A non-contact automatic identification method and system for weak and dying chickens in cages
By using an infrared camera to collect images of the bottom of the chicken coop and combining them with a target detection model and historical location comparison, the problem of high equipment cost, high false alarm rate and cumbersome operation in the existing technology for identifying dead and weak chickens is solved, and efficient and accurate identification of dead and weak chickens is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUZHOU MUJILANG INTELLIGENT TECH CO LTD
- Filing Date
- 2022-08-26
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for identifying dead or weak chickens in chicken houses suffer from problems such as high equipment costs, high false alarm rates, cumbersome operation, and low accuracy, especially in low-light environments where it is difficult to effectively distinguish between live and dead chickens.
Infrared cameras are used to capture images of the bottom of chicken coops. A trained target detection model is used to identify dead or weak chickens, and historical location comparison is used for further confirmation. The IoU algorithm is combined to determine the overlap of chickens, thereby improving the accuracy of identification.
It enables efficient and accurate identification of dead and weak chickens in low-light environments, reducing equipment costs, improving identification accuracy and efficiency, and reducing the waste of human resources.
Smart Images
Figure CN115457525B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of non-contact detection of chickens in cages, and more particularly to a non-contact automatic identification method and system for dead or weak chickens in cages. Background Technology
[0002] Chicken farms raise thousands of chickens in their coops, with a large number of chickens in each coop. The coops are dimly lit. If manual inspection of dead and weak chickens is used, it will not only consume a lot of manpower, but also present problems such as difficulty in identification due to the large number of chickens in low light and the fact that sick and live chickens are not much different in appearance.
[0003] There are two existing non-contact methods for identifying dead chickens: 1. Using an infrared temperature scanner to scan the chickens and obtain thermal images of the cage. An alarm is triggered if the number of chickens with temperatures above 35°C is less than the total number of chickens in the cage. Since the temperature of a dead chicken one day after death differs from that of a healthy chicken by only about 3°C, this method requires high precision from the infrared scanner, and high-precision infrared scanners must be stationary. Mobile infrared scanners are difficult to improve in terms of precision and are very expensive. Furthermore, the high density of chickens in the cage can lead to obstruction and a reduction in the number of chickens that can be identified. Therefore, this method is rarely used in practice due to its high equipment cost and high probability of false alarms. 2. Attaching a leg band to each chicken. A detection unit collects the chicken's movement trajectory data through the leg band, measures the maximum displacement and three-dimensional total variance, and uses machine learning methods to identify the chicken's condition, thus distinguishing between dead and sick chickens. In this technical solution, each chicken needs to be fitted with a leg band in advance. The leg band needs to contain a sensor that can provide movement acceleration data and a battery. Therefore, for a chicken house with thousands of chickens, it is time-consuming and labor-intensive to operate and maintain thousands of leg bands to work properly, resulting in poor application effect. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to provide a non-contact automatic identification method for dead or weak chickens in cages. This method involves acquiring images of the bottom of the chicken cage, automatically identifying obviously dead or weak chickens in the bottom images using a trained target detection model, and further identifying dead or weak chickens by comparing historical chicken locations. This improves the accuracy and efficiency of non-contact identification of sick or weak chickens.
[0005] To achieve the above objectives, the present invention employs the following three technical solutions:
[0006] Technical Solution 1
[0007] A non-contact automatic identification method for dead or weak chickens in a cage includes the following steps: a camera periodically acquires bottom images of the chickens' feet in the cage; the bottom images are input into a trained target detection model, which identifies bottom images containing chickens; the target detection model further identifies whether the chickens in the bottom images are dead or weak. If so, they are determined to be dead or weak, and an alarm is issued; if not, they are determined to be suspected dead or weak, and historical bottom images are extracted. If chickens are present in the historical bottom images, their positions are compared to determine if they have moved. If not, they are determined to be dead or weak, and an alarm is issued; if so, they are determined to be live chickens. If no chickens are present in the historical bottom images, they are determined to be live chickens.
