WO2019227617A1 - 动物种类的分类方法、装置、计算机设备和存储介质 - Google Patents

动物种类的分类方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019227617A1
WO2019227617A1 PCT/CN2018/095668 CN2018095668W WO2019227617A1 WO 2019227617 A1 WO2019227617 A1 WO 2019227617A1 CN 2018095668 W CN2018095668 W CN 2018095668W WO 2019227617 A1 WO2019227617 A1 WO 2019227617A1
Authority
WO
WIPO (PCT)
Prior art keywords
animal
classified
convolutional neural
training
vector
Prior art date
Application number
PCT/CN2018/095668
Other languages
English (en)
French (fr)
Inventor
马进
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 WO2019227617A1 publication Critical patent/WO2019227617A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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

Definitions

  • the present application relates to the field of computer technology, and in particular, to a method, an apparatus, a computer device, and a storage medium for classifying animal species.
  • the main purpose of this application is to provide a method, device, computer equipment and storage medium for classifying animal species, which overcomes the defect of low accuracy in identifying animal species.
  • this application provides a method for classifying animal species, which includes the following steps:
  • the merged vector is input into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • This application also provides a device for classifying animal species, including:
  • a first extraction unit configured to obtain an animal image to be classified, and extract feature vectors of the animal image through multiple different network models
  • a first stitching unit configured to stitch the feature vectors extracted by the multiple different network models into one merge vector
  • a first classification unit is configured to input the merged vector into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • the present application further provides a computer device including a memory and a processor, where the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, implements the steps of any of the foregoing methods.
  • the present application also provides a computer non-volatile readable storage medium having computer-readable instructions stored thereon, which are executed by a processor to implement the steps of the method according to any one of the foregoing.
  • the animal species classification method, device, computer equipment, and storage medium provided in this application, obtain animal images to be classified, and extract feature vectors of the animal images respectively through multiple different network models; connect multiple different networks
  • the feature vectors extracted by the model are spliced into a merged vector; the merged vector is input into the trained convolutional neural network to output the classification result of the animal image to be classified, and the accuracy rate of identifying animal species is improved To overcome the current shortcomings of low recognition accuracy.
  • FIG. 1 is a schematic diagram of steps of a method for classifying animal species in an embodiment of the present application
  • FIG. 2 is a schematic diagram of steps of a method for classifying an animal species in another embodiment of the present application.
  • FIG. 3 is a schematic structural diagram of an animal species classification device according to an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of an animal species classification device according to another embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of an animal species classification device according to another embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of an animal species classification device according to another embodiment of the present application.
  • FIG. 7 is a schematic block diagram of a structure of a computer device according to an embodiment of the present application.
  • a schematic diagram of steps of a method for classifying animal species provided in an embodiment of the present application includes the following steps:
  • step S1 an animal image to be classified is obtained, and feature vectors of the animal image are separately extracted through a plurality of different network models.
  • the image of the animal to be classified is a photo of the animal that needs to be identified.
  • the animal in this embodiment is described using a dog as an example.
  • the network models in this embodiment are all trained network models, and the training process will be described below.
  • the last layer of the above network model is usually a softmax classification layer, which has a softmax classifier.
  • the output after passing through the softmax layer is the classification result of the input data, but before the last layer of the network model (that is, the softmax classification layer).
  • the output is a feature vector. It can be understood that the above feature vector is input in the softmax classification layer, and the feature vector can be output by inputting the feature vector to the softmax classification layer.
  • feature vectors of the animal images are separately extracted through multiple different network models, and the classification results are not directly output through the softmax classification layer.
  • the above-mentioned feature vectors can be extracted by using multiple different network models to extract the output results before their softmax classification layers, respectively.
  • each network model can extract a feature vector. Due to the different structure of the network model, the feature vectors it extracts are also different.
  • Step S2 stitch the feature vectors extracted by the multiple different network models into a merge vector.
  • the feature vectors extracted by the multiple different network models are spliced end-to-end into a merge vector, so that the merge vector includes the features of all the feature vectors described above.
  • the above merge vector is also a vector, and the stitched merge vector is used as the network input of the convolutional neural network model.
  • step S3 the merged vector is input into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • the above-mentioned convolutional neural network is a trained model.
  • the merged vector obtained by the above stitching is input to the trained convolutional neural network for learning, and the output result of the convolutional neural network is the classification result of the animal image to be classified.
  • the feature vectors extracted by each of the above network models are different, the feature vectors obtained after splicing have different characteristics of the above feature vectors. If the above feature vectors are input into the convolutional neural network singly, the classification results obtained are different because of the different feature vectors.
  • the merged vector is input to a convolutional neural network for calculation, which is consistent with the calculation principle when a single feature vector is input, except that the input vectors are different.
  • the merged vector of the above three feature vectors is a new vector.
  • the merged vector is convolved and finally classified.
  • the classification results are combined with the characteristics of different feature vectors to obtain The classification result is more accurate than the recognition classification when the above feature vectors are input separately. Therefore, the classification accuracy rate of animal species is improved, and the shortcoming of the current low identification accuracy rate is overcome.
  • each fully connected layer in the above convolutional neural network is set to: three fully connected layers, of which the first fully connected layer It has 4096 nodes, the second fully-connected layer has 512 nodes, and the third fully-connected layer has 120 nodes.
  • Linear correction layers (Relu) are added before the second fully-connected layer and the third fully-connected layer, respectively.
  • Rectified Linear Unit linear rectification function, also known as linear correction unit
  • Dropout deep learning layer
  • the network model includes a DenseNet (Densely Connected Convolutional Networks), a ResNet (Deep Residual Network), and an Inception network model.
  • the feature vectors of the animal images to be classified are respectively extracted through the above three network models, which respectively correspond to Fd, Fr, Fi; wherein the above three features are each a vector with 1024 numbers.
  • the same dog picture is input to the three different network models of DenseNet, ResNet, and Inception, and the extracted feature vectors are also different, that is, the 1024 numbers in the vector are different. If only one network model is input, , The extracted feature vectors are relatively one-sided, and the information is not comprehensive enough. When three or more network models are used for extraction, the feature vectors express the feature information of the dog more comprehensively, which is more conducive to the classification and recognition of subsequent dog species.
  • the same dog picture is input to the three network models of DenseNet, ResNet, and Inception, and is calculated by multi-layer convolutional layers and then in the penultimate layer (that is, the previous layer of the softmax classification layer).
  • the above feature vectors are extracted (the last layer outputs the classification results, and in this embodiment, the output is not passed through the last layer). Therefore, the result of extracting the output of the penultimate layer is the above-mentioned feature vector.
  • the outputs of all previous convolutional layers are connected to all subsequent convolutional layers.
  • This connection method makes the transfer of features and gradients more efficient, and the network is easier to train. Since each layer is directly connected to the input (input layer) and loss (loss function layer), the phenomenon of gradient disappearance can be reduced.
  • the network model has the characteristics of effectively solving the vanishing gradient problem, enhancing feature propagation, supporting feature reuse, and greatly reducing the number of parameters.
  • the input and output of the block are subjected to an element-wise superposition by shortcut.
  • This simple addition does not add additional parameters and calculations to the network, but it can greatly increase
  • the training speed of the model improves the training effect, and when the number of layers of the model deepens, this simple structure can well solve the degradation problem.
  • the input of the next layer is the output of the previous layer plus the non-linear transformation of the output of the previous layer. It solves the problem of degradation, and the deeper the network error rate on both the training set and the verification set is, the smaller it is.
  • the traditional convolutional layer in the network is modified to increase the depth and width of the network and improve the performance of the deep neural network.
  • stacking 1x1, 3x3, 5x5 convolution kernels and 3x3 maxpooling together increases the network width on the one hand, and convolution kernels of different sizes on the other can enhance the adaptability of the network.
  • a 1x1 convolution kernel is added before the 3x3 convolution kernel, before the 5x5 convolution kernel, and after max pooling.
  • the method before step S1 of obtaining the animal image to be classified and extracting feature vectors of the animal image through multiple different network models, the method includes:
  • Step S101 input each sample picture in a training set into a plurality of different network models to train each of the network models.
  • a dog data set (a large number of dog pictures) of a known dog classification is divided into a training set and a test set in advance.
  • the training set is used to train the network model and the convolutional neural network
  • the test set is used to train the The completed network model and convolutional neural network are tested to verify whether the trained model is accurate.
  • the ratio of the sample data amount of the training set and the test set is 8: 2, that is, when the dog's data set includes 10,000 sample pictures, 8,000 of which are used as the training set and 2000 are used as the test set.
  • training parameters of the above network model need to be obtained. Therefore, each sample picture in the training set needs to be input into the multiple network models and trained separately to obtain corresponding training parameters.
  • the network model after the training parameters are obtained can be used as the network model in the above step S1.
  • the training set is input to the network model for training, and the classification result is output through the last layer of the network model. Since the classifications of dogs in the training set are all known, when training the network model, as long as The classification result output by the last layer is close to / same as the classification of the dog in the training set, and the training of the above network model can be completed.
  • the trained network model is the network model used in the above step S1.
  • step S102 a feature vector of each sample picture in the training set is extracted through the multiple trained network models.
  • the feature vector of each dog picture in the training set is extracted through the trained network model in this step S102.
  • the dog picture in this step S102 may be the above step S101
  • the dog pictures in the used training set can also be other dog pictures in the training set.
  • the process of extracting the feature vector of the dog picture in the training set in this step S102 is similar to that in the above step S1, except that the target dog pictures are different, which will not be repeated here.
  • Step S103 stitch multiple feature vectors corresponding to each sample picture in the training set into a training merge vector.
  • the stitching in this step is to combine the multiple feature vectors corresponding to each of the sample pictures into a merge vector, and the above-mentioned training merge vector obtained by the stitching is input as training data of the convolutional neural network.
  • the stitching process in this step is the same as the stitching process in step S2 described above, and details are not described herein again.
  • Step S104 input the training merge vector into a convolutional neural network to train the convolutional neural network.
