WO2022008576A1 - Détection de mammite clinique chez des mammifères laitiers - Google Patents

Détection de mammite clinique chez des mammifères laitiers Download PDF

Info

Publication number
WO2022008576A1
WO2022008576A1 PCT/EP2021/068780 EP2021068780W WO2022008576A1 WO 2022008576 A1 WO2022008576 A1 WO 2022008576A1 EP 2021068780 W EP2021068780 W EP 2021068780W WO 2022008576 A1 WO2022008576 A1 WO 2022008576A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
clots
milking
milk
detection device
Prior art date
Application number
PCT/EP2021/068780
Other languages
English (en)
Inventor
Igor VAN DEN BRULLE
Sofie PIEPERS
Glenn VAN STEENKISTE
Original Assignee
Universiteit Gent
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Universiteit Gent filed Critical Universiteit Gent
Publication of WO2022008576A1 publication Critical patent/WO2022008576A1/fr

Links

Classifications

    • A—HUMAN NECESSITIES
    • A01—AGRICULTURE; FORESTRY; ANIMAL HUSBANDRY; HUNTING; TRAPPING; FISHING
    • A01J—MANUFACTURE OF DAIRY PRODUCTS
    • A01J5/00—Milking machines or devices
    • A01J5/013—On-site detection of mastitis in milk
    • A01J5/0131—On-site detection of mastitis in milk by analysing the milk composition, e.g. concentration or detection of specific substances
    • A—HUMAN NECESSITIES
    • A01—AGRICULTURE; FORESTRY; ANIMAL HUSBANDRY; HUNTING; TRAPPING; FISHING
    • A01J—MANUFACTURE OF DAIRY PRODUCTS
    • A01J5/00—Milking machines or devices
    • A01J5/013—On-site detection of mastitis in milk
    • A01J5/0134—On-site detection of mastitis in milk by using filters or decanters
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/68—Food, e.g. fruit or vegetables