[0008] More preferably, after the bottom image is input into the target detection model, a recognition image is output. The chicken's position is located in the recognition image. The chicken position comparison is specifically as follows: extract the chicken's position information in the currently output recognition image, extract the historical recognition images of the chickens in the cage, if the chicken's position is located in the historical recognition image, then extract the chicken's position information in the historical recognition image, calculate the intersection value (IoU) between the current position information and the historical position information, and determine whether the position has been occupied by chickens 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 the chicken is determined to be a dead or weak chicken; otherwise, it is determined to be a live chicken. If the chicken's position is not located in the historical recognition image, then the chicken in the current recognition image is determined to be a live chicken.
[0009] More preferably, the steps for establishing a training set during the training of the target detection model are as follows: collecting bottom images containing chickens, labeling the bottom images according to the state of the chickens in the bottom images, and generating label images, specifically: 1) labeling bottom images containing confirmed dead chickens as dead chickens; 2) labeling bottom images containing live chickens as live chickens; 3) labeling bottom images containing suspected dead or weak chickens as live chickens; using the label images to create a training set, a test set, and a validation set; and using the training set, test set, and validation set for the training phase, testing phase, and validation phase of the target detection model, respectively.
[0010] More preferably, the camera is an infrared camera.
[0011] More preferably, the target detection model consists of an EfficientNetLite0 backbone network, an FPN extended encoder, and a nanohead head.
[0012] Technical Solution Two
[0013] A non-contact automatic identification system for dead chickens in a cage includes a camera, a processor, and a memory. The memory stores a computer program that can be run by the processor. When the processor runs the computer program, it controls the camera to capture images from the bottom, thus implementing the steps described in technical solution one.
[0014] Technical Solution 3
[0015] A computer-readable storage medium is characterized in that: the computer-readable storage medium stores a computer program that can be executed by a processor, and when the processor runs the computer program, it controls a camera to capture bottom images, thereby implementing the steps of the method described in Scheme 1.
[0016] The present invention has the following beneficial effects:
[0017] 1. The present invention provides a non-contact automatic identification method and system for dead or weak chickens in cages. Based on the characteristics of sick or weak chickens, the system uses a camera to collect images of only the bottom of the chicken cage, eliminating the influence of high chicken density in the cage on the identification of sick or weak chickens, reducing the difficulty of identification, and improving the recognition accuracy of the target detection model.
[0018] 2. The present invention provides a non-contact automatic identification method and system for dead or weak chickens in cages, which identifies obviously sick or dead chickens quickly and accurately by using a trained target detection model.
[0019] 3. The present invention provides a non-contact automatic identification method and system for dead or weak chickens in cages. By comparing the current location of suspected dead chickens with the historical locations of chickens, the method further identifies dead or weak chickens, greatly improving the detection rate of dead or weak chickens.
[0020] 4. The present invention provides a non-contact automatic identification method and system for dead or weak chickens in cages. It makes full use of the chicken positions framed in the identification image output by the target detection model, combines them with the IoU algorithm, and judges the overlap between the current chicken and the historical chickens based on the intersection value IoU, thereby further identifying the dead chickens from the suspected dead chickens. It is simple and fast.
[0021] 5. The present invention provides a non-contact automatic identification method and system for dead or weak chickens in cages, which uses an infrared camera to enhance image capture capabilities under low light conditions and obtain clear images.
[0022] 6. This invention provides a non-contact automatic identification method and system for dead or weak chickens in cages. It collects and utilizes a target detection model to identify chickens in images at the bottom of the cage, avoiding problems such as high chicken density and overlapping chickens making identification difficult in overall cage images. This improves the accuracy of the target detection model in identifying obviously dead or weak chickens. Furthermore, by comparing the current chicken position with historical chicken positions, it further identifies dead or weak chickens from among suspected dead or weak chickens, significantly improving the detection rate and accuracy of non-contact identification of dead or weak chickens. This not only frees up manpower but also greatly enhances the efficiency of dead or weak chicken detection. The detection results have high accuracy and high robustness to different breeds of chickens. Attached Figure Description
[0023] Figure 1 A flowchart of a non-contact automatic identification method for dead or weak chickens in a cage according to the present invention;
[0024] Figure 2 and Figure 3 The image shows a clearly dead chicken captured by the camera of this invention.