  • the training merge vector is input to the convolutional neural network for training, and the classification result is output through the last layer of the convolutional neural network. Since the classification of the dog in the training set corresponding to the training merge vector is known When training the above network model, as long as the classification result output by the last layer of the convolutional neural network is close to / same as the classification of the dog in the training set, the training of the above convolutional neural network can be completed.
  • the trained convolutional neural network is the convolutional neural network used in the above step S3.
  • step S104 of inputting the training merge vector into the convolutional neural network to train the convolutional neural network the method includes:
  • Step S105 input each sample picture in the test set into a plurality of trained network models to extract feature vectors of each sample picture in the test set separately;
  • Step S106 stitch multiple feature vectors corresponding to each sample picture in the test set into a test merge vector
  • Step S107 input the test merge vector into the convolutional neural network that has been trained to output the classification result of the test set, and verify whether it is consistent with the known classification of the test set.
  • the trained network model and the convolutional neural network are the network model and the convolutional neural network used in the above steps S1 and S3.
  • the trained network model and the convolutional neural network need to be tested.
  • the trained network model and the convolutional neural network are tested using a test set, and the picture data in the test set and the picture data in the training set are pictures of known animal species.
  • the specific implementation of steps S105, S106, and S107 in this embodiment is similar to the implementation process of steps S1, S2, and S3 described above, and details are not described herein again. The only difference is that the target animal image data is different, and in step S107 of this embodiment, it is necessary to compare whether the classification result of the output of the convolutional neural network is consistent with the known classification of the pictures in the test set.
  • step S3 of inputting the merged vector into the trained convolutional neural network to output a classification result of the animal image to be classified the method includes:
  • Step S41 Determine the type name of the animal in the animal image to be classified according to the classification result of the animal image to be classified;
  • step S51 according to the type name of the animal, the animal database information is queried out.
  • the above-mentioned classification and identification of animal species are applied in the field of science popularization. Specifically, after classifying an unknown type of dog according to the above method, the type of the dog in the unknown picture is determined, so as to determine the type name of the dog according to the type in step S41, as in step S51 in The animal database is queried for information about the animal. Among them, the above-mentioned animal database stores a large amount of science-related science-related data, such as life habits, physical characteristics, and major settlements.
  • the relevant material information is automatically matched from the animal database to facilitate pushing the information to the user, and the user can easily understand the unknown picture Information about dogs.
  • step S3 of inputting the merged vector into the trained convolutional neural network to output a classification result of the animal image to be classified the method includes:
  • Step S401 Determine a type name of an animal in the animal image to be classified according to a classification result of the animal image to be classified;
  • step S501 according to the type name of the animal, a policy matching the type name of the animal is queried in the policy database, and at least one of a policy amount and a compensation amount is obtained from the policy.
  • the aforementioned classification and identification of animal species is applied in the field of insurance.
  • more and more users are insuring their own dogs. Different dogs have different insurance policies and compensation amounts.
  • the name of the type of dog is determined according to the classification result; as described in step S501, According to the kind name of the dog, a policy matching the kind name of the animal is queried in the policy database, and at least one of a policy amount and a compensation amount is obtained from the policy.
  • the above policy database stores a large number of insurance policies for animals, and the policy records information such as the policy amount, compensation amount, and dog type name corresponding to different dogs.
  • the type of dog is needed to quickly obtain information such as the policy amount and compensation amount, which is convenient for users to apply for insurance.
  • step S3 of inputting the merged vector into the trained convolutional neural network to output a classification result of the animal image to be classified the method includes:
  • Step S4a Determine the type name of the animal in the animal image to be classified according to the classification result of the animal image to be classified;
  • step S5a according to the type name of the animal, a compensation case matching the type name of the animal is queried in the compensation case database, and a compensation amount is obtained from the compensation case.
  • the aforementioned classification and identification of animal species is applied in the field of insurance. Dogs are likely to bite others when they are outdoors, and different dogs have different compensation amounts. The user wants to obtain the compensation amount corresponding to different dogs. Then he can upload pictures of dogs and classify the dogs by the classification method in the above embodiment, and then classify the dogs as described in step S4a above. As a result, the type name of the dog is determined. Finally, as described in step S5a above, according to the type name of the dog, a compensation case matching the type name of the animal is queried from the compensation case database, and obtained from the compensation case. To the amount of compensation.
  • the method for classifying animal species provided in the embodiments of the present application is to obtain animal images to be classified, and to extract feature vectors of the animal images through multiple different network models, respectively.
  • the feature vectors extracted by the model are spliced into a merged vector; the merged vector is input to the trained convolutional neural network to output the classification result of the animal image to be classified, and the accuracy rate of identifying animal species is improved, Overcome the current shortcomings of low recognition accuracy.
  • an embodiment of the present application further provides a device for classifying animal species, including:
  • a first extraction unit 10 configured to obtain an animal image to be classified, and extract feature vectors of the animal image respectively through multiple different network models
  • the animal images to be classified are photos of animals that need to be identified.
  • the animals in this embodiment are described using dogs as an example.
  • the network models in this embodiment are all trained network models, and the training process will be described below.
  • the last layer of the above network model is usually a softmax classification layer, which has a softmax classifier.
  • the output after passing through the softmax layer is the classification result of the input data, but before the last layer of the network model (that is, the softmax classification layer)
  • the output is a feature vector. It can be understood that the above feature vector is input in the softmax classification layer, and the feature vector can be output by inputting the feature vector to the softmax classification layer.
  • the first extraction unit 10 extracts feature vectors of the animal image through multiple different network models, instead of directly outputting classification through the softmax classification layer. result. Specifically, the first extraction unit 10 extracts the output results before the softmax classification layer of each of them through a plurality of different network models, so as to extract the above feature vectors. For the same animal image, each network model can extract a feature vector. Due to the different structure of the network model, the feature vectors it extracts are also different.
  • a first stitching unit 20 configured to stitch the feature vectors extracted by multiple different network models into one merge vector
  • the first stitching unit 20 stitches the feature vectors extracted by the different network models into a merge vector in sequence, so that the merge vector includes the features of all the feature vectors described above.
  • the above merge vector is also a vector, and the stitched merge vector is used as the network input of the convolutional neural network model.
  • a first classification unit 30 is configured to input the merged vector into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • the above-mentioned convolutional neural network is a trained model.
  • the first classification unit 30 inputs the merged vector obtained by the stitching into a trained convolutional neural network, and the output result of the convolutional neural network is the classification result of the animal image to be classified.
  • the feature vectors extracted by the above network model are different, the feature vectors obtained after splicing have different characteristics of the above feature vectors. If the above feature vectors are input into the convolutional neural network singly, the classification results obtained are different because of the different feature vectors.
  • the merged vector is input to a convolutional neural network for calculation, which is consistent with the calculation principle when a single feature vector is input, except that the input vectors are different.
  • the merged vector of the above three feature vectors is a new vector.
  • the first classification unit 30 performs a convolution calculation on the merged vector, and finally performs classification.
  • the classification results incorporate different features. The characteristics of the vector, the obtained classification results are more accurate than the recognition and classification when the above feature vectors are input separately. Therefore, the classification accuracy rate of animal species is improved, and the shortcoming of the current low identification accuracy rate is overcome.
  • each fully connected layer in the above convolutional neural network is set to: three fully connected layers, of which the first fully connected layer It has 4096 nodes, the second fully-connected layer has 512 nodes, and the third fully-connected layer has 120 nodes.
  • Linear correction layers (Relu) are added before the second fully-connected layer and the third fully-connected layer, respectively.
  • Rectified Linear Unit linear rectification function, also known as linear correction unit
  • Dropout deep learning layer
  • the network models used in the first extraction unit 10 include DenseNet, ResNet, and Inception network models.
  • the feature vectors of the animal images to be classified are respectively extracted through the above three network models, which respectively correspond to Fd, Fr, Fi; wherein the above three features are each a vector with 1024 numbers.
  • the same dog picture is input to the three different network models of DenseNet, ResNet, and Inception, and the extracted feature vectors are also different, that is, the 1024 numbers in the vector are different. If only one network model is input, , The extracted feature vectors are relatively one-sided, and the information is not comprehensive enough. When three or more network models are used for extraction, the feature vectors express the feature information of the dog more comprehensively, which is more conducive to the classification and recognition of subsequent dog species.
  • the same dog picture is input to three network models of DenseNet, ResNet, and Inception, and after multi-layer convolution layer convolution calculation, the first extraction unit 10 is in the penultimate layer (that is, the softmax classification layer).
  • the above feature vector is extracted from the previous layer (the last layer outputs the classification result, and in this embodiment, it does not go through the last layer). Therefore, the result of extracting the output of the penultimate layer is the above-mentioned feature vector.
  • the outputs of all previous convolutional layers are connected to all subsequent convolutional layers.
  • This connection method makes the transfer of features and gradients more efficient, and the network is easier to train. Since each layer is directly connected to the input (input layer) and loss (loss function layer), the phenomenon of gradient disappearance can be reduced.
  • the network model has the characteristics of effectively solving the vanishing gradient problem, enhancing feature propagation, supporting feature reuse, and greatly reducing the number of parameters.
  • the input and output of the block are subjected to an element-wise superposition by shortcut.
  • This simple addition does not add additional parameters and calculations to the network, but it can greatly increase
  • the training speed of the model improves the training effect, and when the number of layers of the model deepens, this simple structure can well solve the degradation problem.
  • the input of the next layer is the output of the previous layer plus the non-linear transformation of the output of the previous layer. It solves the problem of degradation, and the deeper the network error rate on both the training set and the verification set is, the smaller it is.
  • the traditional convolutional layer in the network is modified to increase the depth and width of the network and improve the performance of the deep neural network.
  • stacking 1x1, 3x3, 5x5 convolution kernels and 3x3 maxpooling together increases the network width on the one hand, and convolution kernels of different sizes on the other can enhance the adaptability of the network.
  • a 1x1 convolution kernel is added before the 3x3 convolution kernel, before the 5x5 convolution kernel, and after max pooling.
  • the above-mentioned animal species classification device further includes:
  • the first training unit 101 is configured to input each sample picture in a training set into a plurality of different network models and train each of the network models.
  • a dog data set (a large number of dog pictures) of a known dog classification is divided into a training set and a test set in advance.
