WO2019200735A1 - 牲畜特征向量的提取方法、装置、计算机设备和存储介质 - Google Patents

牲畜特征向量的提取方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019200735A1
WO2019200735A1 PCT/CN2018/095317 CN2018095317W WO2019200735A1 WO 2019200735 A1 WO2019200735 A1 WO 2019200735A1 CN 2018095317 W CN2018095317 W CN 2018095317W WO 2019200735 A1 WO2019200735 A1 WO 2019200735A1
Authority
WO
WIPO (PCT)
Prior art keywords
cnn model
feature vector
livestock
identified
picture
Prior art date
Application number
PCT/CN2018/095317
Other languages
English (en)
French (fr)
Inventor
刘奡智
王健宗
肖京
Original Assignee
平安科技（深圳）有限公司
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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019200735A1 publication Critical patent/WO2019200735A1/zh

Links

Classifications

    • 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
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08—Insurance
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/02—Agriculture; Fishing; Forestry; Mining

Definitions

  • the present application relates to the field of livestock identification, and in particular to a method, device, computer device and storage medium for extracting a feature vector of a livestock.
  • the main object of the present application is to provide a method, device, computer device and storage medium for extracting livestock feature vectors for quickly identifying livestock.
  • the present application provides a method for extracting a livestock feature vector, including:
  • the livestock picture to be identified is input into the first CNN model for processing, and the feature vector corresponding to the animal image to be identified is obtained.
  • the application also provides an apparatus for extracting a livestock feature vector, comprising:
  • An acquiring unit configured to acquire a first CNN model that is completed based on a specified picture sample and a feature vector corresponding to the picture sample;
  • a freezing unit configured to freeze a weight parameter of a specified layer in the first CNN model
  • a processing unit configured to input a picture of the animal to be identified into the CNN model for processing, and the feature vector corresponding to the picture of the animal to be identified.
  • the application further provides a computer device comprising a memory and a processor, the memory storing computer readable instructions, the processor executing the computer readable instructions to implement the steps of any of the methods described above.
  • the present application also provides a computer non-transitory readable storage medium having stored thereon computer readable instructions that, when executed by a processor, implement the steps of any of the methods described above.
  • the method, device, computer equipment and storage medium for extracting livestock feature vectors of the present application are based on the concept of migration learning, using the first CNN model that has been trained, combined with the fine-tune method (so-called fine-tune is already trained) Models, plus our own data, to train new models, equivalent to using the first few layers of the already trained model to extract shallow features, and finally fall into the classification of new data sets), identify livestock
  • the last layers of the first CNN model need to be processed, and the weight parameters of other layers can be frozen, that is, in the process of recognizing the livestock picture, the frozen layer of the first CNN model does not need to participate in the process of processing, and further Greatly improved the speed of livestock picture recognition.
  • FIG. 1 is a schematic flow chart of a method for extracting a feature vector of a livestock according to an embodiment of the present application
  • FIG. 2 is a schematic flow chart of a method for extracting a feature vector of a livestock according to an embodiment of the present application
  • FIG. 3 is a schematic flow chart of a method for extracting a feature vector of a livestock according to an embodiment of the present application
  • FIG. 4 is a schematic block diagram showing the structure of an apparatus for extracting livestock feature vectors according to an embodiment of the present application
  • FIG. 5 is a schematic block diagram showing the structure of an apparatus for extracting livestock feature vectors according to an embodiment of the present application.
  • FIG. 6 is a schematic block diagram showing the structure of an apparatus for extracting livestock feature vectors according to an embodiment of the present application
  • FIG. 7 is a schematic block diagram showing the structure of an apparatus for extracting livestock feature vectors according to an embodiment of the present application.
  • FIG. 8 is a schematic block diagram showing the structure of a computer device according to an embodiment of the present application.
  • an embodiment of the present application provides a method for extracting a feature vector of a livestock, including the following steps:
  • the basic structure of the CNN (Convolutional Neural Network) model includes two layers, one of which is a feature extraction layer, and the input of each neuron is connected to the local accepting domain of the previous layer, and Extract the local features.
  • the second is the feature mapping layer
  • each computing layer of the network is composed of multiple feature maps
  • each feature map is a plane. The weights of all neurons on the plane are equal.
  • the feature mapping structure uses a small sigmoid function that affects the function kernel (Sigmoid function is a function of S type that is common in biology, also known as S-type growth curve. In information science, due to its single increase and inverse function single increase, etc.
  • the Sigmoid function is often used as a threshold function of the neural network, mapping variables to between 0, 1) as an activation function of the convolutional network, so that the feature map has displacement invariance.
  • the neurons on one mapping surface share weights, the number of network free parameters is reduced.
  • Each convolutional layer in the convolutional neural network is followed by a computational layer for local averaging and quadratic extraction. This unique two-feature extraction structure reduces feature resolution.
  • the first CNN model described above is a well-trained model, that is, there is no need to re-train through the image samples, but to use the CNN model trained in other fields or scenes to be used directly.
  • the first CNN model includes one of a VGG19 model, a VGG16 model, a VGG-F model, a ResNet50 model, a ResNet152 model, a DPN131 model, an InceptionV3 model, an Xception model, a DenseNet model, and an AlexNet model.
  • the first CNN model includes multiple layers (for example, a volume base layer and a full connection layer), and according to a preset rule, at least one of the layers may be used as a designated layer (excluding the last layer), and frozen. Specifies the weight parameter of the layer.
  • the volume base layer or/and of the pre-sequence N may be sorted according to the order among the layers.
  • the fully connected layer acts as a designated layer, where N is a positive integer greater than 1 less than the total number of layers of the first CNN model, and N is a preset value. That is to say, the weight parameter of the first specified number of layers of the first CNN model is frozen, that is, the frozen weight parameter is not changed when the livestock picture to be identified is processed later.
  • the first CNN model described above is modeled after ImageNet is trained using VGG19.
  • the above ImageNet is a computer vision system identification project name.
  • the VGG19 model includes a total of 19 layers of neural network layers, wherein the first 16 layers are convolution layers, and the last three layers are fully connected layers.
  • the weight parameters of the first 16 layers of the convolutional layer and the first 2 layers of the fully connected layer are frozen, that is, the weight parameters of the first eighteen layers of the VGG19 model are frozen.
  • the picture of the animal to be identified refers to a picture taken by a farmer or the like according to a certain requirement. For example, if the animal is a pig, a picture with more characteristics of the pig, such as a front or side view of the pig, is usually taken. Instead of taking a partial photo of the pig, such as taking only a photo of the belly.
  • the picture of the animal to be identified is input into the first CNN model for processing. Since the weight value of the specified layer is frozen, it can only be processed in the unfrozen network layer, and the processing speed is faster.
  • the first CNN model is a VGG19 model that is trained in the ImageNet game, which freezes the first 18 layers of the network layer.
  • the processing only the livestock image to be identified needs to be input to the VGG19 model.
  • the last layer can be processed, and the result of processing the completed output is the feature vector of the identified livestock.
  • the processing is ended, and the processing result is output.
  • the loss function of the neural network will become smaller and smaller until it converges to near zero.
  • the output of the feature vector of the animal image to be recognized can be considered.
  • the output eigenvector is optimal.
  • the method before the step S3 of inputting the image of the animal to be identified into the first CNN model for obtaining the feature vector corresponding to the image of the animal to be identified, the method includes:
  • the number of nodes of the last layer of the first CNN model corresponds to the number of the maximum number of pictures that can be processed, because the number of nodes of the last layer of different first CNN models is different, which can be performed.
  • the number of pictures processed is also different.
  • the first CNN model is a VGG19 model trained in an ImageNet game, and the number of nodes in the last layer is 4096, so that it can process up to 4096 pictures, if more than 4096 pictures are exceeded, If the VGG19 model is not applicable, you need to select another model.
  • the number of nodes of the last layer of the first CNN model is first determined, and the current operator is notified by display, voice, etc., to prevent the current operator from inputting too many pictures. It is impossible to complete the corresponding training.
  • the specific judging process is: reading the current network structure of the first CNN model, and then searching for a model similar to the network structure in the preset model database, thereby determining a specific model, and then presetting according to the model. Find the number of nodes in the list that correspond to the last layer.
  • a plurality of different first CNN models are pre-stored in the user's computer device, such as a trained VGG19 model, a VGG16 model, a VGG-F model, a ResNet50 model, a ResNet152 model, a DPN131 model, and an InceptionV3.
  • Model, Xception model, DenseNet model and AlexNet model, etc. and add a label to each model, the label corresponds to the value of the node number of the last layer of the model.
  • look for the corresponding one according to its label look for the corresponding one according to its label. The number of nodes is used to know how many pictures of livestock to be identified can be entered at present.
  • the method includes:
  • the feature vector corresponding to the animal image to be identified is subjected to softmax normalization processing, and the probability distribution output is obtained, which is convenient for calculation.
  • the method before the step S3 of inputting the image of the animal to be identified into the first CNN model for obtaining the feature vector corresponding to the animal image to be identified, the method includes:
  • the first CNN model is a VGG19 model that is trained for training in the ImageNet game, and the weight parameters of the first 18 layers are frozen, so when the picture is recognized, the weight parameter of the last layer is reset to zero. To increase processing speed.
  • the method before the step S1 of acquiring the first CNN model based on the specified picture sample and the feature vector corresponding to the picture sample, the method includes:
  • the list of farmers in the designated area refers to the list of farmers that can be obtained within a specified range.
  • the information of the farmers who have insured the livestock insurance will be recorded by the insurance company in the form of electronic documents.