Definitions

  • the invention relates to the field of clinical mastitis in dairy mammals. More specifically it relates to methods and systems which are configured for automatically detecting the presence of clinical mastitis in dairy mammals.
  • the above objective is accomplished by a method and device according to the present invention.
  • a detection device for detecting one or more clots in milk which are indicative for clinical mastitis in dairy mammals.
  • the detection device comprises:
  • an acquisition interface configured for acquiring an image of a milk filter after or during pre-milking or milking
  • a processing device configured for executing a deep learning algorithm on the acquired image to determine the presence of one or more clots in the image.
  • the deep learning algorithm is trained for determining whether or not one or more clots, indicative for clinical mastitis, are present in the image.
  • the deep learning algorithm which is trained for determining whether or not one or more clots are present in the image, is able to distinguish the clot(s) from the rest of the image (background) including the other detriments on the image.
  • the deep learning algorithm does not necessarily need to recognize background items (e.g. detriments) to distinguish the clot(s) from the rest of the image.
  • inventions of the present invention relate to a method for determining the risk for clinical mastitis in dairy mammals.
  • the method comprises:
  • the deep learning algorithm is trained for determining whether or not one or more clots, indicative for clinical mastitis, are present in the image.
  • a deep learning algorithm for determining whether or not one or more clots, indicative for clinical mastitis, are present in an image, the deep learning algorithm comprising:
  • FIG. 1 shows a schematic block diagram of the basic building blocks of a detection device in accordance with embodiments of the present invention and of a milking machine comprising such a detection device.
  • FIG. 2 shows a flow chart of a method in accordance with embodiments of the present invention.
  • FIG. 3 shows a picture of a filter on which besides clots indicative for clinical mastitis also other detriments are present.
  • FIG. 6 shows a schematic overview of a milking process in an automatic milking system and possible locations to implement the image acquisition of a detection device according to embodiments of the present invention.
  • FIG. 7 shows a schematic overview of a pre-milking and a milking process in an automatic milking system and possible locations to implement the image acquisition of a detection device according to embodiments of the present invention.
  • FIG. 8 shows a schematic drawing of a deep learning algorithm in accordance with embodiments of the present invention.
  • FIG. 9 shows a possible architecture of the neural network.
  • FIG. 10 shows the results of the statistical parameters of an exemplary detection device in accordance with embodiments of the present invention in case of a balanced test.
  • FIG. 11 shows the results of the statistical parameters of an exemplary detection device in accordance with embodiments of the present invention in case of a realistic test with a prevalence of 3%.
  • FIG. 12 shows a residual network with a branched architecture with 2 different outputs, in accordance with embodiments of the present invention.
  • the detection device 100 comprises: an acquisition interface 110 configured for acquiring an image of a milk filter after or during pre-milking or milking, and a processing device 120 configured for executing a deep learning algorithm on the acquired image to determine the presence of one or more clots in the image.
  • the processing device may be a local processing device.
  • the acquisition interface may be part of the processing device.
  • the processing device may for example be an on board processing device such as an embedded microcontroller.
  • the processing device may be a server to which the images are sent. The latter is more flexible for implementing new algorithms and for correlating with other data.
  • a filter can be mounted per udder quarter (such as in dairy cows, buffaloes, camels) or udder halve (such as in sheep, goats, horses, donkeys) to detect clinical mastitis. As will be discussed later, this can be done during the pre-milking stage as well as during the milking stage.
  • the deep learning algorithm is trained for determining whether or not one or more clots (also referred to as flake(s)), indicative for clinical mastitis, are present in the image.
  • Such a method can be applied during the pre-milking (the process before the real milking to trigger the oxytocin reflex) as well as during the real milking process to detect one or more clots.
  • a method can be applied during the pre-milking or milking it can be avoided that milk from a cow with clinical mastitis is added to the bulk milk tank (which would be illegal).
  • the milk from a cow with clinical mastitis may for example be drained to a waste tank.
  • the processing device 120 may for example be configured for triggering a signal to the milking machine, based on the presence of one or more clots in the image, which allows the milking machine to separate the milk or to stop milking to avoid that mastitic milk is added to the bulk milk tank that contains the milk of the herd/farm to be collected by the milk buyer.
  • a part of the milk may be removed, e.g. for bacteriological examination of the milk.
  • the processing device 120 may be configured for storing the determined presence of one or more clots in an image and for comparing the presence of one or more clots in a new image with the determined presence of one or more clots in an earlier image to determine whether or not one or more new clots are present in the new image.
  • Different parameters may be stored when storing the determined presence of one or more clots in an image.
  • exemplary parameters such as number of clots, area of clot(s), clot volume, position of the clot(s), identification number of the dairy mammal (e.g. cow), lactation stage of the dairy mammal (e.g. cow), quarter/udder halve position, parity of the dairy mammal, dairy mammal breed, date and time of applying treatment, number of quarters/udder halves infected, types of bedding material, volume of detriments (e.g. bedding material), date and time of milking, date and time of detection of clinical mastitis, and identification numberof the milking machine and herd/farm can be taken into account.