  • the training set is used to train the network model and the convolutional neural network
  • the test set is used to train the The completed network model and convolutional neural network are tested to verify whether the trained model is accurate.
  • the ratio of the sample data amount of the training set and the test set is 8: 2, that is, when the dog's data set includes 10,000 sample pictures, 8,000 of which are used as the training set and 2000 are used as the test set.
  • training parameters of the above network model need to be obtained. Therefore, the first training unit 101 needs to input each sample picture in the training set into the multiple network models and perform training separately to obtain corresponding training parameters.
  • the network model after the training parameters are obtained can be used as the network model used in the first extraction unit 10 described above.
  • the first training unit 101 inputs the training set into the network model for training, and outputs the classification result through the last layer of the network model. Since the classifications of the dogs in the training set are all known, the network is being trained. In the model, as long as the classification result output by the last layer is close to / same as the classification of the dog in the training set, the training of the above network model can be completed.
  • the trained network model is the network model used in the first extraction unit 10 described above.
  • the second extraction unit 102 is configured to extract a feature vector of each sample picture in the training set through a plurality of the network models that are completed by training.
  • the second extraction unit 102 uses the trained network model to extract the feature vector of each dog picture in the training set.
  • the dog picture in this process may be the first
  • the dog pictures in the training set used in the training unit 101 may also be other dog pictures in the training set.
  • the process of extracting the feature vector of the dog picture in the training set by the second extraction unit 102 is similar to the extraction process of the first extraction unit 10 described above, the difference is that the target dog pictures are different, and details are not described herein again.
  • the second stitching unit 103 is configured to stitch multiple feature vectors corresponding to each sample picture in the training set into a training merge vector.
  • the stitching in this step is to combine the multiple feature vectors corresponding to each of the sample pictures into a merge vector, and the above-mentioned training merge vector obtained by the stitching is input as training data of the convolutional neural network.
  • the stitching process of the second stitching unit 103 is the same as the stitching process of the first stitching unit 20 described above, and details are not described herein again.
  • the second training unit 104 inputs the training merge vector into a convolutional neural network to train the convolutional neural network.
  • the training parameters of the convolutional neural network described above need to be obtained. Therefore, the second training unit 104 needs to input the training merge vector into the convolutional neural network for training to obtain corresponding training parameters.
  • the convolutional neural network after the training parameters are obtained can be used as the convolutional neural network used in the first classification unit 30 described above.
  • the second training unit 104 inputs the training merge vector into the convolutional neural network for training, and outputs the classification result through the last layer of the convolutional neural network. Because the dog in the training set corresponding to the training merge vector The classification of is known. When training the above network model, as long as the classification result output by the last layer of the convolutional neural network is close to / same as the classification of the dog in the training set, the training of the above convolutional neural network can be completed.
  • the trained convolutional neural network is the convolutional neural network used in the first classification unit 30 described above.
  • the apparatus for classifying an animal species further includes:
  • a third extraction unit configured to input each sample picture in the test set into multiple trained network models to extract feature vectors of each sample picture in the test set separately;
  • a third stitching unit configured to stitch multiple feature vectors corresponding to each sample picture in the test set into a test merge vector
  • a verification unit is configured to input the test merge vector into the trained convolutional neural network to output the classification result of the test set, and verify whether it is consistent with the known classification of the test set.
  • the trained network model and the convolutional neural network are the network model and the convolutional neural network used in the first extraction unit 10 and the first classification unit 30 described above.
  • the trained network model and the convolutional neural network need to be tested.
  • the trained network model and the convolutional neural network are tested using a test set, and the picture data in the test set and the picture data in the training set are pictures of known animal species.
  • the specific implementation of the third extraction unit, the third splicing unit, and the verification unit in this embodiment is similar to the implementation process of the first extraction unit 10, the first splicing unit 20, and the first classification unit 30, and is not repeated here. .
  • the only difference is that the animal image data is different, and the verification unit in this embodiment needs to compare whether the classification result of the output of the convolutional neural network is consistent with the known classification of the pictures in the test set.
  • the apparatus for classifying an animal species further includes:
  • a first determining unit 40 configured to determine, according to a classification result of the animal image to be classified, a category name of the animal in the animal image to be classified;
  • the first querying unit 50 is configured to query the animal database information according to the type name of the animal.
  • the above-mentioned classification and identification of animal species are applied in the field of science popularization. Specifically, after classifying an unknown type of dog according to the above-mentioned classification device, the type of the dog in the unknown picture is determined, so that the first determining unit 40 determines the type name of the dog according to the type as described above. For example, the first query unit 50 queries the animal database for information about the animal. Among them, the above-mentioned animal database stores a large amount of science-related science-related data, such as life habits, physical characteristics, and major settlements. In this embodiment, after the dogs in the unknown picture are automatically classified by the machine, the relevant material information is automatically matched from the animal database to facilitate pushing the information to the user, and the user can easily understand the unknown picture Information about dogs.
  • the apparatus for classifying an animal species further includes:
  • a second determining unit 41 configured to determine, according to a classification result of the animal image to be classified, a type name of the animal in the animal image to be classified;
  • a second querying unit 51 is configured to query a policy database matching the type name of the animal in the policy database according to the type name of the animal, and obtain at least one of the policy amount and the compensation amount from the policy. Species.
  • the aforementioned classification and identification of animal species is applied in the field of insurance.
  • more and more users are insuring their own dogs. Different dogs have different insurance policies and compensation amounts.
  • the second determination unit 41 determines the type name of the dog according to the classification result; the second query unit 51 Then, according to the type name of the dog, an insurance policy matching the type name of the animal is queried in the policy database, and at least one of the policy amount and the compensation amount is obtained from the policy.
  • the above policy database stores a large number of insurance policies for animals, and the policy records information such as the policy amount, compensation amount, and dog type name corresponding to different dogs.
  • the type of dog is needed to quickly obtain information such as the policy amount and compensation amount, which is convenient for users to apply for insurance.
  • the apparatus for classifying an animal species further includes:
  • a third determining unit configured to determine, according to a classification result of the animal image to be classified, a type name of the animal in the animal image to be classified;
  • a third query unit is configured to query a compensation case database that matches the type name of the animal according to the type name of the animal, and obtain a compensation amount from the compensation case.
  • the aforementioned classification and identification of animal species is applied in the field of insurance. Dogs are likely to bite others when they are outdoors, and different dogs have different compensation amounts. The user wants to obtain the compensation amount corresponding to different dogs. Then he can upload pictures of dogs and classify the dogs by the classification method in the above embodiment. Then, according to the third determination unit described above, according to the dogs, The classification result of the animal, determine the type name of the dog; finally, as described in the third query unit, according to the type name of the dog, query the compensation case database that matches the type name of the animal, and from the compensation, The amount of compensation obtained in the case.
  • the first extraction unit 10 obtains an image of an animal to be classified, and extracts feature vectors of the animal image through multiple different network models, respectively;
  • a stitching unit 20 stitches the feature vectors extracted by a plurality of different network models into a merge vector;
  • the first classification unit 30 inputs the merge vector into a convolutional neural network to output the animal image to be classified
  • the classification results can improve the accuracy of identifying animal species and overcome the current shortcomings of low accuracy.
  • an embodiment of the present application further provides a computer device.
  • the computer device may be a server, and its internal structure may be as shown in FIG.
  • the computer device includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the computer design processor is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in a non-volatile storage medium.
  • the computer equipment database is used to store data such as network models.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by a processor to implement a method of classifying an animal species.
  • the processor executes the steps of the method for classifying an animal species:
  • the merged vector is input into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • the above network models include DenseNet, ResNet, and Inception network models.
  • the processor before the step of acquiring, by the processor, an animal image to be classified and extracting feature vectors of the animal image through a plurality of different network models, the processor includes:
  • the training merge vector is input into a convolutional neural network to train the convolutional neural network.
  • the method includes:
  • the test merge vector is input into the convolutional neural network that has been trained to output the classification result of the test set, and it is verified whether it is consistent with the known classification of the test set.
  • the ratio of the sample data amount of the training set to the test set is 8: 2.
  • the method includes:
  • the animal database information is queried in the animal database.
  • the method includes:
  • a policy matching the kind name of the animal is queried in the policy database, and at least one of a policy amount and a compensation amount is obtained from the policy.
  • FIG. 7 is only a block diagram of a part of the structure related to the solution of the application, and does not constitute a limitation on the computer equipment to which the solution of the application is applied.
  • An embodiment of the present application further provides a computer non-volatile readable storage medium having computer readable instructions stored thereon.
  • a method for classifying an animal species is specifically:
  • the merged vector is input into a trained convolutional neural network to output a classification result of the animal image to be classified.
  • the above network models include DenseNet, ResNet, and Inception network models.
  • the processor before the step of acquiring, by the processor, an animal image to be classified and extracting feature vectors of the animal image through a plurality of different network models, the processor includes:
  • the training merge vector is input into a convolutional neural network to train the convolutional neural network.
  • the method includes:
  • the test merge vector is input into the convolutional neural network that has been trained to output the classification result of the test set, and it is verified whether it is consistent with the known classification of the test set.
  • the ratio of the sample data amount of the training set to the test set is 8: 2.
  • the method includes:
  • the animal database information is queried in the animal database.
  • the method includes:
  • a policy matching the kind name of the animal is queried in the policy database, and at least one of a policy amount and a compensation amount is obtained from the policy.
  • the animal species classification method, device, computer equipment, and storage medium provided in the embodiments of the present application, obtain animal images to be classified, and extract features of the animal images respectively through multiple different network models.
  • Vector stitching the feature vectors extracted by the multiple different network models into a merge vector; inputting the merge vector into the trained convolutional neural network to output the classification of the animal image to be classified.