[0025] Figure 4 Images of live chickens captured by the camera of this invention;
[0026] Figure 5 and Figure 6 Images of suspected dead chickens captured by the camera of this invention;
[0027] Figure 7 This is another flowchart of a non-contact automatic identification method for dead or weak chickens in a cage according to the present invention. Detailed Implementation
[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0029] Example 1
[0030] In existing poultry farms, chicken cages are typically equipped with feed troughs at the front. The bottom of the cages is sloped and includes an egg collection trough located directly below the feed trough. After a chicken lays an egg, the egg rolls down the sloped bottom into the trough due to gravity, facilitating egg collection. Because the feed trough is deep, most of the chickens' bodies are obscured in the bottom images. Most bottom images only show chicken feet, eggs, the feed trough, and the egg collection trough. A smaller portion of the bottom images may also include live chickens lying down, as well as dead or weak chickens. Dead chickens may fall to the bottom of the cage or slide to the egg collection trough due to the slope. Weak chickens may also lie prone at the bottom of the cage or near the egg collection trough due to their limited body control. Therefore, this invention focuses solely on analyzing images of dead or weak chickens at the bottom of the cage, significantly reducing the difficulty of non-contact identification of dead or weak chickens.
[0031] The specific implementation of this invention is as follows:
[0032] Please see Figure 1 A non-contact automatic identification method for dead or weak chickens in cages includes the following steps: First, a camera periodically collects bottom images of the chickens' feet at a preset interval. The camera can be an infrared camera to collect infrared images. Since the illumination in the chicken coop is low, and the light is even dimmer at the chicken's feet area below the feed trough, an infrared camera can capture clear images in low light conditions. Furthermore, since this invention requires color recognition during image identification, using an infrared camera is a preferred choice. The "period" generally refers to collecting bottom images at regular intervals throughout the day, for example, every 6-10 hours. This can be achieved by collecting bottom images of the chicken cage at 8:00 AM using a camera on an automatic inspection device, and then again at 4:00 PM using the same camera. Next, the bottom image is input into the trained target detection model. The target detection model first identifies bottom images containing chickens, and then further identifies whether the chickens in the bottom image are dead or weak. If so, it is determined to be a dead or weak chicken and an alarm is issued. If not, it is determined to be a suspected dead or weak chicken, and historical bottom images are extracted. If chickens are present in the historical bottom images, the chicken positions are compared to determine whether the chickens have moved. If not, it is determined to be a dead or weak chicken and an alarm is issued. If so, it is determined to be a live chicken. If no chickens are present in the historical bottom images, it is determined to be a live chicken.
[0033] In order to improve the recognition accuracy of the target detection model, this embodiment also performs the following steps:
[0034] First, create a training sample set. Collect bottom images containing chickens, and label the bottom images according to the state of the chickens in the bottom images to generate labeled images. Specifically: 1) Label bottom images containing chickens that are definitely dead as dead chickens, such as... Figure 2 and Figure 3 As shown, Figure 2 The chickens were already lying sideways with their legs up, and had slid onto the egg net. Figure 3 1) Dead chickens that have been left uncleaned for a long time after death, resulting in a dried-out state; 2) such as Figure 4 As shown, bottom images containing live chickens are marked as live chickens; 3) Bottom images containing suspected dead or weak chickens are marked as live chickens, such as... Figure 5 and Figure 6 As shown, Figure 5 The chicken in the middle is lying motionless and located in the egg net. Figure 6For chickens lying on the bottom of the coop inside the cage, these chickens may be dead or weak, requiring further clarification using historical bottom images. Training, testing, and validation sets are created using the labeled images; these sets are then used for the training, testing, and validation phases of the object detection model, respectively.