  • the database of the animal insurance policy can be registered, and then the information of the electronic documents in the database can be read.
  • the contact information of the farmers There are two types of contact information for the farmers here, one is the mobile phone number, and the other is the application with instant messaging function (APP), such as an electronic insurance application or an instant messaging application such as WeChat and QQ. If the contact information of the farmer is the mobile phone number, the collection task will be sent to the farmer via SMS.
  • APP instant messaging function
  • the collection task is sent to the farmer through the APP.
  • the above method for extracting livestock feature vectors is applied to the field of livestock insurance, so a corresponding application program is developed, which integrates a livestock insurance, a risk detection system, and an instant messaging system, and is insured by the application.
  • the user's information is automatically saved to the default database.
  • the single animal species to be identified above refers to a single species of livestock, such as a single species of pigs, cattle, sheep, and the like.
  • the above-mentioned facial picture refers to a photo that is easy to identify and has a single species feature, and is generally a positive face photo of a livestock, such as a positive face photo of a pig.
  • the crowdsourcing task is also issued through a crowdsourcing platform, which is to collect facial images of a single animal species to be identified, and the crowdsourced user is the list of the designated regional farmers obtained above. The corresponding farmer.
  • the picture sent by the farmer is saved in the designated storage space, such as the designated database.
  • the above-mentioned formation of the picture samples means that the pictures will be used as samples of deep neural network learning, and the picture samples are first vectorized to correspond to the pictures, so as to be input to the second CNN model. Training.
  • the picture is simply screened, and the screening rule is whether the quality of the picture reaches the preset requirement, specifically:
  • the fuzzy determination is determined by calculating the difference between the current pixel point and the mean value of the pixel points in the picture.
  • f(x,y) we use f(x,y) to represent the image, where .
  • the mean value of the absolute difference in the level of the entire picture is:
  • Sun blur and Blur cnt are the number of Inverse Blueriness and fuzzy points, respectively, and Edge cnt is the total number of edges.
  • the candidate noise estimates are as follows:
  • the corresponding value can be calculated in the vertical direction.
  • the candidate noise is:
  • N_cand(x, y) represents the candidate noise, which is 0 in the edge region.
  • the noise mean and ratio are:
  • Sum_Noise is the sum of N(x, y) and Noise_cnt is the total number of noises.
  • C a combination of noise and blur.
  • weights are obtained by linear regression analysis. In the embodiment, these weights are: .
  • the metric is greater than the preset value, the above picture is considered to be a picture whose quality is not up to standard, and is deleted.
  • the sample picture is divided into a training set and a test set, and the picture of the test set is input into the second CNN model for training.
  • the second CNN model refers to the initial state of the model, that is, there is no A model that has been entered with any data for calculation.
  • the model of the training is tested by the picture of the test set and the corresponding feature vector of the picture. If the test is qualified, the training is completed.
  • the model is a model that can be used, namely the first CNN model described above.
  • the first CNN model established by the farmer to provide a picture of the animal species to be identified is more accurate.
  • the pictures provided by the farmers are all pictures of the pigs, so the first CNN model trained is relatively more accurate in the picture recognition of the pigs.
  • step S13 the step of dividing the picture sample into the training set and the test set based on the training to the specified second CNN model to obtain the training of the first CNN model before step S13 includes:
  • the picture acquisition task is stopped from being sent without continuing to acquire the picture samples.
  • the image acquisition task is generally sent in the form of a crowdsourcing task, that is, payment is required, and when the acquired image reaches a specified number, the publishing task can be stopped, money is saved, and the like.
  • the money of the crowd-occupied user can be not paid, and the discount amount and other preferential policies can be provided according to the number of pictures provided by the farmer to achieve mutual benefit.
  • the facial image collection task of the single animal species to be identified includes a task of collecting a picture of a sick pig, establishing a collection of pig face images of the dead pig; and then training the dead face image of the dead pig to target death.
  • the first CNN model of the pig is used to determine whether the identified pig is a sick pig. Because the farmers are in the process of insurance, in order to save the premium, some pigs may be insured, while the other pigs are not insured. When insured, the farmers are insured online. Specifically, the farmers log in to the insurance system, register (fill in the name, trade name, contact information of the farmer), and then mark the insured pigs by uploading pictures of the insured pigs.
  • the pictures are uploaded to the insurance system for risk detection, and the first CNN model is used to determine whether the pigs in the pictures after the death are insured.
  • the first CNN model is used to determine whether the pigs in the pictures after the death are insured.
  • the pigs that have not been insured by the farmers are dying, and the insured and alive pigs are photographed to deceive the system, so that the pigs that the system thinks are dead are only the insured pigs, thereby The insurance company compensates for insurance costs.
  • a first CNN model for identifying a sick pig is established. After the farmer uploads a picture of the insured pig, it is first determined whether the pig in the picture is a dead pig, and if so, the judgment is made. Whether the pig is insured with pigs reduces the risk of the insurance company.
  • a system for insurance for pigs is set up, and the specific operation process includes:
  • the task includes two kinds, one is to collect photos of any pig face, and the other is to collect photos of sick pigs. After the image is collected, the quality of the image is detected, and the unclear deletion is performed. At the same time, the information on the quality of the uploaded image is recorded, so as to reduce the degree of preferential treatment in the later period of the farmer's insurance.
  • the first type of pig face photo When the first type of pig face photo reaches the specified number, it is input into the VGG19 neural network model for processing to obtain the initial pig face recognition model; when the photo of the dead pig reaches the specified number, the initial dead pig is obtained by the same method. Identify the model;
  • the insured farmer insures, accept the photo of the sick pig uploaded by the farmer, and then judge whether the dead pig is the insured pig through the above-mentioned pig face recognition model, and if so, enter the photo of the dead pig into the above-mentioned sick death.
  • the pig identification model it is judged whether the dead pig photo is dead or not. If yes, the follow-up insurance process is executed. Otherwise, the current risk reporting process is stopped, and the farmer may report fraudulent behavior for the investigator to investigate.
  • the method for extracting livestock feature vectors in the embodiment of the present application is based on the concept of migration learning, using the first CNN model that has been trained, combined with the fine-tune method, in the process of identifying livestock, only the first CNN model is needed. The last few layers are processed, and the weight parameters of other layers can be frozen. That is, in the process of recognizing the livestock picture, the frozen layer of the first CNN model does not need to participate in the processing process, thereby greatly improving the speed of livestock picture recognition.
  • the present application further provides an apparatus for extracting a livestock feature vector, including:
  • the obtaining unit 10 is configured to acquire a first CNN model that is completed based on the specified picture sample and the feature vector corresponding to the picture sample.
  • a freezing unit 20 configured to freeze a weight parameter of a specified layer in the first CNN model
  • a processing unit 30 configured to input a picture of the animal to be identified into the first CNN model for processing, and obtain a feature vector corresponding to the picture of the animal to be identified
  • the basic structure of the CNN (Convolutional Neural Network) model includes two layers, one of which is a feature extraction layer, and the input of each neuron is connected to the local accepting domain of the previous layer, and Extract the local features. Once the local feature is extracted, its positional relationship with other features is also determined; the second is the feature mapping layer, each computing layer of the network is composed of multiple feature maps, and each feature map is a plane. The weights of all neurons on the plane are equal.
  • the feature mapping structure uses a small sigmoid function that affects the function kernel (Sigmoid function is a function of S type that is common in biology, also known as S-type growth curve. In information science, due to its single increase and inverse function single increase, etc.
  • the Sigmoid function is often used as a threshold function of the neural network, mapping variables to between 0, 1) as an activation function of the convolutional network, so that the feature map has displacement invariance.
  • the neurons on one mapping surface share weights, the number of network free parameters is reduced.
  • Each convolutional layer in the convolutional neural network is followed by a computational layer for local averaging and quadratic extraction. This unique two-feature extraction structure reduces feature resolution.
  • the first CNN model described above is a well-trained model, that is, there is no need to re-train through the image samples, but to use the CNN model trained in other fields or scenes to be used directly.
  • the first CNN model includes one of a VGG19 model, a VGG16 model, a VGG-F model, a ResNet50 model, a ResNet152 model, a DPN131 model, an InceptionV3 model, an Xception model, a DenseNet model, and an AlexNet model.
  • the first CNN model includes multiple layers (for example, a volume base layer and a full connection layer), and according to a preset rule, at least one of the layers may be designated as a designated layer (excluding the last layer), and frozen. Specifies the weight parameter of the layer.
  • the volume base layer or/and of the pre-sequence N may be sorted according to the order among the layers.
  • the fully connected layer acts as a designated layer, where N is a positive integer greater than 1 less than the total number of layers of the first CNN model, and N is a preset value. That is to say, the weight parameter of the first specified number of layers of the first CNN model is frozen, that is, the frozen weight parameter is not changed when the livestock picture is processed later.
  • the first CNN model described above is modeled after ImageNet is trained using VGG19.
  • the above ImageNet is a computer vision system identification project name.
  • the VGG19 model includes a total of 19 layers of neural network layers, wherein the first 16 layers are convolution layers, and the last three layers are fully connected layers.
  • the weight parameters of the first 16 layers of the convolutional layer and the first 2 layers of the fully connected layer are frozen, that is, the weight parameters of the first eighteen layers of the VGG19 model are frozen.
  • the picture of the animal to be identified refers to a picture taken by a farmer or the like according to a certain requirement. For example, if the animal is a pig, a picture with more features of the pig, such as a front or side view of the pig, is usually taken. Instead of taking a partial photo of the pig, such as taking only a photo of the belly.