Abstract

本申请中提供了一种动物种类的分类方法、装置、计算机设备和存储介质，包括：获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；将多个特征向量拼接为一个合并向量；将合并向量输入至训练完成的卷积神经网络中以输出待分类的动物图像的分类结果。本申请提升识别动物种类的准确率。

Description

动物种类的分类方法、装置、计算机设备和存储介质
本申请要求于2018年6月1日提交中国专利局、申请号为2018105572236，发明名称为“动物种类的分类方法、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
本申请涉及计算机技术领域，特别涉及一种动物种类的分类方法、装置、计算机设备和存储介质。
 背景技术
目前，常常用到对人脸进行识别，对狗等动物的种类进行识别分类却很少，通常为人工识别，人工识别准确性不高，且繁琐，造成人力成本太高。
狗的种类繁多，从斯坦福大学的狗种类数据集来看，狗的种类多达120种，其它动物的品种更加多种多样。在保险、科普等行业，识别动物的种类显得尤为重要。而目前，正确识别动物种类的准确率很低。
 技术问题
本申请的主要目的为提供一种动物种类的分类方法、装置、计算机设备和存储介质，克服识别动物种类准确率低的缺陷。
 技术解决方案
为实现上述目的，本申请提供了一种动物种类的分类方法，其包括以下步骤：
获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
本申请还提供了一种动物种类的分类装置，包括：
第一提取单元，用于获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
第一拼接单元，用于将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
第一分类单元，用于将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类动物图像的分类结果。
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
本申请中提供的动物种类的分类方法、装置、计算机设备和存储介质，获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；将多个不同的网络模型分别提取出的所述特征向量拼接为一个合并向量；将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果，提升识别动物种类的准确率，克服目前识别准确率低的缺陷。
 附图说明
图1 是本申请一实施例中动物种类的分类方法步骤示意图；
图2 是本申请另一实施例中动物种类的分类方法步骤示意图；
图3 是本申请一实施例中动物种类的分类装置结构示意图；
图4 是本申请另一实施例中动物种类的分类装置结构示意图；
图5 是本申请又一实施例中动物种类的分类装置结构示意图；
图6 是本申请又一实施例中动物种类的分类装置结构示意图；
图7 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，为本申请一实施例中提供的动物种类的分类方法步骤示意图，其包括以下步骤：
步骤S1，获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量。
在本步骤中，待分类的动物图像即为需要识别种类的动物照片，本实施例中的动物以狗为例进行阐述。本实施例中的网络模型均是已经训练完成的网络模型，其训练过程将在下文中进行阐述。其中，上述网络模型的最后一层通常为softmax分类层，其具有softmax分类器，经过该softmax层输出的结果为输入数据的分类结果，而在网络模型的最后一层（即softmax分类层）之前输出的是一个特征向量，可以理解的是，上述softmax分类层中输入的即是上述特征向量，将特征向量输入至softmax分类层则可以输出分类结果。因此，在本实施例中，获取到待分类的动物图像时，则通过多个不同的网络模型分别提取所述动物图像的特征向量，而并不直接通过softmax分类层输出分类结果。具体地，即通过多个不同的网络模型分别提取其softmax分类层之前的输出结果，便可以提取到上述特征向量。对于同一张动物图像，每一个网络模型均可以提取出一个特征向量，由于网络模型的结构不同，其提取出的特征向量也不同。
步骤S2，将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量。
在本步骤中，将上述多个不同的网络模型分别提取出的特征向量依次首尾拼接为一个合并向量，使得该合并向量中包括有上述所有特征向量的特征。上述合并向量也是一个向量，该拼接的合并向量作为卷积神经网络模型的网络输入。
步骤S3，将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
在本实施例中，上述卷积神经网络为训练完成的模型。将上述拼接得到的合并向量输入至训练完成的卷积神经网络中学习，该卷积神经网络的输出结果即为上述待分类的动物图像的分类结果。
由于上述每一个网络模型提取出的特征向量均不同，拼接之后得到的特征向量具有上述特征向量的不同特点。若将上述特征向量单一输入至卷积神经网络中，由于其特征向量不同，得出的分类结果也不尽相同。而本步骤中，将所述合并向量输入至卷积神经网络中进行计算，与单一特征向量输入时的计算原理一致，只是输入的向量不同。上述三个特征向量拼接后的合并向量是一个新的向量，输入至卷积神经网络中时，对该合并向量进行卷积计算，最后进行分类，分类结果中融合了不同特征向量的特点，得到的分类结果相比于单独输入上述特征向量时的识别分类更准确。因此，提升对动物种类的分类准确率，克服了目前识别准确率低的缺陷。
在一具体实施例中，狗的种类具有120种，为了针对狗的种类进行分类，将上述卷积神经网络中各个全连接层结构设置为：三个全连接层，其中第一个全连接层具有4096个节点，第二个全连接层具有512个节点，第三个全连接层具有120个节点，在第二个全连接层以及第三个全连接层之前分别加入了线性修正层（Relu，Rectified Linear Unit, 线性整流函数,又称线性修正单元）以及深度学习层（Dropout），以分别依次进行线性修正以及深度学习。
在一个实施例中，上述网络模型包括DenseNet（Densely Connected Convolutional Networks），ResNet（Deep Residual Network）以及Inception网络模型。通过上述三个网络模型分别提取出上述待分类的动物图像的特征向量，其分别对应为Fd、Fr、Fi；其中，上述的三个特征分别为一个具有1024个数字的向量。
在本实施例中，同一张狗图片输入至上述DenseNet，ResNet以及Inception三种不同的网络模型中，提取出的特征向量也不同，即向量中的1024个数字不同，若只输入至一个网络模型，则提取出的特征向量比较片面，信息不够全面，使用三个或多个网络模型提取时，特征向量表达狗的特征信息更加全面，更有利于后续狗种类的分类识别。
在本实施例中，同一张狗图片输入至DenseNet，ResNet以及Inception三种网络模型中，经过多层卷积层卷积计算，然后在倒数第二层（即softmax分类层的前一层）中提取出上述特征向量（最后一层输出的是分类结果，本实施例中不经过最后一层输出）。因此，提取倒数第二层输出的结果即提取出上述特征向量。
具体地，在上述DenseNet网络模型中，其前面所有卷积层的输出都连接至其后的所有卷积层，这种连接方式使得特征和梯度的传递更加有效，网络也就更加容易训练，相当于每一层都直接连接input（输入层）和loss（损失函数层），因此就可以减轻梯度消失现象。该网络模型具有有效解决梯度消失问题、强化特征传播、支持特征重用、大幅度减少参数数量的特点。在提取上述Fd特征时，后一个卷积层的输入都来自于其前面所有层的输出。
在上述ResNet网络模型中，ResNet网络中，通过shortcut将block的输入和输出进行一个element-wise的加叠，这个简单的加法并不会给网络增加额外的参数和计算量，同时却可以大大增加模型的训练速度、提高训练效果，并且当模型的层数加深时，这个简单的结构能够很好的解决退化问题。对每一层的输出做一个非线性变换，在提取特征Fr时，后一层的输入为其前一层的输出加上对其前一层的输出的非线性变换。其解决了退化的问题，在训练集和校验集上，都证明了的更深的网络错误率越小。
在上述Inception网络模型中，对网络中的传统卷积层进行了修改，用于增加网络深度和宽度，提高深度神经网络性能。具体地，将1x1，3x3，5x5的卷积核和3x3的maxpooling，堆叠在一起，一方面增加了网络的宽度，另一方面多个不同 size 的卷积核能够增强网络的适应力。为了降低特征图厚度，提取特征Fi时，在3x3卷积核前，5x5卷积核前，max pooling后分别加上了1x1的卷积核。
上述多个网络模型以及卷积神经网络均为预先训练完成的，下文则对其具体训练过程进行阐述。
参照图2，在一实施例中，上述获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤S1之前，包括：
步骤S101，将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型。
在本实施例中，预先将一个已知狗分类的狗数据集（大量的狗图片）分成训练集以及测试集，训练集用于训练上述网络模型以及卷积神经网络，测试集用于对训练完成的上述网络模型以及卷积神经网络进行测试，验证训练得到的模型是否准确。在一个实施例中，上述训练集以及测试集的样本数据量比例为8:2，即当狗的数据集中包括10000个样本图片时，其中8000个作为训练集，2000个作为测试集。
本实施例中，为了得到适用于本实施例中对狗种类进行分类的网络模型，则需要得到上述网络模型的训练参数。因此，需要将上述训练集中的每一张样本图片输入至上述多个网络模型中分别进行训练，以得到相应的训练参数。得出训练参数之后的网络模型则可以作为上述步骤S1中使用的网络模型。
具体地，将上述训练集输入至上述网络模型中训练，并经上述网络模型的最后一层输出分类结果，由于训练集中的狗的分类均是已知的，在训练上述网络模型时，只要使得其最后一层输出的分类结果逼近于/相同于训练集中的狗的分类，则可以完成上述网络模型的训练。该训练完成的网络模型即为上述步骤S1中使用的网络模型。
步骤S102，通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量。
在如上述步骤S101中对上述网络模型进行训练之后，本步骤S102中则通过训练完成的网络模型去提取训练集中每一张狗图片的特征向量，该步骤S102中的狗图片可以是上述步骤S101中的使用过的训练集中的狗图片，也可以是训练集中的其它狗图片。本步骤S102中提取训练集中狗图片的特征向量的过程与上述步骤S1中类似，不同之处在于针对的狗图片不同，在此不再进行赘述。
步骤S103，将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量。