  • specific parameters such as number of clots, area of clot(s), clot volume, position of the clot(s), identification number of the dairy mammal (e.g. cow), lactation stage of the dairy mammal (e.g. cow
  • filters not necessarily need to be changed after detecting the presence of one or more clots. Storing earlier results allows to keep track of differential changes in the clot(s) between images. This, moreover, allows to monitor diseased animals with clinical mastitis (e.g. recovery and severity of the disease) over time. Furthermore it is possible to do a retrospective analysis of the dairy mammal's (e.g. cow's) health status and to make a connection with other systems (such as milk production losses, electrical conductivity, Dairy Herd Improvement results, color of the milk (milk colored by blood), temperature of the dairy mammal (e.g.
  • the data of all these systems can be implemented in a final deep learning algorithm which is capable of interpreting all these data and acts as the final decision maker.
  • a detection device can recognize one or more clots without the need for taking a reference image of an earlier milking.
  • a reference image is required and needs to be taken in advance (e.g. to determine detriments).
  • fuzzy logic typically, a reference image is taken and examined for defects, scratches, etc. and thus intended to ensure that defect objects are excluded from subsequent analysis.
  • the reference image is not used to exclude defect objects since the deep learning algorithm is capable to do the object classification without this step.
  • the reference image allows to differentiate animals of different mastitis cases and do a follow up of the animals overtime.
  • a detection device can detect if new clot(s) are present, if the clot(s) are tilted, if they changed size, if they changed shape or position on the current image by comparing the clot(s) volume of the new image with the reference image.
  • regression is applied on the image. Therefore a neural network is particularly advantageous.
  • the image can, for example, be created via Stereovision (3D camera), Time of Flight camera, Structured light.
  • fuzzy logic it is not possible to define all features and rules by hand to calculate the volume of the object of interest on the acquired 3D images because every image is different, and hence there is an endless number of options. Moreover, calculation of regression via e.g. a Fuzzy logic algorithm is not possible since the outcome of Fuzzy logic is "Fuzzy" with a range between upper and lower boundaries. Thus, the estimation of the clot(s) volume will never be exact enough to compare between images. Also calculation of the clot(s) volume via for example 'support vector regression' will not provide as good results as a neural network since the inference time increases linear with the number of observations.
  • a deep learning algorithm in accordance with embodiments of the present invention that it allows to determine the clot(s). Firstly, if a clinical animal is detected, the filter does not necessarily need to be changed. This gives the farmer flexibility in the timing of the filter changes. Secondly, calculation of the animals clot(s) volume allows to monitor diseased animals with clinical mastitis (e.g. recovery and severity of the disease) over time and to estimate the severity of the disease. For example, if the clot(s) volume of a diseased animal is decreasing between consecutive milkings, the animal is recovering. Finally, if only a few clot(s) are present, increasing the number of milkings per day can be a sufficient treatment. This will contribute to a decrease in antibiotic use. On the other hand, treatment can be necessary when many clots in milk are detected.
  • clinical mastitis e.g. recovery and severity of the disease
  • the deep learning algorithm comprises a volume estimator network (to execute the volume regression) which uses tensors (containing information) from the convolutional layers and the second to last dense layer (after the last dropout layer (see e.g. FIG. 12)).
  • the data from the convolutional layers can be used to reconstruct a volumetric image using a decoder network architecture. Since the data at the second last dense layer mainly contains the information of which parts of the image contain clot(s) and parts don't, this information can be used to be fed to the decoder network architecture to recreate a volumetric image out of the data stream.
  • this approach resembles an auto encoder network architecture with the initial part of the clot(s) detection algorithm as the encoder and the separate data stream with convolutional layers as the decoder.
  • the dense layer at the start of the decoder layers is used for regression of the data coming from the dense and convolutional layers.
  • a deep learning algorithm for determining whether or not one or more clots, indicative for clinical mastitis, are present in an image.
  • a detection device may be configured for combining the results of the deep learning algorithm with one or more of these characteristics.
  • a detection device in addition, it is possible to add an extra sensor and connect this with a detection device according to embodiments of the present invention to detect swelling, redness and temperature (e.g. via infrared) of the teats. Also these characteristics may be combined with the results of the deep learning algorithm. It may also be possible to add and connect an image acquisition of a detection device according to embodiments of the present invention to detect watery milk which may be an indication of clinical mastitis and which would use a similar algorithm as the embodiments of present invention. This can for example be implemented in the pre-milking system of an automatic milking system. In embodiments of the present invention the deep learning algorithm may be trained for determining whether the milk is watery.
