alarm during the identification process.

[0045] See also Figure 4 , this embodiment performs the image detection step while performing the voiceprint detection step: please refer to Figure 5 and Figure 6 , the upper image of the chicken cage and the bottom image of the chicken cage are collected by the camera, and then input into the trained second target detection network; the second target detection network recognizes the chicken head and comb from the upper image and the bottom image, such as Figure 7 and Figure 8As shown, an identification image with the positions of the chicken's head and comb borders is output. The comb feature values are then used to determine whether the comb is atrophied. If so, a sick chicken alert is issued. Simultaneously, the second target detection network further identifies the comb. If the comb has white spots, a sick chicken alert is issued. The specific process for determining whether the comb is atrophied based on the comb feature values is as follows: analyzing the identification image and performing a one-to-one binding of the combs to the chicken's head. This embodiment uses the IoU algorithm to perform a one-to-one binding of the head and comb: determining whether there are more than one comb positions that intersect with the same head position. If not, the head and comb are bound. If so, the intersection value of each comb with the head is calculated one by one, and then the ratio of the intersection value to the comb area is calculated. If the ratio is less than a preset intersection threshold, the comb is determined to be unbound to the head. If the ratio is greater than the preset intersection threshold, the comb with the highest ratio is bound to the head. For laying hens, the intersection threshold can be set to 0.7. The calculation process of the comb characteristic value is as follows: the comb characteristic value is the ratio of comb size to head size. If this ratio is less than a preset characteristic threshold, the comb is determined to be atrophied and a sick chicken alarm is issued. The comb characteristic value can also be multidimensional data, including the ratio of comb size to head size and any combination of the following items: head size, comb size, comb length-to-width ratio, and head length-to-width ratio. After standardizing the multidimensional data using standard deviation, it is input into a trained LSTM neural network. The LSTM neural network analyzes the data and identifies whether the comb is atrophied. If so, a sick chicken alarm is issued.

[0046] Since the state of the chicken's head is uncertain, the chicken's head may be turned to the side, the chicken's head may face the camera directly, or the chicken's head may be partially blocked, etc. Therefore, the comb feature value uses multi-dimensional data to judge whether the comb is atrophied from multiple angles, which can greatly improve the accuracy of the identification results of dead chickens. The comb feature value can be obtained according to the length and width of the box corresponding to the chicken head and comb in the recognition image. In this embodiment, the comb feature value is taken as five-dimensional data as an example, including the size of the chicken head, the size of the comb, the comb size / the size of the chicken head, the length-to-width ratio of the comb, and the length-to-width ratio of the chicken head. In order to better analyze the data, standard Scale is used to pre-process the data so that the processed data conforms to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1. Its conversion function is: x *=(x-μ) / σ, where μ is the mean of all sample data corresponding to x, σ is the standard deviation of all sample data corresponding to x, and x is a parameter in the comb eigenvalue, for example, x is the length-to-width ratio of the comb. The standardized comb eigenvalues are input into a trained deep learning network, and the data is analyzed by the deep learning network to determine whether the comb is atrophied and identify sterile chickens. The deep learning network can adopt a convolutional neural network, an LSTM algorithm model or a machine learning model, etc. Taking the LSTM algorithm model as an example, the input size is 5, the level is set to 1 layer, and the target labels are: 1) normal chicken, 2) sterile chicken. The labels are converted to Onehot format for easy training.

[0047] The bottom image is input into the trained third object detection network, the image containing the chicken body is identified from the bottom image, and the bottom image with the chicken body position marked is output; the bottom image containing the chicken appearing at the bottom is shown in Figures 9 to Figure 11 As shown, Figure 9 For healthy chicken, Figure 10 For sick chickens, Figure 11 These chickens are likely sick and near death. The system extracts historical bottom images of the cage, typically the last identified image of the chickens in the cage. If a chicken is present in the historical bottom image, a position comparison is performed to determine whether the chicken has moved. If not, a sick chicken alert is issued. If so, the chicken is considered healthy. If the chicken is not present in the historical bottom image, the chicken is considered healthy. The position comparison involves extracting the chicken's position information from the historical bottom image. This position information refers to the bounding box information that locates the chicken's position. The intersection value (IoU) of the current position information and the historical position information is calculated. Based on the IoU, the system determines whether a chicken has been present at that location for a long time. If the IoU exceeds a preset intersection threshold, the overlap is high and the chicken is considered sick. Since healthy chickens are relatively active and rarely remain in the same position for an extended period of time, they can be identified as sick. The intersection threshold can typically be set to 0.7. The intersection value IoU calculates the overlap rate of the chicken's border in the current bottom image and the chicken's border in the historical bottom image, that is, the ratio of their intersection and union.

[0048] Both the second and third object detection networks utilize a high-scale object detection model: EfficientNetLite + FPN + nanohead. The backbone network is EfficientNetLite, and only the features from layers 2, 3, 5, and 6 of EfficientNetLite are used as features from EfficientNetLite to the dilated FPN encoder. The original image resolution is 1280x720p. When feeding it to EfficientNetLite, the input image is resized to 720x720, and the output layer parameters are adjusted from 128 to 512. Compared to the original network's 256x256 image size, the 720x720 resolution maximizes the preservation of the chicken's features, improving recognition accuracy.