[0035] Secondly, the selection of the target detection model. This embodiment improves the original target detection model Nanodet. The improved structure consists of an EfficientNetLite0 backbone network, an FPN dilated encoder, and a nanohead. The EfficientNetLite0 backbone network extracts the basic features of the input image, the FPN dilated encoder fuses features from different levels to obtain new features, and finally the nanohead predicts the type and location of the target. Specifically: 1) The backbone network is changed from the original ShuffleNetV2 to EfficientNetLite0. The reason for this choice is that although EfficientNetLite0 will lose a few milliseconds in running speed, it has a better recognition effect on the recognition of chicken coop bottom images compared to the original ShuffleNetV2. The EfficientNetLite0 backbone network is a Convolutional Neural Network (CNN). When the bottom image is input to the CNN, the input size is adjusted from 256*256p in the original object detection model Nanodet to 720*720p. The 720*720p image can retain the chicken features to the maximum extent and maintain the original clarity of the bottom image. The original size of the bottom image captured by the camera is 1280*720. 2) The dilated encoder is changed from the original PAN to FPN. FPN consists of two parts: the first part is the bottom-up process, and the second part is the fusion process of top-down and lateral connections. Bottom-up process: In FPN, each stage corresponds to a feature pyramid level, and the last layer feature of each stage is selected as the feature of the corresponding level in FPN. For example, the last feature of conv2, conv3, conv5, and conv6 layers can be selected as the features of FPN, denoted as {C2, C3, C5, C6}. Top-down process and lateral connections: The top-down process upsamples the small feature map of the top layer to the feature of the previous stage. Figure 1The sample size is determined. Layer C6 first undergoes a 1x1 convolution to obtain M6 features. M6 features are then upsampled and combined with layer C5 through a 1x1 convolution to obtain M5 features. This process is repeated twice to obtain M3 and M2. The M-layer features are then subjected to a 3x3 convolution to obtain the final P2, P3, P5, and P6 layer features. The number of channels in layer M is set to 256 in this invention. The optimization method for the object detection model can be stochastic gradient descent (SGD), and a warmup is used during training to pre-adjust the learning rate.
[0036] In this embodiment, for bottom images with obvious dead chicken characteristics, the target detection model can quickly identify them. For bottom images with characteristics close to live chickens, this embodiment uses a method of comparing the current position with historical positions to further determine whether it is a dead or weak chicken. There are various methods for comparing the current position of the chicken with historical positions. Preferably, using the chicken position outlined in the recognition image output by the target detection model for comparison is not only fast but also has a higher comparison accuracy. Please refer to [link to relevant documentation]. Figure 7 Specifically: After the bottom image is input into the target detection model, a recognition image is output. The location of the chickens is located in the recognition image, and the location information of the chickens in the currently output recognition image is extracted. Historical recognition images of the chickens in the cage are extracted; generally, the previous recognition image of the chickens in the cage is extracted. If the location of the chickens is located in the historical recognition image, the location information of the chickens in the historical recognition image is extracted. The location information refers to the border information of the chickens' locations. The intersection value (IoU) between the current location information and the historical location information is calculated. Based on the intersection value (IoU), it is determined whether chickens have been present at that location for a long time. If the intersection value (IoU) is greater than a preset intersection threshold, the overlap is high, and the chickens are identified as dead or weak. Since normal live chickens are relatively active, it is rare for chickens to remain in the same position for a long time; therefore, they can be identified as dead or weak chickens. If the location of the chickens is not located in the historical recognition image, the chickens in the current recognition image are identified as live chickens. Generally, the intersection threshold can be set to 0.7. The intersection value IoU is calculated as the overlap rate between the borders of chickens in the current recognition image and the borders of chickens in the historical recognition images, that is, the ratio of their intersection to their union.
[0037] Another implementation method for comparing the current location of the chicken with its historical location is as follows: Extract historical bottom images of the chickens in the cage suspected of being dead or weak; extract a local image of the chicken's location from the currently acquired bottom image; and use an image similarity comparison algorithm to calculate the similarity between this local image and local images of the same location in the historical bottom images. If the similarity is greater than a preset similarity threshold, the chicken is determined to be dead or weak, and an alarm is issued. The implementation method for comparing the current location of the chicken with its historical location can also be achieved through deep learning algorithms or convolutional neural networks, which will not be elaborated upon here.
[0038] Please refer to Table 1, which is a comparative analysis of the experimental results of the accuracy of identifying dead and weak chickens in this invention with the experimental results of other target detection models.
[0039] Experimental input data: 100 images, including 20 images of dead or weak chickens;
[0040] Image recognition models: 1) SSD model; 2) Nanodet original object detection model (image input size is 320*320); 3) Nanodet original object detection model (image input size is 720*720); 4) The improved object detection model of this invention;
[0041] Experimental results: The accuracy rate of identifying dead or weak chickens is shown in Table 1:
[0042] Table 1:
[0043] method accuracy SSD model 90% Original Nanodet (320*320) 88% Original Nanodet (720*720) 92% This patent (720*720) 99%
[0044] The experimental data above show that the improved target detection model of this invention can achieve an accuracy rate of up to 99% in identifying dead or weak chickens.