  • the custom made residual network comprises residual blocks such as two- dimensional matrix convolutions with shortcuts (e.g. max pooling) between convolutional layers. Residual thereby refers to the shortcuts between several blocks of convolutional layers. These layers may be designed of 2D convolutional layers (Conv2D).
  • the convolutional matrix parameters may be obtained empirically or using genetic algorithms such as NSGA-II.
  • Keras may be used as application programming interface (API). Keras provides a Conv2D class which represents a 2D convolutional layer.
  • API application programming interface
  • the max pooling and the "filter” will compress the images and only retain the features.
  • the residual neural network comprises a branched architecture connecting multiple stages of architecture, resulting in 2 types of output variables (classifier and regression) specific for a detection device according to embodiments of the present invention.
  • FIG. 12 shows a possible residual neural network with a branched architecture and 2 different outputs (in the figure BN is the acronym for branch normalization, RELU is the acronym for rectified linear unit, and 2D conv is the acronym for 2D convolutional layer).
  • BN is the acronym for branch normalization
  • RELU is the acronym for rectified linear unit
  • 2D conv is the acronym for 2D convolutional layer.
  • Training the deep learning algorithm may imply optimizing it with thousands of pictures (e.g. more than thousand, or even more than ten thousand) of images of a matrix or surface comprising one or more clots and images without one or more clots before it can start classifying new clot pictures with relative accuracy.
  • the training set preferably comprises images of milk filters with milk (with or without one or more clots) from different dairy mammals (e.g. cows) to develop/train a good deep learning model.
  • Deep learning algorithms may have a specificity of 99.9% and a sensitivity of 98% and a positive predicting value of 99% and a negative predicting value of 99.9%.
  • FIG. 10 shows the results of a balanced test and
  • FIG. 11 shows the results of a realistic test with a prevalence of 3% obtained using an exemplary deep learning algorithm in accordance with embodiments of the present invention.
  • CM clinical mastitis
  • no CM clinical mastitis
  • the algorithm is capable to generalize or interpolate from the training dataset to real world samples.
  • Clots may for example be described and computed as usually light, tending to have a circular outer contour and the size of a clot in its largest dimension ranges from approx. 0.1 mm up to several millimeters. Hence, if a clot (which is very heterogeneous) does not match this description, this means that the clot will not be detected by a fuzzy logic based algorithm.
  • learning may be supervised or (partially) unsupervised.
  • the relationship between the input pictures and decision label may be assessed using an integrated gradients method.
  • the training dataset can be optimized for improved generalization of the neural network.
  • the integrated gradients method thus helps to optimize the training dataset for improved generalization of the neural network. It is an advantage of embodiments of the present invention that the integrated gradients method allows humans to see what the deep learning algorithm actually uses for its final decision.
  • the advantages of both deep learning (the generalization and robustness of the algorithm) and fuzzy logic (the interpretability for humans) are combined.
  • a method according to embodiments of the present invention comprises an integrated gradients method which is configured for reporting to a user whether the deep learning algorithm uses non-relevant features for determining whether there is clinical mastitis or not, and wherein the method is configured for receiving an adjusted set of images for the training.
  • the integrated gradients method allows the user to detect if the deep learning algorithm actually uses non-relevant features for its final decision, and hence allows the user to subsequently adjust the training dataset without these non-relevant features.
  • Such a method in accordance with embodiments of the present invention, thus displays the state of the system such that the system can be adjusted.
  • the integrated gradients method may display how much a local pixel attributes to the model overall output prediction (negative or positive) by giving it a score.
  • FIG. 13 An example of the integrated gradients method, applied to the training dataset, is illustrated in FIG. 13 (positive example), 14 (negative example), 15 (negative example).
  • the left pictures of these figures show the original image and the right pictures show the result after applying the integrated gradients method.
  • clots are present on the image.
  • white spots are displayed (at the image on the right). These white spots present the focus of the deep learning algorithm. It can be seen that the algorithm focuses on the edges of the clots. The algorithm is clearly not focusing on the detriments, hence the algorithm is doing a correct interpretation of the image classification.
  • the integrated gradients method displays features that are considered important for an interpretation, also empty areas receive a higher (negative) score. Detriments, on the other hand, receive no score at all. In this example the clots are striped on the image.
  • FIG. 14 illustrates a negative example.
  • no clots are present on the image.
  • the white spots in the right figure are not focusing on the detriments, but are focusing on all non-detriment parts (negative score). This means that the algorithm is looking at the right spots to detect clot(s) and detriments are not scored.
  • a deep learning neural network makes an evidence based decision based on its education via the training part (the used samples for the training part are real clinical mastitis samples of different animals) and its generalization and interpolation capabilities. Fuzzy logic, on the other hand, makes an educated guess based on the human predefined features and decision rules. In fuzzy logic no real clinical mastitis samples are used to develop the algorithm. Therefore fuzzy logic is less robust and less generalized. Due to the parallel structure of a deep learning neural network, in accordance with embodiments of the present invention, this robustness is even further enhanced. The dual pathway ensures redundancy in case one of the two pathways produces an unreliable result.