本步骤中的拼接，即是将上述每一张样本图片对应的多个特征向量拼合成一个合并向量，拼接得到的上述训练合并向量作为卷积神经网络的训练数据输入。本步骤中的拼接过程与上述步骤S2中的拼接过程相同，在此不再进行赘述。
步骤S104，将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
在本步骤中，为了得到适用于本实施例中对狗种类进行分类的卷积神经网络，则需要得到上述卷积神经网络的训练参数。因此，需要将上述训练合并向量输入至上述卷积神经网络中进行训练，以得到相应的训练参数。得出训练参数之后的卷积神经网络则可以作为上述步骤S3中使用的卷积神经网络。
具体地，将上述训练合并向量输入至上述卷积神经网络中进行训练，并经上述卷积神经网络的最后一层输出分类结果，由于上述训练合并向量所对应训练集中的狗的分类是已知的，在训练上述网络模型时，只要使得卷积神经网络最后一层输出的分类结果逼近于/相同于训练集中的狗的分类，则可以完成上述卷积神经网络的训练。该训练完成的卷积神经网络即为上述步骤S3中使用的卷积神经网络。
在一实施例中，上述将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络的步骤S104之后，包括：
步骤S105，将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
步骤S106，将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
步骤S107，将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
在本实施例中，上述训练完成的网络模型以及卷积神经网络即为上述步骤S1、步骤S3中使用的网络模型以及卷积神经网络。为了验证上述训练过程中得到的网络模型以及卷积神经网络对动物种类的分类准确性，则需要对上述训练完成的网络模型以及卷积神经网络进行测试。本实施例中，使用测试集对上述训练完成的网络模型以及卷积神经网络进行测试，该测试集中图片数据与上述训练集中的图片数据均是已知动物种类的图片。本实施例中的步骤S105、步骤S106以及步骤S107的具体实现与上述步骤S1、步骤S2、步骤S3的实现过程类似，在此不再进行赘述。不同点仅在于针对的动物图像数据不同，且本实施例的步骤S107中需要比对卷积神经网络的输出的分类结果是否与测试集中图片的已知分类一致。
在另一实施例中，上述将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤S3之后，包括：
步骤S41，根据所述待分类动物图像的分类结果，确定所述待分类动物图像中动物的种类名称；
步骤S51，根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
在本实施例中，上述对动物种类的分类识别应用于科普领域中。具体地，在对一只未知种类的狗按照上述方法进行种类分类之后，即确定出未知图片中的狗的种类类型，从而如步骤S41根据种类类型确定该只狗的种类名称，如步骤S51在动物数据库中查询出所述动物的资料信息。其中，上述动物数据库中存储有大量的与动物相关的科普资料，例如生活习性、形体特征、主要聚居地等资料。本实施例中，通过机器自动对未知图片中的狗进行分类之后，自动从动物数据库中匹配出与之相关的资料信息，便于推送该资料信息至用户，用户则可以很方便的了解到未知图片中的狗的相关信息。
在又一实施例中，上述将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤S3之后，包括：
步骤S401，根据所述待分类动物图像的分类结果，确定所述待分类动物图像中动物的种类名称；
步骤S501，根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。
在本实施例中，上述对动物种类的分类识别应用于保险领域中。目前越来越多的用户对自己饲养的狗进行投保，不同的狗，其保单金额以及赔偿金额相应的也不相同。为了便于用户快速获取到上述保单金额以及赔偿金额等信息，本实施例中，使用上述方法识别狗的分类之后，如步骤S401，根据分类结果，确定狗的种类名称；再如步骤S501所述，根据狗的种类名称在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。上述保单数据库中存储有大量的为动物投保的保单，该保单中记载有与不同的狗相对应的保单金额、赔偿金额以及狗种类名称等信息。本实施例中，只需要根据狗种类名称，便可以快速获取到保单金额、赔偿金额等信息，便于用户投保。
在又一实施例中，上述将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤S3之后，包括：
步骤S4a，根据所述待分类动物图像的分类结果，确定所述待分类动物图像中动物的种类名称；
步骤S5a，根据所述动物的种类名称，在赔偿案例数据库中查询出与所述动物的种类名称相匹配的赔偿案例，并从所述赔偿案例中获取到赔偿金额。
在本实施例中，上述对动物种类的分类识别应用于保险领域中。狗在户外时，容易咬伤他人，被不同的狗咬时，其赔偿额度不同。用户想要获取到被不同的狗咬时其对应的赔偿金额；则可以上传狗的图片，并通过上述实施例中的分类方法对该狗进行分类，再如上述步骤S4a所述根据狗的分类结果，确定狗的种类名称；最后如上述步骤S5a所述，根据狗的种类名称，从赔偿案例数据库中查询出与所述动物的种类名称相匹配的赔偿案例，并从所述赔偿案例中获取到赔偿金额。
综上所述，为本申请实施例中提供的动物种类的分类方法，获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；将多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果，提升识别动物种类的准确率，克服目前识别准确率低的缺陷。
参照图3，本申请一实施例中还提供了一种动物种类的分类装置，包括：
第一提取单元10，用于获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
在本实施例中，待分类的动物图像即为需要识别种类的动物照片，本实施例中的动物以狗为例进行阐述。本实施例中的网络模型均是已经训练完成的网络模型，其训练过程将在下文中进行阐述。其中，上述网络模型的最后一层通常为softmax分类层，其具有softmax分类器，经过该softmax层输出的结果为输入数据的分类结果，而在网络模型的最后一层（即softmax分类层）之前输出的是一个特征向量，可以理解的是，上述softmax分类层中输入的即是上述特征向量，将特征向量输入至softmax分类层则可以输出分类结果。因此，在本实施例中，获取到待分类的动物图像时，第一提取单元10则通过多个不同的网络模型分别提取所述动物图像的特征向量，而并不直接通过softmax分类层输出分类结果。具体地，即第一提取单元10通过多个不同的网络模型分别提取其softmax分类层之前的输出结果，便可以提取到上述特征向量。对于同一张动物图像，每一个网络模型均可以提取出一个特征向量，由于网络模型的结构不同，其提取出的特征向量也不同。
第一拼接单元20，用于将多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
在本实施例中，第一拼接单元20将上述多个不同的网络模型分别提取的特征向量依次首尾拼接为一个合并向量，使得该合并向量中包括有上述所有特征向量的特征。上述合并向量也是一个向量，该拼接的合并向量作为卷积神经网络模型的网络输入。
第一分类单元30，用于将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
在本实施例中，上述卷积神经网络为训练完成的模型。第一分类单元30将上述拼接得到的合并向量输入训练完成的至卷积神经网络中学习，该卷积神经网络的输出结果即为上述待分类的动物图像的分类结果。
由于上述网络模型提取出的特征向量不同，拼接之后得到的特征向量具有上述特征向量的不同特点。若将上述特征向量单一输入至卷积神经网络中，由于其特征向量不同，得出的分类结果也不尽相同。而本步骤中，将所述合并向量输入至卷积神经网络中进行计算，与单一特征向量输入时的计算原理一致，只是输入的向量不同。上述三个特征向量拼接后的合并向量是一个新的向量，输入至卷积神经网络中时，第一分类单元30对该合并向量进行卷积计算，最后进行分类，分类结果中融合了不同特征向量的特点，得到的分类结果相比于单独输入上述特征向量时的识别分类更准确。因此，提升对动物种类的分类准确率，克服了目前识别准确率低的缺陷。
在一具体实施例中，狗的种类具有120种，为了针对狗的种类进行分类，将上述卷积神经网络中各个全连接层结构设置为：三个全连接层，其中第一个全连接层具有4096个节点，第二个全连接层具有512个节点，第三个全连接层具有120个节点，在第二个全连接层以及第三个全连接层之前分别加入了线性修正层（Relu，Rectified Linear Unit, 线性整流函数,又称线性修正单元）以及深度学习层（Dropout），以分别依次进行线性修正以及深度学习。
在又一具体实施例中，上述第一提取单元10中使用的网络模型包括DenseNet，ResNet以及Inception网络模型。
通过上述三个网络模型分别提取出上述待分类的动物图像的特征向量，其分别对应为Fd、Fr、Fi；其中，上述的三个特征分别为一个具有1024个数字的向量。
在本实施例中，同一张狗图片输入至上述DenseNet，ResNet以及Inception三种不同的网络模型中，提取出的特征向量也不同，即向量中的1024个数字不同，若只输入至一个网络模型，则提取出的特征向量比较片面，信息不够全面，使用三个或多个网络模型提取时，特征向量表达狗的特征信息更加全面，更有利于后续狗种类的分类识别。
在本实施例中，同一张狗图片输入至DenseNet，ResNet以及Inception三种网络模型中，经过多层卷积层卷积计算，然后第一提取单元10在倒数第二层（即softmax分类层的前一层）中提取出上述特征向量（最后一层输出的是分类结果，本实施例中不经过最后一层输出）。因此，提取倒数第二层输出的结果即提取出上述特征向量。
具体地，在上述DenseNet网络模型中，其前面所有卷积层的输出都连接至其后的所有卷积层，这种连接方式使得特征和梯度的传递更加有效，网络也就更加容易训练，相当于每一层都直接连接input（输入层）和loss（损失函数层），因此就可以减轻梯度消失现象。该网络模型具有有效解决梯度消失问题、强化特征传播、支持特征重用、大幅度减少参数数量的特点。在提取上述Fd特征时，后一个卷积层的输入都来自于其前面所有层的输出。
在上述ResNet网络模型中，ResNet网络中，通过shortcut将block的输入和输出进行一个element-wise的加叠，这个简单的加法并不会给网络增加额外的参数和计算量，同时却可以大大增加模型的训练速度、提高训练效果，并且当模型的层数加深时，这个简单的结构能够很好的解决退化问题。对每一层的输出做一个非线性变换，在提取特征Fr时，后一层的输入为其前一层的输出加上对其前一层的输出的非线性变换。其解决了退化的问题，在训练集和校验集上，都证明了的更深的网络错误率越小。