  • the picture of the animal to be identified is input into the first CNN model for processing. Since the weight value of the specified layer is frozen, it can only be processed in the unfrozen network layer, and the processing speed is faster.
  • the first CNN model is a VGG19 model that is trained in the ImageNet game, which freezes the first 18 layers of the network layer.
  • the process only the livestock image to be identified needs to be input to the VGG19 model.
  • the last layer can be processed, and the result of the training completion output is the feature vector of the identified livestock.
  • the recognition process is ended, and the recognition result is output.
  • the output characteristic vector of the animal image to be recognized can be considered as When the required feature vector is reached, in theory, when the convergence value of the loss function of the neural network is 0, the output feature vector is optimal.
  • the apparatus for extracting the livestock feature vector further includes:
  • the obtaining node unit 301 is configured to acquire the number of nodes of the last layer of the first CNN model, and generate, according to the number of nodes, the reminding information that inputs the maximum number of livestock pictures to be identified.
  • the number of nodes of the last layer of the first CNN model corresponds to the number of sheets of the largest picture that can be processed, because the number of nodes of the last layer of different first CNN models is different, The number of pictures processed is also different.
  • the first CNN model is a VGG19 model trained in an ImageNet game, and the number of nodes in the last layer is 4096, so that it can process up to 4096 pictures, if more than 4096 pictures are exceeded, If the VGG19 model is not applicable, you need to select another model.
  • the number of nodes of the last layer of the first CNN model is first determined, and the current operator is notified by display, voice, etc., to prevent the current operator from inputting too many pictures. It is impossible to complete the corresponding training.
  • the specific judging process is: reading the current network structure of the first CNN model, and then searching for a model similar to the network structure in the preset model database, thereby determining a specific model, and then presetting according to the model. Find the number of nodes in the list that correspond to the last layer.
  • a plurality of different first CNN models are pre-stored in the user's computer device, such as a trained VGG19 model, a VGG16 model, a VGG-F model, a ResNet50 model, a ResNet152 model, a DPN131 model, and an InceptionV3.
  • Model, Xception model, DenseNet model and AlexNet model, etc. and add a label to each model, the label corresponds to the value of the node number of the last layer of the model.
  • look for the corresponding one according to its label look for the corresponding one according to its label. The number of nodes is used to know how many pictures of livestock to be identified can be entered at present.
  • the apparatus for extracting the livestock feature vector further includes:
  • the normalization unit 40 performs softmax normalization processing on the feature vector corresponding to the animal image to be recognized.
  • the feature vector corresponding to the animal image to be identified is subjected to softmax normalization processing, and the probability distribution output is obtained, which is convenient for calculation.
  • the apparatus for extracting the livestock feature vector further includes:
  • the resetting unit 31 is configured to reset the weight parameter of each layer in the first CNN model that is not frozen.
  • the reset unit 301 resetting the weight parameter of each layer of the un-freezing neural network in the first CNN model to 0 increases the speed of training, and does not have much influence on the accuracy of the result, that is, The impact of the results is within acceptable limits.
  • the first CNN model is a VGG19 model that is trained for training in the ImageNet game, and the weight parameters of the first 18 layers are frozen, so when the picture is recognized, the weight parameter of the last layer is reset to zero. To increase processing speed.
  • the apparatus for extracting the livestock feature vector further includes:
  • the classification training unit 13 is configured to divide the picture sample into a training set and a test set, and perform training based on the specified second CNN model to obtain the trained first CNN model.
  • the list of farmers in the above designated area refers to a list of farmers that can be obtained within a specified range.
  • the information of the farmers who have insured the livestock insurance will be recorded by the insurance company in the form of electronic documents.
  • the database of the animal insurance policy can be registered, and then the information of the electronic documents in the database can be read.
  • the contact information of the farmers There are two types of contact information for the farmers here, one is the mobile phone number, and the other is the application with instant messaging function (APP), such as an electronic insurance application or an instant messaging application such as WeChat and QQ. If the contact information of the farmer is the mobile phone number, the collection task will be sent to the farmer via SMS.
  • APP instant messaging function
  • the collection task is sent to the farmer through the APP.
  • the above-mentioned livestock feature vector extracting device is applied to the field of livestock insurance, so a corresponding application program is developed, which integrates a livestock insurance, a risk reporting system, and an instant messaging system, and is insured by the application.
  • the user's information is automatically saved to the default database.
  • the single animal species to be identified above refers to a single species of livestock, such as a single species of pigs, cattle, sheep, and the like.
  • the above-mentioned facial picture refers to a photo that is easy to identify and has a single species feature, and is generally a positive face photo of a livestock, such as a positive face photo of a pig.
  • the crowdsourcing task is also issued through a crowdsourcing platform, which is to collect facial images of a single animal species to be identified, and the crowdsourced user is the list of the designated regional farmers obtained above. The corresponding farmer.
  • the pictures sent by the farmer are saved in a designated storage space, such as a designated database.
  • the above-mentioned image sample formation means that the images will be used as samples processed by the deep neural network, and the image samples are first vectorized to correspond to the images, so as to be input to the second CNN model. Training.
  • the picture is simply filtered, and the screening rule is whether the quality of the picture reaches the preset requirement. The specific process is described in the above method embodiment, and will not be described again. .
  • the above sample picture is divided into a training set and a test set, and the picture of the test set is input into the second CNN model for training, and the second CNN model refers to the initial state of the model, that is, A model that has not entered any data for calculation.
  • the model of the training is tested by the picture of the test set and the corresponding feature vector of the picture. If the test is qualified, the training is completed.
  • the model is a model that can be used, namely the first CNN model described above.
  • the first CNN model established by the farmer to provide a picture of the animal species to be identified is more accurate.
  • the pictures provided by the farmers are all pictures of the pigs, so the first CNN model trained is relatively more accurate in the picture recognition of the pigs.
  • the apparatus for extracting the livestock feature vector further includes:
  • the determining threshold unit 131 is configured to determine whether the number of the picture samples reaches a preset value
  • the publishing unit 132 is configured to stop sending the image acquisition task if the preset value is reached.
  • the picture acquisition task is stopped from being sent without continuing to acquire the picture samples.
  • the image acquisition task is generally sent in the form of a crowdsourcing task, that is, payment is required, and when the acquired image reaches a specified number, the publishing task can be stopped, money is saved, and the like.
  • the money of the crowd-occupied user can be not paid, and the discount amount and other preferential policies can be provided according to the number of pictures provided by the farmer to achieve mutual benefit.
  • the facial image collection task of the single animal species to be identified includes a task of collecting a picture of a sick pig, establishing a collection of pig face images of the dead pig; and then training the dead face image of the dead pig to target death.
  • the first CNN model of the pig is used to determine whether the identified pig is a sick pig. Because the farmers are in the process of insurance, in order to save the premium, some pigs may be insured, while the other pigs are not insured. When insured, the farmers are insured online. Specifically, the farmers log in to the insurance system, register (fill in the name, trade name, contact information, etc.) of the farmer, and then mark the insured pig by uploading pictures of the insured pigs.
  • the pictures are uploaded to the insurance system for risk detection, and the first CNN model is used to determine whether the pigs in the pictures after the death are insured.
  • the first CNN model is used to determine whether the pigs in the pictures after the death are insured.
  • the pigs that have not been insured by the farmers are dying, and the insured and alive pigs are photographed to deceive the system, so that the pigs that the system thinks are dead are only the insured pigs, thereby The insurance company compensates for insurance costs.
  • a first CNN model for identifying a sick pig is established. After the farmer uploads a picture of the insured pig, it is first determined whether the pig in the picture is a dead pig, and if so, the judgment is made. Whether the pig is insured with pigs reduces the risk of the insurance company.
  • the apparatus for extracting livestock feature vectors in the embodiment of the present application utilizes the first CNN model that has been trained, combined with the fine-tune method, in the process of identifying livestock, only the first CNN model is needed. The last few layers are processed, and the weight parameters of other layers can be frozen. That is, in the process of recognizing the livestock picture, the frozen layer of the first CNN model does not need to participate in the processing process, thereby greatly improving the speed of livestock picture recognition.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 8.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the computer designed processor is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions, and a database.
  • the memory provides an environment for the operation of operating systems and computer readable instructions in a non-volatile storage medium.
  • the database of the computer device is used to store data such as the first CNN model that identifies the animal.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection.
  • the computer readable instructions are executed by a processor to implement the processes of the various method embodiments described above.
  • An embodiment of the present application further provides a computer non-volatile readable storage medium having stored thereon computer readable instructions, which are implemented by a processor to implement the processes of the foregoing method embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • General Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Agronomy & Crop Science (AREA)
  • Animal Husbandry (AREA)
  • Marine Sciences & Fisheries (AREA)
  • Mining & Mineral Resources (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Mathematical Physics (AREA)
  • Primary Health Care (AREA)
  • Computational Linguistics (AREA)
  • Tourism & Hospitality (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Development Economics (AREA)
  • Technology Law (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种牲畜特征向量的提取方法、装置、计算机设备和存储介质，其中牲畜特征向量的提取方法包括：获取基于指定图片样本、以及图片样本对应的特征向量训练完成的第一CNN模型(S1)；冻结第一CNN模型的指定层的权重参数(S2)；将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出待识别的牲畜图片对应的特征向量(S3)。