[0049] The second embodiment of the present invention provides a non-contact automatic identification method for sick chickens in cages. Sick chickens that make abnormal calls are identified through voiceprint recognition. At the same time, sick chickens with abnormal behaviors and abnormal combs are discovered through image recognition. Sick chickens are identified from multiple angles. Therefore, this embodiment not only has the beneficial effects described in the first embodiment, but also greatly improves the recognition rate of sick chickens.

[0050] Based on the same inventive concept, the present application provides non-contact automatic identification systems corresponding to embodiment 1 and embodiment 2.

[0051] Implementation Three

[0052] See also Figures 1 to 12 A non-contact automatic identification system for sick chickens in cages includes an acoustic sensor, a memory, and a processor; the acoustic sensor collects audio data near the chicken cage, and the memory stores instructions, which are suitable for being loaded by the processor and executed by voiceprint detection steps: collecting audio data near the chicken cage by the acoustic sensor, and then using STFT to convert the audio data into a spectrogram; inputting the spectrogram into a trained first target detection network, and locating the calling area in the spectrogram; extracting features of each calling area from the spectrogram one by one, and then inputting the features into a trained convolutional neural network for voiceprint classification, outputting the calling type corresponding to each calling area, and issuing a sick chicken alarm if the sound contains chicken coughing.

[0053] More preferably, the size of the call area image obtained by feature extraction is matched to the original size of the spectrogram, and then input into the trained convolutional neural network.

[0054] Preferably, the system further comprises a camera, which captures images of the upper portion and the bottom portion of the chicken cage, and the instructions are loaded by the processor and further execute an image detection step: the captured upper and bottom images are input into a trained second target detection network; the second target detection network recognizes the chicken head and comb from the upper image, outputs a recognition image marked with the position of the chicken head and comb frame, and determines whether the comb is atrophied based on the comb feature value, and if so, issues a sick chicken alarm; at the same time, the second target detection network further identifies the comb, and if it is a comb with white spots, issues an alarm; the bottom image is input into a trained third target detection network, and an image containing the chicken body is identified from the bottom image, and outputs a bottom image marked with the position of the chicken body; the historical bottom image of the chicken cage is extracted, and if there are chickens in the historical bottom image, the chicken body position is compared to determine whether the chicken body has moved, and if not, issues a sick chicken alarm, and if so, it is determined to be a healthy chicken; if there is no chicken body in the historical bottom image, it is determined to be a healthy chicken.

[0055] The specific process of judging whether the comb is shrunk based on the comb characteristic value is as follows: analyzing the recognition image, binding the comb and the chicken head one by one: judging whether there are more than one comb positions intersecting with the same chicken head position, if not, binding the chicken head and the comb; if so, calculating the intersection value of each comb and the chicken head one by one, and then calculating the ratio of the intersection value to the comb area; if the ratio is less than the preset intersection threshold, judging that the comb and the chicken head have no binding relationship; if the ratio is greater than the preset intersection threshold, taking the comb with the highest ratio and binding it to the chicken head; calculating the comb characteristic value: the comb characteristic value is the ratio of the comb size to the chicken head size, if the ratio is less than the preset characteristic threshold, judging that the comb is shrunk, and issuing a sick chicken alarm. The comb characteristic value is multidimensional data, including the ratio of the comb size to the chicken head size and a combination of any of the following items: chicken head size, comb size, comb length-to-width ratio, and chicken head length-to-width ratio; after the multidimensional data is standardized using standard deviation, it is input into a trained LSTM neural network, and the LSTM neural network analyzes the data to identify whether the comb is atrophied. If so, a sick chicken alarm is issued.

[0056] For the implementation steps not detailed in this embodiment 3, please refer to embodiment 1 and embodiment 2. This embodiment 3 has the same beneficial effects as embodiment 1 and embodiment 2.

[0057] The above description is only a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A non-contact automatic identification method for sick chickens in cages, characterized by: The system includes the following steps: collecting audio data near the chicken coop using an acoustic sensor, converting the audio data into a spectrogram using STFT; inputting the spectrogram into a trained first target detection network to locate the calling area in the spectrogram; cutting out each calling area from the spectrogram one by one, and then inputting the trained convolutional neural network for voiceprint classification, outputting the corresponding calling type of each calling area. If the call type contains a chicken coughing sound, a sick chicken alarm is issued; The process also includes an image detection step: using a camera to capture images of the upper and lower parts of the chicken coop, and then inputting them into a trained second object detection network; The second object detection network identifies the chicken head and comb from the upper and lower images, outputs an identification image marked with the positions of the chicken head and comb frames, and determines whether the comb is atrophied based on the comb feature value. If so, a sick chicken alarm is issued. At the same time, the second object detection network further identifies the comb and issues an alarm if it is a comb with white spots. The bottom image is input into a trained third object detection network, and an image containing a chicken body is identified from the bottom image, and a bottom image with the chicken body position marked is output; historical bottom images of the chicken cage are extracted, and if a chicken is present in the historical bottom image, the chicken body position is compared to determine whether the chicken body has moved. If not, a sick chicken alarm is issued; if so, the chicken is determined to be healthy; if no chicken body is present in the historical bottom image, the chicken is determined to be healthy; The specific process of judging whether the comb is atrophied based on the comb characteristic value is as follows: Analyze the recognition image and bind the combs and chicken heads one by one: determine whether there is more than one comb position that intersects with the same chicken head position; if not, bind the chicken head and comb; if so, calculate the intersection value of each comb and the chicken head one by one, and then calculate the ratio of the intersection value to the area of the comb; if the ratio is less than a preset intersection threshold, determine that there is no binding relationship between the comb and the chicken head; if the ratio is greater than the preset intersection threshold, take the comb with the highest ratio and bind it to the chicken head; Calculate the comb characteristic value: the comb characteristic value is the ratio of the comb size to the chicken head size. If the ratio is less than the preset characteristic threshold, it is determined that the comb is atrophied and a sick chicken alarm is issued.