在上述Inception网络模型中，对网络中的传统卷积层进行了修改，用于增加网络深度和宽度，提高深度神经网络性能。具体地，将1x1，3x3，5x5的卷积核和3x3的maxpooling，堆叠在一起，一方面增加了网络的宽度，另一方面多个不同 size 的卷积核能够增强网络的适应力。为了降低特征图厚度，提取特征Fi时，在3x3卷积核前，5x5卷积核前，max pooling后分别加上了1x1的卷积核。
上述多个网络模型以及卷积神经网络均为预先训练完成的，下文则对其具体训练过程进行阐述。
参照图4，在一实施例中，上述动物种类的分类装置还包括：
第一训练单元101，用于将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型。
在本实施例中，预先将一个已知狗分类的狗数据集（大量的狗图片）分成训练集以及测试集，训练集用于训练上述网络模型以及卷积神经网络，测试集用于对训练完成的上述网络模型以及卷积神经网络进行测试，验证训练得到的模型是否准确。在一个实施例中，上述训练集以及测试集的样本数据量比例为8:2，即当狗的数据集中包括10000个样本图片时，其中8000个作为训练集，2000个作为测试集。
本实施例中，为了得到适用于本实施例中对狗种类进行分类的网络模型，则需要得到上述网络模型的训练参数。因此，第一训练单元101需要将上述训练集中的每一张样本图片输入至上述多个网络模型中分别进行训练，以得到相应的训练参数。得出训练参数之后的网络模型则可以作为上述第一提取单元10中使用的网络模型。
具体地，第一训练单元101将上述训练集输入至上述网络模型中训练，并经上述网络模型的最后一层输出分类结果，由于训练集中的狗的分类均是已知的，在训练上述网络模型时，只要使得其最后一层输出的分类结果逼近于/相同于训练集中的狗的分类，则可以完成上述网络模型的训练。该训练完成的网络模型即为上述第一提取单元10中使用的网络模型。
第二提取单元102，用于通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量。
在上述第一训练单元101对上述网络模型进行训练之后，第二提取单元102则通过训练完成的网络模型去提取训练集中每一张狗图片的特征向量，该过程中的狗图片可以是第一训练单元101中使用过的训练集中的狗图片，也可以是训练集中的其它狗图片。第二提取单元102提取训练集中狗图片的特征向量的过程与上述第一提取单元10的提取过程类似，不同之处在于针对的狗图片不同，在此不再进行赘述。
第二拼接单元103，用于将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量。
本步骤中的拼接，即是将上述每一张样本图片对应的多个特征向量拼合成一个合并向量，拼接得到的上述训练合并向量作为卷积神经网络的训练数据输入。第二拼接单元103的拼接过程与上述第一拼接单元20的拼接过程相同，在此不再进行赘述。
第二训练单元104，将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
在本实施例中，为了得到适用于本实施例中对狗种类进行分类的卷积神经网络，则需要得到上述卷积神经网络的训练参数。因此，第二训练单元104需要将上述训练合并向量输入至上述卷积神经网络中进行训练，以得到相应的训练参数。得出训练参数之后的卷积神经网络则可以作为上述第一分类单元30中使用的卷积神经网络。
具体地，第二训练单元104将上述训练合并向量输入至上述卷积神经网络中进行训练，并经上述卷积神经网络的最后一层输出分类结果，由于上述训练合并向量所对应训练集中的狗的分类是已知的，在训练上述网络模型时，只要使得卷积神经网络最后一层输出的分类结果逼近于/相同于训练集中的狗的分类，则可以完成上述卷积神经网络的训练。该训练完成的卷积神经网络即为上述第一分类单元30中使用的卷积神经网络。
在一实施例中，上述动物种类的分类装置还包括：
第三提取单元，用于将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
第三拼接单元，用于将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
验证单元，用于将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
在本实施例中，上述训练完成的网络模型以及卷积神经网络即为上述第一提取单元10、第一分类单元30中使用的网络模型以及卷积神经网络。为了验证上述训练过程中得到的网络模型以及卷积神经网络对动物种类的分类准确性，则需要对上述训练完成的网络模型以及卷积神经网络进行测试。本实施例中，使用测试集对上述训练完成的网络模型以及卷积神经网络进行测试，该测试集中图片数据与上述训练集中的图片数据均是已知动物种类的图片。本实施例中的第三提取单元、第三拼接单元以及验证单元的具体实现与上述第一提取单元10、第一拼接单元20、第一分类单元30的实现过程类似，在此不再进行赘述。不同点仅在于针对的动物图像数据不同，且本实施例的验证单元中需要比对卷积神经网络的输出的分类结果是否与测试集中图片的已知分类一致。
参照图5，在一实施例中，上述动物种类的分类装置还包括：
第一确定单元40，用于根据所述待分类的动物图像的分类结果，确定所述待分类动物图像中动物的种类名称；
第一查询单元50，用于根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
在本实施例中，上述对动物种类的分类识别应用于科普领域中。具体地，在对一只未知种类的狗按照上述分类装置进行种类分类之后，即确定出未知图片中的狗的种类类型，从而如上述第一确定单元40根据种类类型确定该只狗的种类名称，如第一查询单元50在动物数据库中查询出所述动物的资料信息。其中，上述动物数据库中存储有大量的与动物相关的科普资料，例如生活习性、形体特征、主要聚居地等资料。本实施例中，通过机器自动对未知图片中的狗进行分类之后，自动从动物数据库中匹配出与之相关的资料信息，便于推送该资料信息至用户，用户则可以很方便的了解到未知图片中的狗的相关信息。
参照图6，在另一实施例中，上述动物种类的分类装置还包括：
第二确定单元41，用于根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
第二查询单元51，用于根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。
在本实施例中，上述对动物种类的分类识别应用于保险领域中。目前越来越多的用户对自己饲养的狗进行投保，不同的狗，其保单金额以及赔偿金额相应的也不相同。为了便于用户快速获取到上述保单金额以及赔偿金额等信息，本实施例中，使用上述分类装置识别狗的分类之后，第二确定单元41根据分类结果，确定狗的种类名称；第二查询单元51再根据狗的种类名称在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。上述保单数据库中存储有大量的为动物投保的保单，该保单中记载有与不同的狗相对应的保单金额、赔偿金额以及狗种类名称等信息。本实施例中，只需要根据狗种类名称，便可以快速获取到保单金额、赔偿金额等信息，便于用户投保。
在另一实施例中，上述动物种类的分类装置还包括：
第三确定单元，用于根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
第三查询单元，用于根据所述动物的种类名称，在赔偿案例数据库中查询出与所述动物的种类名称相匹配的赔偿案例，并从所述赔偿案例中获取到赔偿金额。
在本实施例中，上述对动物种类的分类识别应用于保险领域中。狗在户外时，容易咬伤他人，被不同的狗咬时，其赔偿额度不同。用户想要获取到被不同的狗咬时其对应的赔偿金额；则可以上传狗的图片，并通过上述实施例中的分类方法对该狗进行分类，再如上述第三确定单元所述根据狗的分类结果，确定狗的种类名称；最后如第三查询单元所述，根据狗的种类名称，从赔偿案例数据库中查询出与所述动物的种类名称相匹配的赔偿案例，并从所述赔偿案例中获取到赔偿金额。
综上所述，为本申请实施例中提供的动物种类的分类装置，第一提取单元10获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；第一拼接单元20将多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；第一分类单元30将所述合并向量输入至卷积神经网络中以输出所述待分类的动物图像的分类结果，提升识别动物种类的准确率，克服目前识别准确率低的缺陷。
参照图7，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7所示。该计算机设备包括通过系统总线连接的处理器、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系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统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网络模型等数据。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一种动物种类的分类方法。
上述处理器执行上述动物种类的分类方法的步骤：
获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
将所述多个不同的网络模型分别提取的特征向量拼接为一个合并向量；
将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
在一实施例中，上述网络模型包括DenseNet，ResNet以及Inception网络模型。
在一实施例中，所述处理器获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤之前，包括：
将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
在一实施例中，所述处理器将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络的步骤之后，包括：
将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
在一实施例中，所述训练集与所述测试集的样本数据量比例为8:2。
在一实施例中，所述处理器将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
在一实施例中，所述处理器将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。
本领域技术人员可以理解，图7中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定。
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一种动物种类的分类方法，具体为：
获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