Description

牲畜特征向量的提取方法、装置、计算机设备和存储介质
本申请要求于2018年4月17日提交中国专利局、申请号为2018103446686，申请名称为“牲畜特征向量的提取方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及到牲畜识别领域，特别是涉及到一种牲畜特征向量的提取方法、装置、计算机设备和存储介质。
 背景技术
在牲畜保险业务中，养殖户往往利用没有成熟的识别牲畜的技术，而使用未投保的牲畜作为已投保的牲畜进行骗保，保险公司的保险风险较大。而区别报险牲畜是否为投保牲畜，可以通过图像进行区分，因此，在投保时拍摄一组照片，报险时拍摄一组照片，然后对两组照片进行比对，以判定当前的报险牲畜是否为投保的牲畜。由于养殖户在投保过程中要对大批量的牲畜进行拍照识别，以得到对应投保牲畜的特征向量，方便在报险时对病死牲畜的识别，判断病死牲畜是否为投保牲畜。因为需要大批量地识别牲畜，所以需要一种可以提高识别牲畜速度的方法。
 技术问题
本申请的主要目的为提供一种快速识别牲畜的牲畜特征向量的提取方法、装置、计算机设备和存储介质。
 技术解决方案
为了实现本申请提出的目的，本申请提出一种牲畜特征向量的提取方法，包括：
获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
冻结所述第一CNN模型中指定层的权重参数；
将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
本申请还提供一种牲畜特征向量的提取装置，包括：
获取单元，用于获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
冻结单元，用于冻结所述第一CNN模型中指定层的权重参数；
处理单元，用于将待识别的牲畜图片输入到所述CNN模型中进行处理，所述待识别的牲畜图片对应的特征向量。
本申请还提供一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现上述任一项所述方法的步骤。
本申请还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，所述计算机可读指令被处理器执行时实现上述任一项所述的方法的步骤。
 有益效果
本申请的牲畜特征向量的提取方法、装置、计算机设备和存储介质，基于迁移学习的理念，利用已经训练过的第一CNN模型，结合fine-tune方法（所谓fine-tune就是用已经训练好的模型，加上我们自己的数据，来训练新的模型，相当于使用已经训练好的的模型的前几层，来提取浅层特征，最后再落入新数据集的分类中），在识别牲畜过程中，只需要对第一CNN模型的最后几层进行处理，其它层的权重参数可以冻结，即，识别牲畜图片的过程中，第一CNN模型被冻结的层不需要参与处理的过程，进而大大地提高了牲畜图片识别的速度。
 附图说明
图1 为本申请一实施例的牲畜特征向量的提取方法的流程示意图；
图2 为本申请一实施例的牲畜特征向量的提取方法的流程示意图；
图3 为本申请一实施例的牲畜特征向量的提取方法的流程示意图；
图4 为本申请一实施例的牲畜特征向量的提取装置的结构示意框图；
图5 为本申请一实施例的牲畜特征向量的提取装置的结构示意框图；
图6 为本申请一实施例的牲畜特征向量的提取装置的结构示意框图；
图7 为本申请一实施例的牲畜特征向量的提取装置的结构示意框图；
图8 为本申请一实施例的计算机设备的结构示意框图。
 本发明的最佳实施方式
参照图1，本申请实施例提供一种牲畜特征向量的提取方法，包括步骤：
S1、获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
S2、冻结所述第一CNN模型中指定层的权重参数；
S3、将待识别的牲畜图片输入到所述CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
如上述步骤S1所述，CNN（Convolutional Neural Network，卷积神经网络）模型的基本结构包括两层，其一为特征提取层，每个神经元的输入与前一层的局部接受域相连，并提取该局部的特征。一旦该局部特征被提取后，它与其它特征间的位置关系也随之确定下来；其二是特征映射层，网络的每个计算层由多个特征映射组成，每个特征映射是一个平面，平面上所有神经元的权值相等。特征映射结构采用影响函数核小的sigmoid函数（Sigmoid函数是一个在生物学中常见的S型的函数，也称为S型生长曲线。在信息科学中，由于其单增以及反函数单增等性质，Sigmoid函数常被用作神经网络的阈值函数，将变量映射到0,1之间）作为卷积网络的激活函数，使得特征映射具有位移不变性。此外，由于一个映射面上的神经元共享权值，因而减少了网络自由参数的个数。卷积神经网络中的每一个卷积层都紧跟着一个用来求局部平均与二次提取的计算层，这种特有的两次特征提取结构减小了特征分辨率。上述第一CNN模型是一种训练好的模型，即，无需重新通过图片样本进行训练，而是拿来主义，将其它领域或场景下训练完成的CNN模型拿过来直接使用。本实施例中，上述第一CNN模型包括VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、 Xception模型、DenseNet模型和AlexNet模型中的一种。
如上述步骤S2所述，第一CNN模型中包括多层（例如，卷基层和全连接层），根据预设规则，可以将其中至少一个层作为指定层（不包括最后一层），并冻结指定层的权重参数。
在一种可能的实施方式中，由于第一CNN模型中的多个卷基层和多个全连接层存在先后顺序，可以按照各层之间的先后顺序，将排序前N的卷基层或/和全连接层作为指定层，其中N为大于1小于第一CNN模型总层数的正整数，且N为预设值。也就是说，将第一CNN模型的前指定层数的权重参数冻结，即，被冻结的权重参数，在之后对待识别的牲畜图片进行处理时，不会进行改变。在一具体实施例中，上述第一CNN模型是ImageNet的使用VGG19训练后得到模型。上述ImageNet是一个计算机视觉***识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别***建立的，其中包含120万张图片，分别包含1000种类别的图片，其中的动物，比如猫，狗，牛等牲畜的图片数量巨大。本实施例中，上述VGG19模型一共包括19层神经网络层，其中，其前16层为卷积层，最后三层为全连接层。在一具体实施例中，冻结前16层卷积层和前2层全连接层的权重参数，即冻结VGG19模型的前十八层的权重参数。
如上述步骤S3所述，上述待识别的牲畜图片是指养殖户等按照一定要求拍摄的图片，比如，牲畜为猪，通常会拍摄猪的正面照或侧面照等含有猪的特征较多的照片，而不会拍摄猪的局部照片，比如只拍摄肚子的照片等。将待识别的牲畜图片输入到第一CNN模型中进行处理，由于其指定层数的权重值被冻结，所以只能在未被冻结的网络层进行处理，处理速度会更快。在一具体实施例中，第一CNN模型为ImageNet比赛中训练完成的VGG19模型，其冻结了前18层网络层，所以，在处理过程中，只需要将待识别的牲畜图片输入到VGG19模型的最后一层进行处理即可，处理完成输出的结果即为被识别牲畜的特征向量。在处理过程中，直到其最后一层的神经网络的损失函数变为一个接近0的阈值时，结束处理的过程，输出处理结果。可以这么理解，随着VGG19的最后一个全连接层参数的不断变化，神经网络的损失函数会越来越小，直到收敛到接近0，此时，可以认为输出的待识别的牲畜图片的特征向量为到达要求的特征向量，理论上，神经网络的损失函数的收敛值为0时，输出的特征向量最优。
本实施例中，上述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤S3之前，包括：
S301、获取所述第一CNN模型最后一层的节点数量，根据所述节点数量生成最大输入所述待识别的牲畜图片数量的提醒信息。
如上述步骤S301所述，上述第一CNN模型的最后一层的节点数数量对应着其最大可以处理图片的张数，因为不同的第一CNN模型的最后一层的节点数量不同，其可以进行处理的图片张数也不同。在一具体实施例中，上述第一CNN模型为ImageNet比赛中训练完成的VGG19模型，其最后一层的节点数量为4096，那么其最多能对4096张图片进行处理，如果超出4096张图片，则不适用VGG19模型，需要选择其它的模型。本实施例中，在输入牲畜图片之前会先判断上述第一CNN模型的最后一层的节点数数量，并通过显示、语音等形式告知当前的操作者，以防止当前操作者输入过多的图片而无法完成相应的训练。具体的判断过程为：读取当前的第一CNN模型的网络结构，然后在预设的模型数据库中查找与该网络结构相同的模型，进而判断出具体的模型，然后根据该模型，在预设列表中查找与其对应的最后一层的节点数量。在另一实施例中，用户的计算机设备中预存有多个不同的第一CNN模型，如分别为训练后的VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、 Xception模型、DenseNet模型和AlexNet模型等，并给每一个模型添加一个标签，该标签对应模型的最后一层的节点数量值，当使用某一具体模型时，根据其标签查找与之对应的节点数量值，从而知道当前最多可输入多少张待识别的牲畜图片。
参照图2，本实施例中，上述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤S3之后，包括：
S4、对所述待识别的牲畜图片对应的特征向量进行softmax归一化处理。
如上述步骤S4所述，对待识别的牲畜图片对应的特征向量进行softmax归一化处理，会得出概率分布输出，方便计算。
参照图3，本实施例中，上述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤S3之前，包括：
S31、将所述第一CNN模型中未冻结的各层的权重参数重置为0。