2. The non-contact automatic identification method for sick chickens in cages according to claim 1, characterized in that: The size of the cropped call area image is matched to the original size of the spectrogram, and then input into the trained convolutional neural network.

3. The non-contact automatic identification method for sick chickens in cages according to claim 1, characterized in that: The comb characteristic value is multidimensional data, including the ratio of the comb size to the chicken head size and a combination of any of the following items: chicken head size, comb size, comb length-to-width ratio, and chicken head length-to-width ratio; after the multidimensional data is standardized using standard deviation, it is input into a trained LSTM neural network, and the LSTM neural network analyzes the data to identify whether the comb is atrophied. If so, a sick chicken alarm is issued.

4. The non-contact automatic identification method for sick chickens in cages according to claim 1, characterized in that: The bottom image is input into a trained third target detection network, an image containing a chicken body is identified from the bottom image, and the position of the chicken body is located. The chicken body position comparison is specifically as follows: extracting the position information of the chicken body in the current bottom image, extracting the historical bottom image of the chicken in the cage, if the chicken body position is located in the historical bottom image, then extracting the position information of the chicken body in the historical bottom image, calculating the intersection value IoU of the current position information and the historical position information, and judging whether there has been a chicken at the position for a long time based on the intersection value IoU. If the intersection value IoU is greater than a preset intersection threshold, the overlap is high and it is judged to be a sick chicken; otherwise, it is judged to be a healthy chicken. If the chicken body position is not located in the historical bottom image, then the chicken in the current recognition image is judged to be a healthy chicken.

5. A non-contact automatic identification system for sick chickens in cages, characterized by: The system comprises an acoustic sensor, a memory and a processor; the acoustic sensor collects audio data near a chicken coop, and the memory stores instructions suitable for being loaded by the processor and executing the voiceprint detection steps: An acoustic sensor collects audio data near the chicken coop, then uses STFT to convert the audio data into a spectrogram. The spectrogram is input into a trained first target detection network to locate the calling area in the spectrogram. Each calling area is cut out from the spectrogram one by one and then input into a trained convolutional neural network for voiceprint classification. The corresponding call type for each calling area is output, and if a chicken coughs, a sick chicken alarm is issued. The system further includes a camera, which collects images of the upper portion and the bottom portion of the chicken cage. The instructions are loaded by the processor and further perform the image detection step: Input the collected upper and bottom images into the trained second object detection network; The second object detection network identifies the chicken head and comb from the upper and lower images, outputs an identification image marked with the positions of the chicken head and comb frames, and determines whether the comb is atrophied based on the comb feature value. If so, a sick chicken alarm is issued. At the same time, the second object detection network further identifies the comb, and if it is a comb with white spots, an alarm is issued. The bottom image is input into a trained third object detection network, and an image containing a chicken body is identified from the bottom image, and a bottom image with the chicken body position marked is output; historical bottom images of the chicken cage are extracted, and if a chicken is present in the historical bottom image, the chicken body position is compared to determine whether the chicken body has moved. If not, a sick chicken alarm is issued; if so, the chicken is determined to be healthy; if no chicken body is present in the historical bottom image, the chicken is determined to be healthy; The specific process of judging whether the comb is atrophied based on the comb characteristic value is as follows: Analyze the recognition image and bind the combs and chicken heads one by one: determine whether there is more than one comb position that intersects with the same chicken head position; if not, bind the chicken head and comb; if so, calculate the intersection value of each comb and the chicken head one by one, and then calculate the ratio of the intersection value to the area of the comb; if the ratio is less than a preset intersection threshold, determine that there is no binding relationship between the comb and the chicken head; if the ratio is greater than the preset intersection threshold, take the comb with the highest ratio and bind it to the chicken head; Calculate the comb characteristic value: the comb characteristic value is the ratio of the comb size to the chicken head size. If the ratio is less than the preset characteristic threshold, it is determined that the comb is atrophied and a sick chicken alarm is issued.

6. The non-contact automatic identification system for sick chickens in cages according to claim 5, characterized in that: The size of the cropped call area image is matched to the original size of the spectrogram, and then input into the trained convolutional neural network.

Images