[0045] This invention provides a non-contact automatic identification method for dead or weak chickens in cages. Utilizing the active nature of live chickens and their infrequent presence at the bottom of the cage, the method collects and uses a target detection model to identify chickens in images of the cage bottom. This avoids the problems of high chicken density and overlapping chickens in the overall cage image, improving the accuracy of the target detection model in identifying obviously dead or weak chickens. Furthermore, by comparing the current chicken position with historical chicken positions, this invention further identifies dead or weak chickens from among suspected dead or weak chickens, significantly improving the accuracy and detection rate of non-contact identification. Simultaneously, this non-contact automatic identification method enables non-contact detection and alarm of dead or weak chickens in the chicken coop, freeing up manpower and greatly enhancing the efficiency of dead or weak chicken detection. The detection results have high accuracy and robustness to different chicken breeds.
[0046] Example 2
[0047] A non-contact automatic identification system for dead chickens in a cage includes a camera, a processor, and a memory. The memory stores a computer program that can be run by the processor. When the processor runs the computer program, it controls the camera to capture images from the bottom and executes the steps described in Embodiment 1.
[0048] Example 3
[0049] A computer-readable storage medium storing a computer program executable by a processor, wherein when the processor runs the computer program, it controls a camera to capture bottom images and performs the steps of the method described in Embodiment 1.
[0050] The beneficial effects produced by Examples 2 and 3 are the same as those produced by Example 1.
[0051] The above description is merely a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A non-contact automatic identification method for dead or weak chickens in a cage, characterized in that: Includes the following steps: The camera periodically captures images of the bottom of the chickens' feet inside the cage according to a preset cycle; The bottom image is input into a trained target detection model. The target detection model identifies bottom images containing chickens and outputs a recognized image. The chicken's position is located in the recognized image. The target detection model further identifies whether the chicken in the bottom image is a dead or weak chicken. If so, it is determined to be a dead or weak chicken and an alarm is issued. If not, it is determined to be a suspected dead or weak chicken. Then, historical bottom images are extracted. If chickens exist in the historical bottom images, the chicken's position is compared to determine whether the chicken has moved. If not, it is determined to be a dead or weak chicken and an alarm is issued. If so, it is determined to be a live chicken. If no chickens exist in the historical bottom images, it is determined to be a live chicken. The chicken location comparison specifically involves: extracting the chicken location information from the currently output recognition image, extracting the historical recognition images of the chickens in the cage, and if the chicken location is located in the historical recognition image, then extracting the chicken location information from the historical recognition image, calculating the intersection value (IoU) between the current location information and the historical location information, and determining whether chickens have been present at the location 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 the chicken is identified as a dead or weak chicken; otherwise, the chicken is identified as a live chicken. If the location of the chicken is not found in the historical recognition image, then the chicken in the current recognition image is determined to be a live chicken.
2. The non-contact automatic identification method for dead or weak chickens in a cage according to claim 1, characterized in that: The steps for establishing the training set during the training of the object detection model are as follows: Collect bottom images containing chickens, and label the bottom images according to the status of the chickens in the bottom images to generate labeled images. Specifically: 1) Label bottom images containing confirmed dead chickens as dead chickens; 2) Label bottom images containing live chickens as live chickens; 3) Label bottom images containing suspected dead or weak chickens as live chickens. Use the labeled images to create a training set, a test set, and a validation set; The training set, test set, and validation set are used for the training, testing, and validation phases of the object detection model, respectively.
3. The non-contact automatic identification method for dead or weak chickens in a cage according to claim 1, characterized in that: The camera is an infrared camera.
4. The non-contact automatic identification method for dead or weak chickens in a cage according to claim 1, characterized in that: The target detection model consists of an EfficientNetLite0 backbone network, an FPN extended encoder, and a nanohead head.
5. A non-contact automatic identification system for dead or weak chickens in cages, characterized in that: The device includes a camera, a processor, and a memory. The memory stores a computer program that can be run by the processor. When the processor runs the computer program, it controls the camera to capture bottom images, thereby implementing the steps of the method as described in any one of claims 1 to 4.
6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that can be executed by a processor. When the processor runs the computer program, it controls the camera to capture bottom images, thereby implementing the steps of the method as described in any one of claims 1 to 4.