如上述步骤S31所述，将第一CNN模型中最后一层的权重参数重置为0，会提高训练的速度，而对结果的准确性并没有很大的应影响，即对结果的影响在可以接受的范围之内。在一具体实施例中，第一CNN模型是为ImageNet比赛中训练完成的VGG19模型，其前18层的权重参数被冻结，那么在识别图片时，将其最后一层的权重参数重置为零，以提高处理速度。
本实施例中，上述获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型的步骤S1之前，包括：
S11、获取指定区域养殖户的名单，并根据名单分配待识别单一牲畜种类的脸部图片采集任务给各养殖户；
S12、收集养殖户反馈的图片，并存储到指定的存储空间内形成所述图片样本；
S13、将所述图片样本分成训练集和测试集，基于指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型。
如上述步骤S11所述，上述指定区域的养殖户名单是指可以在一指定范围内获取到的养殖户名单。比如，投保了牲畜险的养殖户的信息会被保险公司以电子文档的形式进行备案，当获取养殖户的名单时，可以登录存储牲畜保单的数据库，然后读取数据库中电子文档的养殖户信息，以得到养殖户的联系方式。这里养殖户的联系方式包括两种，一种是手机号码，另一种是带即时通讯功能的应用程序（APP）的账户，如电子投保的APP或者如微信、QQ等即时通讯APP。如养殖户的联系方式是手机号码，则通过短信的形式发送采集任务给养殖户。如果是电子投保的APP，则通过APP发送采集任务给养殖户。在本实施例中，上述牲畜特征向量的提取方法应用于牲畜保险领域，所以会开发一款对应的应用程序，其集成有牲畜投保、报险***，以及即时通讯***，通过该应用程序投保的用户的信息会自动保存到预设的数据库中，当发布任务时，直接通过上述应用程序进行发布即可。上述待识别单一牲畜种类是指单一种类的牲畜，比如猪、牛、羊等中的单一物种。上述的脸部图片是指方便识别的，具有单一物种特征较多的照片，一般为牲畜的正脸照片，如猪的正脸照片等。在另一具体实施例中，还以通过众包平台发布众包任务，众包任务即为收集待识别单一牲畜种类的脸部图片，而众包用户即为上述获取的指定区域养殖户的名单上对应的养殖户。
如上述步骤S12所述，将养殖户发送的图片保存到指定的存储空间内，如指定的数据库中等。上述形成所述图片样本，是指这些图片将会作为深度神经网络学习的样本进行使用，同样其会将图片样本先先进行向量化，使其与图片对应，以便于输入到上述第二CNN模型中训练。本实施例中，在接收到养殖户的图片之后，会对图片进行简单的筛选，其筛选规律为图片的质量是否达到预设要求，具体地：
A、对图片进行边缘检测，然后是模糊确定。此处模糊确定是通过计算当前像素点与图片内像素点均值之差来确定。我们用f(x,y) 表示图片，其中
Figure dest_path_image001
。
定义水平绝对差如下：
Figure dest_path_image002
整个图片的水平绝对差的均值为：
Figure dest_path_image003
如果当前像素点的
Figure dest_path_image004
则该像素点就是一个候选的边缘点
Figure dest_path_image005
. 如果
Figure dest_path_image006
比它水平方向两个相邻的点
Figure dest_path_image007
都大，则该像素点就被确认为一个边缘点。边缘点
Figure dest_path_image008
的判断总结如下：
Figure dest_path_image009
接下来我们检测边缘点是否模糊。定义：
Figure dest_path_image010
同理，按照以上的步骤我们可以计算垂直方向的值
Figure dest_path_image011
。
Figure dest_path_image012
两者之大者称作Inverse Blurriness（逆模糊强度），用于最终的模糊判定依据。
Figure dest_path_image013
低于阈值ThB的Inverse Blurriness 被认为是模糊的。实验测试表明此处的阈值ThB取值0.1。最后，边缘模糊的均值和比率为：
Figure dest_path_image014
其中Sun blur和Blur cnt分别是Inverse Blueriness和模糊点的数目，Edge cnt是边缘总点数。
B、对图片进行噪点检测。因为沿边缘的噪点视觉上不明显，因此我们只检测边缘之外的噪点。边缘检测会被噪点影响，因此在检测边缘之前做一个噪点滤波的预处理。在本实施例中，我们应用均值滤波来消除噪点。均值滤波之后的图像g(x,y)为：
Figure dest_path_image015
候选的噪点估计如下：
Figure dest_path_image016
同理可以在垂直方向计算对应的值。然后得到候选的噪点是：
Figure dest_path_image017
其中N_cand(x,y)表示候选噪点，它在边缘区域为0。
Figure dest_path_image018
噪点均值和比率为：
Figure dest_path_image019
其中Sum_Noise是N(x,y)之和，Noise_cnt是噪点总数目。C、噪点和模糊的组合。此处我们的图像质量评价指标定义如下：
Figure dest_path_image020
其中w1、w2、w3、w4是权值。通过线性回归分析获取这些权值。本实施例中这些权值为：
Figure dest_path_image021
。当Metric（度规）大于预设值，则认为上述的图片为质量不达标的图片等，将其删除。
如上述步骤S13所述，将上述的样本图片分成训练集和测试集，现将测试集的图片输入到上述第二CNN模型中进行训练，该第二CNN模型是指模型的初始状态，即没有输入过任何数据进行运算的模型。当第二CNN模型经过上述训练集的图片、以及图片对应特征向量训练完成后，通过测试集的图片、以及图片对应特征向量对该训练完成的模型进行检验，如果检验合格，则认定该训练完成的模型是可以使用的模型，即上述的第一CNN模型。本实施例中，通过养殖户提供待识别的牲畜种类的图片，建立的第一CNN模型会更加准确。比如，养殖户提供的图片全部是猪的图片，那么训练出来的第一CNN模型在对猪的图片识别上，相对会更加准确。
本实施例中，上述将所述图片样本分成训练集和测试集，基于到指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型的步骤S13之前，包括：
S131、判断所述图片样本的数量是否达到预设值；
S132、若达到，则停止发送图片获取任务。
如上述步骤S131和S132所述，图片样本的数量达到预设的值之后，无需继续获取图片样本，则停止发送图片获取任务。本实施例中，上述的图片获取任务一般是通过众包任务的形式进行发送的，即需要进行付费，当获取的图片达到指定数量之后，既可以停止发布任务，节约金钱等。本实施例中，可以不付众包用户（养殖户）金钱，可以根据养殖户提供图片的数量，对其投保时提供打折等优惠政策，以实现互利互惠。
本实施例中，上述的待识别单一牲畜种类的脸部图片采集任务包括专门采集病死猪的图片的任务，建立一个病死猪的猪脸图片集合；然后利用病死猪的猪脸图片集合训练针对病死猪的第一CNN模型，用于判断被识别的猪只是否是病死猪。因为，养殖户在投保过程，为了节约保费，可能只给部分猪只进行投保，而另一部分猪只不进行投保。在投保时，养殖户进行线上投保，具体地，养殖户登录投保***，进行注册（填写养殖户的名称、商号、联系方式），然后通过上传投保猪只的图片对投保猪只进行标记，当投保猪只发生病死状态时，拍摄其病死后的图片上传投保***进行报险，通过上述第一CNN模型判断病死后的图片中的猪只是否为投保猪只。此时存在一种情况，即养殖户未投保的猪只发生病死情况，其将投保过的且活着的猪只进行拍照以欺骗***，使***认为病死的猪只是投保过的猪只，进而使保险公司赔偿保险费用。本实施例中，建立一个专门识别病死猪的第一CNN模型，当养殖户上传投保猪只的图片之后，首先判断该图片中的猪只是否为病死状态的猪只，如果是，才进行判断该猪只是否为投保猪只，降低保险公司的风险。
在一具体实施例中，设置专门针对猪只进行保险的***，其具体操作的过程包括：
1、通过众包的方式发布猪脸图片收集任务，该任务包括两种，一种是收集任意猪脸照片，另一种是收集病死猪照片。在收集到图片之后对图片进行质量检测，将不清楚的进行删除，同时记录上传质量不过关图片的养殖的信息，以便于在后期该养殖户投保过程中降低对其优惠的程度。
2、当第一种猪脸照片达到指定数量之后，分别输入到VGG19神经网络模型中进行处理，得到初始的猪脸识别模型；当病死猪照片达到指定数量后，通过同样的方法得到初始的病死猪识别模型；
3、将上述的猪脸识别模型和病死猪识别模型配置到畜牧保险***中；
4、当猪只进行投保时，调用猪脸识别模型，冻结上述猪脸识模型的前18层的权重参数，以及重置第19层的权重参数为零；然后输入待识别的猪脸照片（最大输入4096张图片）到猪脸识别模型中进行猪脸识别，各猪脸照片在处理过程中直到网络收敛至指定阈值后备案。
5、当投保的养殖户进行保险时，接收养殖户上传的病死猪照片，然后通过上述猪脸识别模型判断该病死猪是否为投保猪只，若是，则将该病死猪照片输入到上述的病死猪识别模型中判断该病死猪照片是否病死猪，若是，则执行后续的报险流程，否者停止当前的报险流程，并上报该养殖户可能存在骗保行为，以供调查人员进行调查。在另一实施例中，还可以先通过病死猪识别模型判断上传的病死猪照片是否是病死猪，若是病死猪，则通过上述猪脸识别模型判断该病死猪是否为投保猪只；若不是病死猪，则上报该养殖户可能存在骗保行为，以供调查人员进行调查。
本申请的实施例中的牲畜特征向量的提取方法，基于迁移学习的理念，利用已经训练过的第一CNN模型，结合fine-tune方法，在识别牲畜过程中，只需要对第一CNN模型的最后几层进行处理，其它层的权重参数可以冻结，即，识别牲畜图片的过程中，第一CNN模型被冻结的层不需要参与处理的过程，进而大大地提高了牲畜图片识别的速度。
参照图4，本申请还提供一种牲畜特征向量的提取装置，包括：