在一实施例中，上述网络模型包括DenseNet，ResNet以及Inception网络模型。
在一实施例中，所述处理器获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤之前，包括：
将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
在一实施例中，所述处理器将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络的步骤之后，包括：
将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
在一实施例中，所述训练集与所述测试集的样本数据量比例为8:2。
在一实施例中，所述处理器将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
在一实施例中，所述处理器将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单，并从所述保单中获取出保单金额、赔偿金额中至少一种。
综上所述，为本申请实施例中提供的动物种类的分类方法、装置、计算机设备和存储介质，获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果，提升识别动物种类的准确率，克服目前识别准确率低的缺陷。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其它变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素，而且还包括没有明确列出的其它要素，或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
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它相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种动物种类的分类方法，其特征在于，包括以下步骤：
    获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
    将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
    将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
  2. 根据权利要求1所述的动物种类的分类方法，其特征在于，所述网络模型包括DenseNet，ResNet和Inception网络模型。
  3. 根据权利要求1所述的动物种类的分类方法，其特征在于，所述获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤之前，包括：
    将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
    通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
    将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
    将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
  4. 根据权利要求3所述的动物种类的分类方法，其特征在于，所述将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络的步骤之后，包括：
    将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
    将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
    将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
  5. 根据权利要求4所述的动物种类的分类方法，其特征在于，所述训练集与所述测试集的样本数据量比例为8:2。
  6. 根据权利要求1所述的动物种类的分类方法，其特征在于，所述将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
    根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
    根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
  7. 根据权利要求1所述的动物种类的分类方法，其特征在于，所述将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果的步骤之后，包括：
    根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
    根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单。
  8. 一种动物种类的分类装置，其特征在于，包括：
    第一提取单元，用于获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
    第一拼接单元，用于将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
    第一分类单元，用于将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
  9. 根据权利要求8所述的动物种类的分类装置，其特征在于，所述网络模型包括DenseNet，ResNet和Inception网络模型。
  10. 根据权利要求8所述的动物种类的分类装置，其特征在于，还包括：
    第一训练单元，用于将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
    第二提取单元，用于通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
    第二拼接单元，用于将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
    第二训练单元，用于将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
  11. 根据权利要求10所述的动物种类的分类装置，其特征在于，还包括：
    第三提取单元，用于将测试集中每一张样本图片输入至多个训练完成的所述网络模型中以分别提取出测试集中每一张样本图片的特征向量；
    第三拼接单元，用于将所述测试集中每一张样本图片对应的多个特征向量拼接为一个测试合并向量；
    验证单元，用于将所述测试合并向量输入至训练完成的所述卷积神经网络中以输出所述测试集的分类结果，并验证其是否与测试集的已知分类一致。
  12. 根据权利要求11所述的动物种类的分类装置，其特征在于，所述训练集与所述测试集的样本数据量比例为8:2。
  13. 根据权利要求8所述的动物种类的分类装置，其特征在于，还包括：
    第一确定单元，用于根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
    第一查询单元，用于根据所述动物的种类名称，在动物数据库中查询出所述动物的资料信息。
  14. 根据权利要求8所述的动物种类的分类装置，其特征在于，还包括：
    第二确定单元，用于根据所述待分类的动物图像的分类结果，确定所述待分类的动物图像中动物的种类名称；
    第二查询单元，用于根据所述动物的种类名称，在保单数据库中查询出与所述动物的种类名称相匹配的保单。
  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动物种类的分类方法，所述方法包括：
    获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
    将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
    将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
  16. 根据权利要求15所述的计算机设备，其特征在于，所述网络模型包括DenseNet，ResNet和Inception网络模型。
  17. 根据权利要求15所述的计算机设备，其特征在于，所述处理器获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤之前，包括：
    将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
    通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
    将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
    将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
  18. 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动物种类的分类方法，所述方法包括：
    获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量；
    将所述多个不同的网络模型分别提取的所述特征向量拼接为一个合并向量；
    将所述合并向量输入至训练完成的卷积神经网络中以输出所述待分类的动物图像的分类结果。
  19. 根据权利要求18所述的计算机非易失性可读存储介质，其特征在于，所述网络模型包括DenseNet，ResNet和Inception网络模型。
  20. 根据权利要求18所述的计算机非易失性可读存储介质，其特征在于，所述处理器获取待分类的动物图像，并通过多个不同的网络模型分别提取所述动物图像的特征向量的步骤之前，包括：
    将训练集中每一张样本图片分别输入至多个不同的所述网络模型中分别训练各所述网络模型；
    通过训练完成的多个所述网络模型分别提取所述训练集中每一张样本图片的特征向量；
    将所述训练集中每一张样本图片对应的多个特征向量拼接为一个训练合并向量；
    将所述训练合并向量输入至卷积神经网络中以训练所述卷积神经网络。
PCT/CN2018/095668 2018-06-01 2018-07-13 动物种类的分类方法、装置、计算机设备和存储介质 WO2019227617A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810557223.6A CN108875811A (zh) 2018-06-01 2018-06-01 动物种类的分类方法、装置、计算机设备和存储介质
CN201810557223.6 2018-06-01

