WO2020052183A1 - 商标侵权的识别方法、装置、计算机设备和存储介质 - Google Patents

商标侵权的识别方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2020052183A1
WO2020052183A1 PCT/CN2019/071363 CN2019071363W WO2020052183A1 WO 2020052183 A1 WO2020052183 A1 WO 2020052183A1 CN 2019071363 W CN2019071363 W CN 2019071363W WO 2020052183 A1 WO2020052183 A1 WO 2020052183A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
trademark
target
feature
trademark image
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2019/071363
Other languages
English (en)
French (fr)
Inventor
叶素兰
窦文伟
徐冰
聂宇昕
汪伟
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020052183A1 publication Critical patent/WO2020052183A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06Q50/10—Services
    • G06Q50/18—Legal services

Definitions

  • the present application relates to a method, device, computer equipment, and storage medium for identifying trademark infringement.
  • the commonly used method is to input the trademark image pair into two pre-trained feature extraction models to perform feature extraction to obtain the corresponding feature vector, and then calculate the Euclidean distance between the two feature vectors to obtain the corresponding The degree of similarity between the trademark image pairs to determine whether it is infringing.
  • a method, a device, a computer device, and a storage medium for identifying a trademark infringement are provided.
  • a method of identifying trademark infringement includes:
  • the target prediction label is a label of trademark infringement
  • determining the target trademark image as an infringing trademark image is a label of trademark infringement
  • a trademark infringement identification device includes:
  • a query module configured to query a pre-stored candidate trademark image according to the target trademark image
  • An extraction module for inputting the target trademark image and the candidate trademark image into a trained feature extraction model for prediction, obtaining a first image feature corresponding to the target trademark image, and corresponding to the candidate trademark image
  • the second image feature
  • a stitching module configured to stitch the first image feature and the second image feature to obtain a third image feature
  • a prediction module configured to input the third image feature into a trained infringement prediction model for prediction to obtain a target prediction label
  • a determining module configured to determine the target trademark image as an infringing trademark image when the target predicted label is a label of a trademark infringement.
  • a computer device includes a memory and one or more processors.
  • the memory stores computer-readable instructions, and the computer-readable instructions, when executed by the one or more processors, cause the one or more Processors implement the steps of the method for identifying a trademark infringement provided in any one of the embodiments of the present application.
  • One or more non-volatile computer-readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to implement any The steps of the method for identifying a trademark infringement provided in one embodiment.
  • FIG. 1 is an application scenario diagram of a method for identifying a trademark infringement according to one or more embodiments.
  • FIG. 2 is a schematic flowchart of a method for identifying a trademark infringement according to one or more embodiments.
  • FIG. 3 is a schematic flowchart of a method for identifying a trademark infringement in another embodiment.
  • FIG. 4 is a schematic flowchart of training steps of a feature extraction model according to one or more embodiments.
  • FIG. 5 is a block diagram of a device for identifying a trademark infringement according to one or more embodiments.
  • FIG. 6 is a block diagram of a device for identifying a trademark infringement in another embodiment.
  • FIG. 7 is a block diagram of a computer device according to one or more embodiments.
  • the method for identifying trademark infringement can be applied to the application environment shown in FIG. 1.
  • the terminal 102 communicates with the server 104 through the network through the network.
  • the server 104 queries a pre-stored candidate trademark image according to the obtained target trademark image, extracts a first image feature from the target trademark image through a trained feature extraction model, and extracts a second image feature from the candidate trademark image.
  • the image feature and the second image feature are stitched to obtain the corresponding third image feature, and the obtained third image feature is input into the trained infringement prediction model to perform prediction to obtain the target prediction label.
  • the target prediction label is a label representing a trademark infringement
  • the target image trademark is an infringing trademark image
  • the determination result is sent to the terminal 102.
  • the terminal 102 may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server 104 may be implemented by an independent server or a server cluster composed of multiple servers.
  • a method for identifying a trademark infringement is provided.
  • the method is applied to the server in FIG. 1 as an example, and includes the following steps:
  • the target trademark image refers to the trademark image to be determined whether it is infringing.
  • the target trademark image may be a newly registered trademark image.
  • a trademark image is an image containing a trademark, that is, an image with the trademark as an image content.
  • the trademark image may specifically be an image formed by the trademark itself, that is, the trademark image itself.
  • a trademark is a mark that distinguishes a brand or service.
  • the trademark may specifically include a graphic trademark and a text trademark.
  • a graphic trademark refers to a trademark expressed in a graphic form
  • a text trademark refers to a trademark expressed in a text form.
  • the server detects the newly registered trademark image in real time, and when the newly registered trademark image is detected, determines the detected trademark image as the target trademark image.
  • the server detects a specified trigger operation, and when a specified trigger operation is detected, obtains a corresponding target trademark image according to the detected specified trigger operation.
  • the designated trigger operation is a pre-designated operation for triggering the trademark infringement recognition process, such as a trigger operation of a preset trigger control.
  • the specified trigger operation may specifically be a trademark image entry operation or a trigger operation on a preset control for triggering a trademark infringement recognition process.
  • the server generates a corresponding query instruction according to the detected specified trigger operation, and obtains a corresponding target trademark image from a local or other computer device according to the generated query instruction.
  • Other computer equipment such as terminals or other servers for storing trademark images.
  • the server may also obtain the corresponding target trademark image from the designated webpage based on the network according to the query instruction.
  • the candidate trademark image is a trademark image that is stored in advance and can be compared with the target trademark image to determine whether the target trademark image is infringing.
  • the candidate trademark image may specifically be a registered trademark image, that is, an existing trademark image that has been pre-registered in a local database.
  • the pre-stored candidate trademark image is correspondingly queried from a local or other server for storing the trademark image according to the obtained target trademark image.
  • the server obtains the target trademark image, it correspondingly obtains one or more candidate trademark images stored in advance.