获取单元10，用于获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
冻结单元20，用于冻结所述第一CNN模型中指定层的权重参数；
处理单元30，用于将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量
在上述获取单元10中，CNN（Convolutional Neural Network，卷积神经网络）模型的基本结构包括两层，其一为特征提取层，每个神经元的输入与前一层的局部接受域相连，并提取该局部的特征。一旦该局部特征被提取后，它与其它特征间的位置关系也随之确定下来；其二是特征映射层，网络的每个计算层由多个特征映射组成，每个特征映射是一个平面，平面上所有神经元的权值相等。特征映射结构采用影响函数核小的sigmoid函数（Sigmoid函数是一个在生物学中常见的S型的函数，也称为S型生长曲线。在信息科学中，由于其单增以及反函数单增等性质，Sigmoid函数常被用作神经网络的阈值函数，将变量映射到0,1之间）作为卷积网络的激活函数，使得特征映射具有位移不变性。此外，由于一个映射面上的神经元共享权值，因而减少了网络自由参数的个数。卷积神经网络中的每一个卷积层都紧跟着一个用来求局部平均与二次提取的计算层，这种特有的两次特征提取结构减小了特征分辨率。上述第一CNN模型是一种训练好的模型，即，无需重新通过图片样本进行训练，而是拿来主义，将其它领域或场景下训练完成的CNN模型拿过来直接使用。本实施例中，上述第一CNN模型包括VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、 Xception模型、DenseNet模型和AlexNet模型中的一种。
在上述冻结单元20中，第一CNN模型中包括多层（例如，卷基层和全连接层），根据预设规则，可以将其中至少一个层作为指定层（不包括最后一层），并冻结指定层的权重参数。
在一种可能的实施方式中，由于第一CNN模型中的多个卷基层和多个全连接层存在先后顺序，可以按照各层之间的先后顺序，将排序前N的卷基层或/和全连接层作为指定层，其中N为大于1小于第一CNN模型总层数的正整数，且N为预设值。也就是说，将第一CNN模型的前指定层数的权重参数冻结，即，被冻结的权重参数，在之后对牲畜图片进行处理时，不会进行改变。在一具体实施例中，上述第一CNN模型是ImageNet的使用VGG19训练后得到模型。上述ImageNet是一个计算机视觉***识别项目名称，是目前世界上图像识别最大的数据库，是美国斯坦福的计算机科学家，模拟人类的识别***建立的，其中包含120万张图片，分别包含1000种类别的图片，其中的动物，比如猫，狗，牛等牲畜的图片数量巨大。本实施例中，上述VGG19模型一共包括19层神经网络层，其中，其前16层为卷积层，最后三层为全连接层。在一具体实施例中，冻结前16层卷积层和前2层全连接层的权重参数，即冻结VGG19模型的前十八层的权重参数。
在上述处理单元30中，上述待识别的牲畜图片是指养殖户等按照一定要求拍摄的图片，比如，牲畜为猪，通常会拍摄猪的正面照或侧面照等含有猪的特征较多的照片，而不会拍摄猪的局部照片，比如只拍摄肚子的照片等。将待识别的牲畜图片输入到第一CNN模型中进行处理，由于其指定层数的权重值被冻结，所以只能在未被冻结的网络层进行处理，处理速度会更快。在一具体实施例中，第一CNN模型为ImageNet比赛中训练完成的VGG19模型，其冻结了前18层网络层，所以，在处理过程中，只需要将待识别的牲畜图片输入到VGG19模型的最后一层进行处理即可，训练完成输出的结果即为被识别牲畜的特征向量。在处理过程中，直到其最后一层的神经网络的损失函数变为一个接近0的阈值时，结束识别的过程，输出识别结果。可以这么理解，随着VGG19的最后一个全连接层参数的不断变化，神经网络的损失函数会越来越小，直到收敛到接近0，此时，可以认为输出的待识别牲畜图片的特征向量为到达要求的特征向量，理论上，神经网络的损失函数的收敛值为0时，输出的特征向量最优。
参照图5，本实施例中，上述牲畜特征向量的提取装置，还包括：
获取节点单元301，用于获取所述第一CNN模型最后一层的节点数量，根据所述节点数量生成最大输入所述待识别的牲畜图片数量的提醒信息。
在上述获取节点单元101中，上述第一CNN模型的最后一层的节点数数量对应着其最大可以处理图片的张数，因为不同的第一CNN模型的最后一层的节点数量不同，其可以进行处理的图片张数也不同。在一具体实施例中，上述第一CNN模型为ImageNet比赛中训练完成的VGG19模型，其最后一层的节点数量为4096，那么其最多能对4096张图片进行处理，如果超出4096张图片，则不适用VGG19模型，需要选择其它的模型。本实施例中，在输入牲畜图片之前会见先判断上述第一CNN模型的最后一层的节点数数量，并通过显示、语音等形式告知当前的操作者，以防止当前操作者输入过多的图片而无法完成相应的训练。具体的判断过程为：读取当前的第一CNN模型的网络结构，然后在预设的模型数据库中查找与该网络结构相同的模型，进而判断出具体的模型，然后根据该模型，在预设列表中查找与其对应的最后一层的节点数量。在另一实施例中，用户的计算机设备中预存有多个不同的第一CNN模型，如分别为训练后的VGG19模型、VGG16模型、VGG-F模型、ResNet50模型、ResNet152模型、DPN131模型、InceptionV3模型、 Xception模型、DenseNet模型和AlexNet模型等，并给每一个模型添加一个标签，该标签对应模型的最后一层的节点数量值，当使用某一具体模型时，根据其标签查找与之对应的节点数量值，从而知道当前最多可输入多少张待识别的牲畜图片。
参照图5，本实施例中，上述牲畜特征向量的提取装置，还包括：
归一单元40，用于对待识别的牲畜图片对应的特征向量进行softmax归一化处理。
在归一单元40中，对待识别的牲畜图片对应的特征向量进行softmax归一化处理，会得出概率分布输出，方便计算。
参照图6，本实施例中，上述牲畜特征向量的提取装置，还包括：
重置单元31，用于将所述第一CNN模型中未冻结的各层的权重参数重置为0。
在重置单元301中，将第一CNN模型中未冻结的各层神经网络的权重参数重置为0，会提高训练的速度，而对结果的准确性并没有很大的应影响，即对结果的影响在可以接受的范围之内。在一具体实施例中，第一CNN模型是为ImageNet比赛中训练完成的VGG19模型，其前18层的权重参数被冻结，那么在识别图片时，将其最后一层的权重参数重置为零，以提高处理速度。
参照图7，本实施例中，上述牲畜特征向量的提取装置，还包括：
获取发布单元11，用于获取指定区域养殖户的名单，并根据名单分配待识别单一牲畜种类的脸部图片采集任务给各养殖户；
收集存储单元12、收集养殖户反馈的图片，并存储到指定的存储空间内形成所述图片样本；
分类训练单元13、用于将所述图片样本分成训练集和测试集，基于指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型。
在上述获取发布单元11中，上述指定区域的养殖户名单是指可以在一指定范围内获取到的养殖户名单。比如，投保了牲畜险的养殖户的信息会被保险公司以电子文档的形式进行备案，当获取养殖户的名单时，可以登录存储牲畜保单的数据库，然后读取数据库中电子文档的养殖户信息，以得到养殖户的联系方式。这里养殖户的联系方式包括两种，一种是手机号码，另一种是带即时通讯功能的应用程序（APP）的账户，如电子投保的APP或者如微信、QQ等即时通讯APP。如养殖户的联系方式是手机号码，则通过短信的形式发送采集任务给养殖户。如果是电子投保的APP，则通过APP发送采集任务给养殖户。在本实施例中，上述牲畜特征向量的提取装置应用于牲畜保险领域，所以会开发一款对应的应用程序，其集成有牲畜投保、报险***，以及即时通讯***，通过该应用程序投保的用户的信息会自动保存到预设的数据库中，当发布任务时，直接通过上述应用程序进行发布即可。上述待识别单一牲畜种类是指单一种类的牲畜，比如猪、牛、羊等中的单一物种。上述的脸部图片是指方便识别的，具有单一物种特征较多的照片，一般为牲畜的正脸照片，如猪的正脸照片等。在另一具体实施例中，还以通过众包平台发布众包任务，众包任务即为收集待识别单一牲畜种类的脸部图片，而众包用户即为上述获取的指定区域养殖户的名单上对应的养殖户。
在上述收集存储单元12中，将养殖户发送的图片保存到指定的存储空间内，如指定的数据库中等。上述形成所述图片样本，是指这些图片将会作为深度神经网络处理的样本进行使用，同样其会将图片样本先先进行向量化，使其与图片对应，以便于输入到上述第二CNN模型中训练。本实施例中，在接收到养殖户的图片之后，会对图片进行简单的筛选，其筛选规律为图片的质量是否达到预设要求，具体地过程在上述方法实施例中已经描述，不再赘述。
在上述分类训练单元13中，将上述的样本图片分成训练集和测试集，现将测试集的图片输入到上述第二CNN模型中进行训练，该第二CNN模型是指模型的初始状态，即没有输入过任何数据进行运算的模型。当第二CNN模型经过上述训练集的图片、以及图片对应特征向量训练完成后，通过测试集的图片、以及图片对应特征向量对该训练完成的模型进行检验，如果检验合格，则认定该训练完成的模型是可以使用的模型，即上述的第一CNN模型。本实施例中，通过养殖户提供待识别的牲畜种类的图片，建立的第一CNN模型会更加准确。比如，养殖户提供的图片全部是猪的图片，那么训练出来的第一CNN模型在对猪的图片识别上，相对会更加准确。
参照图7，本实施例中，上述牲畜特征向量的提取装置，还包括：
判断阈值单元131，用于判断所述图片样本的数量是否达到预设值；
停止发布单元132，用于若达到预设值，则停止发送图片获取任务。
在上述判断阈值单元131和停止发布单元132中，图片样本的数量达到预设的值之后，无需继续获取图片样本，则停止发送图片获取任务。本实施例中，上述的图片获取任务一般是通过众包任务的形式进行发送的，即需要进行付费，当获取的图片达到指定数量之后，既可以停止发布任务，节约金钱等。本实施例中，可以不付众包用户（养殖户）金钱，可以根据养殖户提供图片的数量，对其投保时提供打折等优惠政策，以实现互利互惠。