Publications (1)

Publication Number Publication Date
WO2019227617A1 true WO2019227617A1 (zh) 2019-12-05

Family

ID=64336360

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095668 WO2019227617A1 (zh) 2018-06-01 2018-07-13 动物种类的分类方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108875811A (zh)
WO (1) WO2019227617A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529020A (zh) * 2020-12-24 2021-03-19 携程旅游信息技术（上海）有限公司 基于神经网络的动物识别方法、系统、设备及存储介质
CN113139581A (zh) * 2021-03-23 2021-07-20 广东省科学院智能制造研究所 一种基于多图融合的图像分类方法和系统
CN114898155A (zh) * 2022-05-18 2022-08-12 平安科技（深圳）有限公司 车辆定损方法、装置、设备及存储介质

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109711443A (zh) * 2018-12-14 2019-05-03 平安城市建设科技(深圳)有限公司 基于神经网络的户型图识别方法、装置、设备及存储介质
CN109919005A (zh) * 2019-01-23 2019-06-21 平安科技（深圳）有限公司 牲畜身份识别方法、电子装置及可读存储介质
CN110097564B (zh) * 2019-04-04 2023-06-16 平安科技（深圳）有限公司 基于多模型融合的图像标注方法、装置、计算机设备及存储介质
CN110188595A (zh) * 2019-04-12 2019-08-30 淮阴工学院 一种基于ar和cnn算法的动物园游览系统及游览方法
CN109982051B (zh) * 2019-04-19 2021-06-29 东莞市南星电子有限公司 具有动物识别功能的监控摄像方法、监控相机和存储介质
CN110119694B (zh) * 2019-04-24 2021-03-12 北京百炼智能科技有限公司 一种图片处理方法、装置及计算机可读存储介质
CN110188613A (zh) * 2019-04-28 2019-08-30 上海鹰瞳医疗科技有限公司 图像分类方法及设备
CN110188820B (zh) * 2019-05-30 2023-04-18 中山大学 基于深度学习子网络特征提取的视网膜oct图像分类方法
CN111050315B (zh) * 2019-11-27 2021-04-13 北京邮电大学 一种基于多核双路网络的无线发射机识别方法
CN111178203B (zh) * 2019-12-20 2021-01-29 江苏常熟农村商业银行股份有限公司 签名审核方法、装置、计算机设备和存储介质
CN111046858B (zh) * 2020-03-18 2020-09-08 成都大熊猫繁育研究基地 一种基于图像的动物物种细分类方法、系统及介质
CN111753697A (zh) * 2020-06-17 2020-10-09 新疆爱华盈通信息技术有限公司 智能宠物管理系统及其管理方法
CN112215066A (zh) * 2020-09-08 2021-01-12 北京农业信息技术研究中心 家畜脸部图像识别方法及装置
CN112905793B (zh) * 2021-02-23 2023-06-20 山西同方知网数字出版技术有限公司 一种基于Bilstm+Attention文本分类的案例推荐方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170242879A1 (en) * 2007-07-26 2017-08-24 Hamid Hatami-Hanza Ontological Subjects Of A Universe And Knowledge Representations Thereof
CN107273899A (zh) * 2016-04-07 2017-10-20 富士通株式会社 对象分类方法和对象分类设备
CN107437096A (zh) * 2017-07-28 2017-12-05 北京大学 基于参数高效的深度残差网络模型的图像分类方法
CN108052894A (zh) * 2017-12-11 2018-05-18 北京飞搜科技有限公司 一种目标对象的多属性识别方法、设备、介质及神经网络

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103971097B (zh) * 2014-05-15 2015-05-13 武汉睿智视讯科技有限公司 一种基于多尺度笔画模型的车牌识别方法与系统
CN105469100B (zh) * 2015-11-30 2018-10-12 广东工业大学 基于深度学习的皮肤活检图像病理特性识别方法
CN106803090A (zh) * 2016-12-05 2017-06-06 中国银联股份有限公司 一种图像识别方法和装置
CN107292298B (zh) * 2017-08-09 2018-04-20 北方民族大学 基于卷积神经网络和分类器模型的牛脸识别方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170242879A1 (en) * 2007-07-26 2017-08-24 Hamid Hatami-Hanza Ontological Subjects Of A Universe And Knowledge Representations Thereof
CN107273899A (zh) * 2016-04-07 2017-10-20 富士通株式会社 对象分类方法和对象分类设备
CN107437096A (zh) * 2017-07-28 2017-12-05 北京大学 基于参数高效的深度残差网络模型的图像分类方法
CN108052894A (zh) * 2017-12-11 2018-05-18 北京飞搜科技有限公司 一种目标对象的多属性识别方法、设备、介质及神经网络

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529020A (zh) * 2020-12-24 2021-03-19 携程旅游信息技术（上海）有限公司 基于神经网络的动物识别方法、系统、设备及存储介质
CN113139581A (zh) * 2021-03-23 2021-07-20 广东省科学院智能制造研究所 一种基于多图融合的图像分类方法和系统
CN113139581B (zh) * 2021-03-23 2023-09-01 广东省科学院智能制造研究所 一种基于多图融合的图像分类方法和系统
CN114898155A (zh) * 2022-05-18 2022-08-12 平安科技（深圳）有限公司 车辆定损方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN108875811A (zh) 2018-11-23

Similar Documents

Publication Publication Date Title
WO2019227617A1 (zh) 动物种类的分类方法、装置、计算机设备和存储介质
TWI677852B (zh) 一種圖像特徵獲取方法及裝置、電子設備、電腦可讀存儲介質
WO2019137021A1 (zh) 一种机器学习模型训练方法和装置
WO2019200748A1 (zh) 迁移学习方法、装置、计算机设备和存储介质
WO2019218826A1 (zh) 图像处理方法、装置、计算机设备及存储介质
WO2019227616A1 (zh) 动物身份的识别方法、装置、计算机设备和存储介质
WO2020215571A1 (zh) 一种识别敏感数据的方法、装置、存储介质及计算机设备
WO2019119505A1 (zh) 人脸识别的方法和装置、计算机装置及存储介质
WO2019200781A1 (zh) 票据识别方法、装置及存储介质
WO2020088439A1 (zh) 实现异构图、分子空间结构性质识别的方法、装置和计算机设备
US9342735B2 (en) Facial recognition lost pet identifying system
CN108304435A (zh) 信息推荐方法、装置、计算机设备及存储介质
CN110046706B (zh) 模型生成方法、装置及服务器
Stylianou et al. Hotels-50k: A global hotel recognition dataset
Thompson et al. finFindR: Automated recognition and identification of marine mammal dorsal fins using residual convolutional neural networks
WO2022057309A1 (zh) 肺部特征识别方法、装置、计算机设备及存储介质
CN112862093A (zh) 一种图神经网络训练方法及装置
CN111242083A (zh) 基于人工智能的文本处理方法、装置、设备、介质
CN111325200A (zh) 图像标注方法、装置、设备及计算机可读存储介质
CN114329022A (zh) 一种色情分类模型的训练、图像检测方法及相关装置
CN105740903B (zh) 多属性识别方法及装置
CN116127080A (zh) 描述对象的属性值提取方法及相关设备
CN111597937B (zh) 鱼姿势识别方法、装置、设备及存储介质
CN110826475B (zh) 一种检测近重复视频的方法、装置及计算设备
Kin‐Chung Au et al. Electors voting for fast automatic shape correspondence

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: 18921144

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 25.03.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18921144

Country of ref document: EP

Kind code of ref document: A1