  • a deep learning algorithm in accordance with embodiments of the present invention, focuses on clot(s) in all circumstances in contrast to fuzzy logic which has to take into account what the circumstances are.
  • Fuzzy logic is only based on logical relationships which are within the fuzzy logic framework mathematically implemented as minimum and maximum operators, limiting the complexity and learning-capabilities of the decision framework.
  • fuzzy logic is considered less robust to noise (e.g. sawdust) in the picture.
  • deep learning which is capable to model highly complex (non-linear) functional relationships due to the multi-layer architecture of non-linear activation functions applied on linear combinations of the node-values of the preceding network layer. Therefore, deep learning is able to construct its own (complex) features from the image dataset, including those that could be overlooked by a human expert. Hence, learning complex relationships that are not present as prior knowledge, is therefore advantageously done through neural networks (deep learning).
  • a detection device comprises a branched architecture resulting in 2 types of output variables (classifier + regression). This is advantageous, compared to sequential architectures as these are not capable to handle the classification and regression output variables simultaneously. Moreover, through the parallel structure residual neural network according to embodiments of the present invention is more robust as opposed to the often used sequential architectures.
  • the neural network may be programmed using a neural network library such as Keras which is a high-level application programming interface of Tensorflow. Keras may contain the training whereas Tensorflow is the backend of the algorithm. The invention is, however, not limited thereto. Various optimization and regularizations may be added to improve the model.
  • Regularizers help to avoid overfitting of the network and improve the generalization capabilities of the network.
  • regularizers such as dropout, L2: ridge regression, data augmentation and batch normalization techniques may be used to avoid the problem of overfitting.
  • Images may be random adjusted for each image of the dataset. This may be achieved by turning, by slight lighting variations, by in and out zooming, by moving, and by mirroring.
  • the acquisition interface may provide the necessary functions for controlling a camera such that at least some of the above mentioned operations can be performed.
  • a deep learning algorithm which is trained for determining whether or not one or more clots are present in the image, is capable of discerning new detriments (e.g. a new type of bedding material) from one or more clots.
  • this algorithm is capable of discerning objects similar to clot(s) (such as remainder of internal teat sealants or air bubbles at the filter) from clot(s).
  • At least one camera 230 configured for taking pictures of the filter
  • each teat cup may be connected with a different filter. This allows to diagnose the milk from the different udder quarters/halves separately.
  • the teat cups for the different teats may be connected with a shared filter.
  • a teat cup may be connected to each teat.
  • a separate teat cup may be present for each quarter of the udder or udder halve.
  • Each teat cup may be separately connected to a separate receiver or the separate teat cups may be connected together in a milk claw which is connected to a share receiver.
  • pre-milking and milking may use the same teat cups, milking machine tubes, and filters.
  • the detection device 100 or an additional detection device is configured for retrieving pictures of the filter connected with the pre-milking machine tube, and for determining the presence of one or more clots on this filter.
  • a detection device can be installed on an existing twin filter system or on a milking machine without twin filters.
  • a twin filter system allows to automatically switch between milk filters. This is typically done when there is a cleaning of the milking machine. This allows to continue milking with a clean filter.
  • the cluster may comprise of a clawpiece and four teat cups each with its own shell and liner, short milk and short pulsation tube.
  • FIG. 4 shows a schematic overview of the milking process in a conventional milking system and possible locations to implement the image acquisition of a detection device according to embodiments of the present invention.
  • FIG. 5 shows a schematic overview of a pre-milking process in an automatic milking system and possible locations to implement the image acquisition of a detection device according to embodiments of the present invention.
  • FIG. 7 shows a schematic overview of a pre-milking and a milking process in an automatic milking system and possible locations to implement the image acquisition of a detection device according to embodiments of the present invention.
  • inventions of the present invention relate to a method for detecting one or more clots in milk which are indicative for clinical mastitis in dairy mammals, the method comprises:
  • the deep learning algorithm is trained for determining whether or not one or more clots, indicative for clinical mastitis, are present in the image.
  • the method may comprise training 330 the deep learning algorithm using a training set of images of which the absence or presence of one or more clots is confirmed. It is, thereby, an advantage that by training the deep learning algorithm it becomes automatically capable of determining the most optimal parameters.
  • the training 330 may also be applied during the milking or pre-milking by manually confirming the presence or absence of one or more clots in an image. It is, thereby, an advantage that the algorithm has the ability to easily adjust at herd/farm or individual level via its learning part.
  • the method may comprise determining an area, and/or volume, and/or position, and/or shape and/or color of one or more clots present in the image.