本实施例中，上述的待识别单一牲畜种类的脸部图片采集任务包括专门采集病死猪的图片的任务，建立一个病死猪的猪脸图片集合；然后利用病死猪的猪脸图片集合训练针对病死猪的第一CNN模型，用于判断被识别的猪只是否是病死猪。因为，养殖户在投保过程，为了节约保费，可能只给部分猪只进行投保，而另一部分猪只不进行投保。在投保时，养殖户进行线上投保，具体地，养殖户登录投保***，进行注册（填写养殖户的名称、商号、联系方式等），然后通过上传投保猪只的图片对投保猪只进行标记，当投保猪只发生病死状态时，拍摄其病死后的图片上传投保***进行报险，通过上述第一CNN模型判断病死后的图片中的猪只是否为投保猪只。此时存在一种情况，即养殖户未投保的猪只发生病死情况，其将投保过的且活着的猪只进行拍照以欺骗***，使***认为病死的猪只是投保过的猪只，进而使保险公司赔偿保险费用。本实施例中，建立一个专门识别病死猪的第一CNN模型，当养殖户上传投保猪只的图片之后，首先判断该图片中的猪只是否为病死状态的猪只，如果是，才进行判断该猪只是否为投保猪只，降低保险公司的风险。
本申请的实施例中的牲畜特征向量的提取装置，基于迁移学习的理念，利用已经训练过的第一CNN模型，结合fine-tune方法，在识别牲畜过程中，只需要对第一CNN模型的最后几层进行处理，其它层的权重参数可以冻结，即，识别牲畜图片的过程中，第一CNN模型被冻结的层不需要参与处理的过程，进而大大地提高了牲畜图片识别的速度。
参照图8，本申请实施例中还提供一种计算机设备，该计算机设备可以是服务器，其内部结构可以如图8所示。该计算机设备包括通过***总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作***、计算机可读指令和数据库。该内存器为非易失性存储介质中的操作***和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储识别牲畜的第一CNN模型等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现上述各方法实施例的流程。
本申请一实施例还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，计算机可读指令被处理器执行时实现上述各方法实施例的流程。
以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种牲畜特征向量的提取方法，其特征在于，包括：
    获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
    冻结所述第一CNN模型中指定层的权重参数；
    将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
  2. 根据权利要求1所述的牲畜特征向量的提取方法，其特征在于，所述第一CNN模型包括VGG19模型，所述冻结所述第一CNN模型中指定层的权重参数的步骤，包括：
    冻结所述第一CNN模型中前十八层的权重参数。
  3. 根据权利要求1所述的牲畜特征向量的提取方法，其特征在于，所述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤之前，包括：
    获取所述第一CNN模型最后一层的节点数量，根据所述节点数量生成最大输入所述待识别的牲畜图片数量的提醒信息。
  4. 根据权利要求1所述的牲畜特征向量的提取方法，其特征在于，所述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤之后，包括：
    对所述待识别的牲畜图片对应的特征向量进行softmax归一化处理。
  5. 根据权利要求1所述的牲畜特征向量的提取方法，其特征在于，所述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤之前，包括：
    将所述第一CNN模型中未冻结的各层的权重参数重置为0。
  6. 根据权利要求1所述的牲畜特征向量的提取方法，其特征在于，所述获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型的步骤之前，包括：
    获取指定区域养殖户的名单，并根据名单分配待识别单一牲畜种类的脸部图片采集任务给各养殖户；
    收集养殖户反馈的图片，并存储到指定的存储空间内形成所述图片样本；
    将所述图片样本分成训练集和测试集，基于指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型。
  7. 根据权利要求6所述的牲畜特征向量的提取方法，其特征在于，所述将所述图片样本分成训练集和测试集，基于到指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型的步骤之前，包括：
    判断所述图片样本的数量是否达到预设值；
    若达到，则停止发送图片获取任务。
  8. 一种牲畜特征向量的提取装置，其特征在于，包括：
    获取单元，用于获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
    冻结单元，用于冻结所述第一CNN模型的前指定层数的权重参数；
    处理单元，用于将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
  9. 根据权利要求8所述的牲畜特征向量的提取装置，其特征在于，所述第一CNN模型包括VGG19模型，所述冻结单元，具体用于冻结所述第一CNN模型中前十八层的权重参数。
  10. 根据权利要求1所述的牲畜特征向量的提取装置，其特征在于，所述牲畜特征向量的提取装置，还包括：
    获取节点单元，用于获取所述第一CNN模型最后一层的节点数量，根据所述节点数量生成最大输入所述待识别的牲畜图片数量的提醒信息。
  11. 根据权利要求8所述的牲畜特征向量的提取装置，其特征在于，所述牲畜特征向量的提取装置，还包括：
    归一单元，用于对所述待识别的牲畜图片对应的特征向量进行softmax归一化处理。
  12. 根据权利要求8所述的牲畜特征向量的提取装置，其特征在于，所述牲畜特征向量的提取装置，还包括：
    重置单元，用于将所述第一CNN模型中未冻结的各层的权重参数重置为0。
  13. 根据权利要求8所述的牲畜特征向量的提取装置，其特征在于，所述牲畜特征向量的提取装置，还包括：
    获取发布单元，用于获取指定区域养殖户的名单，并根据名单分配待识别单一牲畜种类的脸部图片采集任务给各养殖户；
    收集存储单元，用于收集养殖户反馈的图片，并存储到指定的存储空间内形成所述图片样本；
    分类训练单元，用于将所述图片样本分成训练集和测试集，基于指定的第二CNN模型进行训练，得到训练完成的所述第一CNN模型。
  14. 根据权利要求13所述的牲畜特征向量的提取装置，其特征在于，所述牲畜特征向量的提取装置，还包括：
    判断阈值单元，用于判断所述图片样本的数量是否达到预设值；
    停止发布单元，用于若达到预设值，则停止发送图片获取任务。
  15. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现牲畜特征向量的提取方法，该提取方法包括：
    获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
    冻结所述第一CNN模型中指定层的权重参数；
    将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
  16. 根据权利要求15所述的计算机设备，其特征在于，所述第一CNN模型包括VGG19模型，所述冻结所述第一CNN模型中指定层的权重参数的步骤，包括：
    冻结所述第一CNN模型中前十八层的权重参数。
  17. 根据权利要求15所述的计算机设备，其特征在于，所述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤之前，包括：
    获取所述第一CNN模型最后一层的节点数量，根据所述节点数量生成最大输入所述待识别的牲畜图片数量的提醒信息。
  18. 根据权利要求15所述的计算机设备，其特征在于，所述将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量的步骤之后，包括：
    对所述待识别的牲畜图片对应的特征向量进行softmax归一化处理。
  19. 一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现牲畜特征向量的提取方法，所述提取方法包括：
    获取基于指定图片样本、以及所述图片样本对应的特征向量训练完成的第一CNN模型；
    冻结所述第一CNN模型中指定层的权重参数；
    将待识别的牲畜图片输入到所述第一CNN模型中进行处理，得出所述待识别的牲畜图片对应的特征向量。
  20. 根据权利要求19所述的计算机非易失性可读存储介质，其特征在于，所述第一CNN模型包括VGG19模型，所述冻结所述第一CNN模型中指定层的权重参数的步骤，包括：
    冻结所述第一CNN模型中前十八层的权重参数。
PCT/CN2018/095317 2018-04-17 2018-07-11 牲畜特征向量的提取方法、装置、计算机设备和存储介质 WO2019200735A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810344668.6A CN108805137A (zh) 2018-04-17 2018-04-17 牲畜特征向量的提取方法、装置、计算机设备和存储介质
CN201810344668.6 2018-04-17

Publications (1)

Publication Number Publication Date
WO2019200735A1 true WO2019200735A1 (zh) 2019-10-24

Family

ID=64094327

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095317 WO2019200735A1 (zh) 2018-04-17 2018-07-11 牲畜特征向量的提取方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108805137A (zh)
WO (1) WO2019200735A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111369372A (zh) * 2020-03-03 2020-07-03 福州农福腾信息科技有限公司 基于猪体识别的保险理赔方法和后台端
CN111639629A (zh) * 2020-06-15 2020-09-08 安徽工大信息技术有限公司 一种基于图像处理的猪只体重测量方法、装置及存储介质
CN112734731A (zh) * 2021-01-11 2021-04-30 牧原食品股份有限公司 一种牲畜温度检测方法、装置、设备和存储介质
CN115862060A (zh) * 2022-11-25 2023-03-28 天津大学四川创新研究院 基于猪脸识别与猪只重识别的猪只唯一化识别方法及***

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754068A (zh) * 2018-12-04 2019-05-14 中科恒运股份有限公司 基于深度学习预训练模型的迁移学习方法及终端设备
CN109726824A (zh) * 2018-12-05 2019-05-07 中科恒运股份有限公司 训练模型的迁移学习方法及终端设备
CN111339306B (zh) * 2018-12-18 2023-05-12 腾讯科技（深圳）有限公司 分类模型训练方法、分类方法及装置、设备和介质
CN109446332B (zh) * 2018-12-25 2023-08-25 银江技术股份有限公司 一种基于特征迁移和自适应学习的人民调解案例分类***及方法
CN110163812B (zh) * 2019-04-10 2023-11-07 深圳大学 肌肉超声图像降噪方法、装置及存储介质
CN110543815B (zh) * 2019-07-22 2024-03-08 平安科技（深圳）有限公司 人脸识别模型的训练方法、人脸识别方法、装置、设备及存储介质
CN112183265A (zh) * 2020-09-17 2021-01-05 国家电网有限公司 基于图像识别的电力施工视频监控告警方法及***
CN112184322A (zh) * 2020-10-13 2021-01-05 中国农业银行股份有限公司四川省分行 一种基于图片识别的活体禽畜抵押贷款贷前估值方法
CN112637165B (zh) * 2020-12-14 2022-08-30 广东电网有限责任公司 模型训练方法、网络攻击检测方法、装置、设备及介质
CN113077485A (zh) * 2021-04-28 2021-07-06 北京神州慧达信息技术有限公司 基于人工智能的无参考物的测量方法
CN113793121A (zh) * 2021-09-15 2021-12-14 平安国际智慧城市科技股份有限公司 诉讼案件的自动立案方法、装置、电子设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104200224A (zh) * 2014-08-28 2014-12-10 西北工业大学 基于深度卷积神经网络的无价值图像去除方法
CN104268524A (zh) * 2014-09-24 2015-01-07 朱毅 一种基于动态调整训练目标的卷积神经网络的图像识别方法
CN104268521A (zh) * 2014-09-23 2015-01-07 朱毅 一种基于非限定类别的卷积神经网络的图像识别方法
CN104991959A (zh) * 2015-07-21 2015-10-21 北京京东尚科信息技术有限公司 一种基于内容检索相同或相似图像的方法与***

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104200224A (zh) * 2014-08-28 2014-12-10 西北工业大学 基于深度卷积神经网络的无价值图像去除方法
CN104268521A (zh) * 2014-09-23 2015-01-07 朱毅 一种基于非限定类别的卷积神经网络的图像识别方法
CN104268524A (zh) * 2014-09-24 2015-01-07 朱毅 一种基于动态调整训练目标的卷积神经网络的图像识别方法
CN104991959A (zh) * 2015-07-21 2015-10-21 北京京东尚科信息技术有限公司 一种基于内容检索相同或相似图像的方法与***

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111369372A (zh) * 2020-03-03 2020-07-03 福州农福腾信息科技有限公司 基于猪体识别的保险理赔方法和后台端
CN111639629A (zh) * 2020-06-15 2020-09-08 安徽工大信息技术有限公司 一种基于图像处理的猪只体重测量方法、装置及存储介质
CN111639629B (zh) * 2020-06-15 2024-02-23 安徽工大信息技术有限公司 一种基于图像处理的猪只体重测量方法、装置及存储介质
CN112734731A (zh) * 2021-01-11 2021-04-30 牧原食品股份有限公司 一种牲畜温度检测方法、装置、设备和存储介质
CN112734731B (zh) * 2021-01-11 2023-07-28 牧原食品股份有限公司 一种牲畜温度检测方法、装置、设备和存储介质
CN115862060A (zh) * 2022-11-25 2023-03-28 天津大学四川创新研究院 基于猪脸识别与猪只重识别的猪只唯一化识别方法及***
CN115862060B (zh) * 2022-11-25 2023-09-26 天津大学四川创新研究院 基于猪脸识别与猪只重识别的猪只唯一化识别方法及***

Also Published As

Publication number Publication date
CN108805137A (zh) 2018-11-13

Similar Documents

Publication Publication Date Title
WO2019200735A1 (zh) 牲畜特征向量的提取方法、装置、计算机设备和存储介质
EP3333768A1 (en) Method and apparatus for detecting target
CN109086811B (zh) 多标签图像分类方法、装置及电子设备
CN109002766B (zh) 一种表情识别方法及装置
WO2020228446A1 (zh) 模型训练方法、装置、终端及存储介质
CN110135231B (zh) 动物面部识别方法、装置、计算机设备和存储介质
CN110909618B (zh) 一种宠物身份的识别方法及装置
CN108985190B (zh) 目标识别方法和装置、电子设备、存储介质
CN110705489B (zh) 目标识别网络的训练方法、装置、计算机设备和存储介质
CN112884782B (zh) 生物对象分割方法、装置、计算机设备和存储介质
CN112862757A (zh) 一种基于计算机视觉技术的重量评估***及实现方法
CN111144398A (zh) 目标检测方法、装置、计算机设备和存储介质
CN111459922A (zh) 用户识别方法、装置、设备及存储介质
CN111507288A (zh) 图像检测方法、装置、计算机设备和存储介质
CN110809768B (zh) 数据清洗***和方法
CN113706558A (zh) 图像分割方法、装置及计算机设备
CN111783997B (zh) 一种数据处理方法、装置及设备
CN111382791B (zh) 深度学习任务处理方法、图像识别任务处理方法和装置
CN113743426A (zh) 一种训练方法、装置、设备以及计算机可读存储介质
CN113902944A (zh) 模型的训练及场景识别方法、装置、设备及介质
CN113762005B (zh) 特征选择模型的训练、对象分类方法、装置、设备及介质
CN112818946A (zh) 年龄识别模型的训练、年龄识别方法、装置及电子设备
CN113569081A (zh) 图像识别方法、装置、设备及存储介质
CN112861743A (zh) 一种掌静脉图像防伪方法、装置和设备
CN117253192A (zh) 用于桑蚕养殖的智能***及方法

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: 18915591

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: 18915591

Country of ref document: EP

Kind code of ref document: A1