Landscapes

  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Animal Husbandry (AREA)
  • Environmental Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Chemical & Material Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Analytical Chemistry (AREA)
  • Image Analysis (AREA)

Abstract

Dispositif de détection pour détecter un ou plusieurs caillots dans le lait qui indiquent une mammite clinique chez des mammifères laitiers. Le dispositif de détection comprend : une interface d'acquisition configurée pour acquérir une image d'un filtre à lait après ou pendant la pré-traite ou la traite ; un dispositif de traitement configuré pour exécuter un algorithme d'apprentissage profond sur l'image acquise pour déterminer la présence d'un ou de plusieurs caillots dans l'image, l'algorithme d'apprentissage profond comprenant un réseau de neurones artificiels résiduel comprenant des blocs résiduels avec des raccourcis entre des couches de convolution et configuré pour extraire des caractéristiques hors de l'image ; des couches denses configurées pour une classification finale, sur la base des caractéristiques extraites, s'il y a une mammite clinique ou non.
PCT/EP2021/068780 2020-07-07 2021-07-07 Détection de mammite clinique chez des mammifères laitiers WO2022008576A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
EP20184397.6 2020-07-07
EP20184397 2020-07-07

Publications (1)

Publication Number Publication Date
WO2022008576A1 true WO2022008576A1 (fr) 2022-01-13

Family

ID=71523016

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2021/068780 WO2022008576A1 (fr) 2020-07-07 2021-07-07 Détection de mammite clinique chez des mammifères laitiers

Country Status (1)

Country Link
WO (1) WO2022008576A1 (fr)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070289364A1 (en) * 2004-01-23 2007-12-20 Magnus Wiethoff Method and Device for Determining the Quality of Milk Produced by Machine Milking
WO2008093344A1 (fr) * 2007-02-01 2008-08-07 E.N.G.S. Systems Ltd. Système pour détecter des particules dans un fluide lacté tel que le lait
US20080259351A1 (en) * 2004-03-24 2008-10-23 Magnus Wiethoff Device and Method for Recognizing Particles in Milk
WO2009072870A1 (fr) * 2007-12-07 2009-06-11 Qlip N.V. Procédé de détection de pathogènes de mastite dans un échantillon de lait

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070289364A1 (en) * 2004-01-23 2007-12-20 Magnus Wiethoff Method and Device for Determining the Quality of Milk Produced by Machine Milking
US20080259351A1 (en) * 2004-03-24 2008-10-23 Magnus Wiethoff Device and Method for Recognizing Particles in Milk
WO2008093344A1 (fr) * 2007-02-01 2008-08-07 E.N.G.S. Systems Ltd. Système pour détecter des particules dans un fluide lacté tel que le lait
WO2009072870A1 (fr) * 2007-12-07 2009-06-11 Qlip N.V. Procédé de détection de pathogènes de mastite dans un échantillon de lait

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
XIAOLONG LIU ET AL: "Recent progress in semantic image segmentation", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 20 September 2018 (2018-09-20), XP081195179, DOI: 10.1007/S10462-018-9641-3 *

Similar Documents

Publication Publication Date Title
CN110839557B (zh) 母猪发情监测方法、装置、系统、电子设备及存储介质
US20230129428A1 (en) System for high performance, ai-based dairy herd management and disease detection
US9922242B2 (en) Systems and methods for predicting the outcome of a state of a subject
US8794182B2 (en) System and a method for controlling an automatic milking system
KR20240053658A (ko) 물고기 부피 추정, 중량 추정, 및 분석 값 생성을 위한 시스템 및 방법
EP2559336B1 (fr) Système et procédé de classification automatique d'alevins de poisson au moyen de procédés optiques
Garcia et al. Lameness detection challenges in automated milking systems addressed with partial least squares discriminant analysis
CN112131921A (zh) 基于立体视觉的生物自动量测系统及其量测方法
DK177925B1 (da) Fremgangsmåde og apparat til kvalitetsbestemmelse af mælk produceret ved maskinel malkning
CN115512215A (zh) 水下生物监测方法、装置及存储介质
Barbedo Computer-aided disease diagnosis in aquaculture: current state and perspectives for the future.
WO2022008576A1 (fr) Détection de mammite clinique chez des mammifères laitiers
US20210068404A1 (en) Method for evaluating a health state of an anatomical element, related evaluation device and related evaluation system
WO2023201443A1 (fr) Procédé et système d'évaluation automatisée d'animaux
CN116543462A (zh) 基于视频骨骼的奶牛行为识别判断奶牛健康状况的方法
Shah et al. Towards Intelligent Detection and Classification of Rice Plant Diseases Based on Leaf Image Dataset.
CN114600793A (zh) 一种奶牛乳房炎自动检测方法、系统、存储介质及设备
Yeşil et al. Recognition of Hereford and Simmental cattle breeds via computer vision.
US20240074400A1 (en) Quality sensor, computer-implemented method of predicting inhomogeneities in milk extracted from an animal, computer program and non-volatile data carrier
US20100316294A1 (en) Method and Data Processing System for Creating Image Segmentation Models
Afridi et al. Leveraging Transfer Learning for Analyzing Cattle Front Teat Placement
CN117409403B (zh) 一种基于深度学习的水稻稻穗成熟度估算方法
Gunaratnam et al. Computer vision in livestock management and production
Van Hertem et al. On farm implementation of a fully automatic computer vision system for monitoring gait related measures in dairy cows
NO347741B1 (en) System and method for sorting animals

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21742797

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21742797

Country of ref document: EP

Kind code of ref document: A1