  • the server when the server obtains the target trademark image, the server searches the pre-stored trademark image locally or at another server for storing the trademark image, and determines the queried trademark image as corresponding to the target trademark image.
  • Candidate trademark image when the server obtains the target trademark image, the server searches the pre-stored trademark image locally or at another server for storing the trademark image, and determines the queried trademark image as corresponding to the target trademark image.
  • Candidate trademark image when the server obtains the target trademark image, the server searches the pre-stored trademark image locally or at another server for storing the trademark image, and determines the queried trademark image as corresponding to the target trademark image.
  • the server uses the preset filtering method according to the target trademark images to select from the multiple queryed trademark images. Filter out one or more candidate trademark images.
  • the preset filtering method may specifically be filtering according to the trademark type corresponding to the trademark in the trademark image, for example, from a plurality of queryed trademark images, the selected trademark type is related to the trademark type corresponding to the trademark in the target trademark image.
  • the types of marks include graphic marks and text marks.
  • the preset filtering method may also specifically be based on the key features of the trademark image, for example, a trademark image with the closest key feature is selected as a candidate trademark image. Key features such as the iconic features of a trademark image.
  • the target trademark image and the candidate trademark image are respectively input into a trained feature extraction model for prediction, and a first image feature corresponding to the target trademark image and a second image feature corresponding to the candidate trademark image are obtained.
  • the feature extraction model is a model obtained by performing model training according to a previously acquired training sample set, and can be used to extract image features from a trademark image.
  • Image characteristics refer to the characteristics of the image, and specifically may be the characteristics of the trademark in the trademark image.
  • the image feature may be a specified dimension feature vector used to characterize the features possessed by the image, and may specifically be a designated dimension feature vector used to characterize the features possessed by the trademark in the trademark image. Specify the number of dimensions, such as 4096.
  • the server inputs the target trademark image into the trained feature extraction model for prediction, obtains the corresponding first image feature, and inputs the candidate trademark image into the trained feature extraction model for prediction, to obtain the corresponding second image feature.
  • the feature extraction model used to extract the first image feature from the target trademark image may be the same model as the feature extraction model used to extract the second image feature from the candidate trademark image, or it may be a weight Shared different models.
  • the server can sequentially input the target trademark image and candidate trademark image into the trained feature extraction model for prediction, and obtain the first image feature corresponding to the target trademark image and the second image feature corresponding to the candidate trademark image.
  • the server may also input the target trademark image into the trained first feature extraction model for prediction, obtain the corresponding first image feature, and input the candidate trademark image into the trained second feature extraction model for prediction, and obtain the corresponding second image feature.
  • the first feature extraction model and the second feature extraction model may be feature extraction models with weight sharing, that is, they may be twin feature extraction models.
  • the feature extraction model is a model obtained by training based on a VGG network (Visual Geometry Group).
  • the feature extraction model may specifically be a model obtained by training based on the ImageNet data set and the VGG network.
  • ImageNet is a natural image dataset.
  • the server extracts the feature from the feature.
  • the feature extraction layer acquires a first image feature corresponding to the target trademark image.
  • Feature extraction layers such as the penultimate layer of the feature extraction model.
  • the server performs model training based on the VGG network to obtain a corresponding VGG network model.
  • the VGG network model is a VGG network that has been trained with various parameters, that is, a feature extraction model in each of the foregoing embodiments.
  • the server When the server enters the target trademark image into the VGG network model for prediction, the first image feature is extracted from the penultimate layer of the VGG network model. Similarly, the server extracts a second image feature from the candidate trademark image through a feature extraction model trained based on the VGG network.
  • the server inputs a target trademark image with a size of 224 * 224 into a feature extraction model trained based on the VGG network for prediction, and extracts a 4096-dimensional feature vector from the penultimate layer of the feature extraction model.
  • the 4096-dimensional feature The vector is the first image feature.
  • the server stitches the extracted first image feature and the corresponding second image feature according to a preset stitching method to obtain a corresponding third image feature.
  • the preset stitching method is a preset stitching method.
  • the stitching method may be stitching a low-dimensional image feature into a high-dimensional image feature, or a vector form image feature into a matrix form image feature.
  • the first image feature is a first feature vector having a specified dimension
  • the second image feature is a second feature vector having the same feature dimension as the first image.
  • the server stitches the first feature vector and the second feature vector according to a preset stitching method to obtain a corresponding matrix.
  • the preset stitching method may specifically align each element in the first feature vector with the corresponding element in the second feature vector, and stitch and combine the first feature vector and the second feature vector after the elements are aligned to obtain the first Three feature matrices.
  • the third feature matrix is the third image feature obtained by stitching.
  • the server aligns the first feature vector with the second feature in a manner that the corresponding elements of the feature vector are aligned respectively.
  • the vectors are stitched to obtain a 2 * 4096 matrix.
  • S210 Input the third image feature into a trained infringement prediction model for prediction, and obtain a target prediction label.
  • the infringement prediction model is a model obtained by training according to a previously acquired training sample set, and can be used to correspondingly determine an unknown target prediction label according to a known third image feature.
  • the target prediction label is a label obtained by a tort prediction model according to a known third image feature prediction.
  • the target predictive label is the basis for determining whether the target trademark image is infringing.
  • the target prediction label may specifically be a character or a character string composed of at least one of characters such as numbers, letters, and symbols.
  • the target prediction label may be a label indicating trademark infringement or a label indicating non-infringement of the trademark, for example, a label indicating 1 for trademark infringement and a label indicating 0 for non-infringement.
  • the target prediction label is 1, indicating that the target prediction label is a label indicating trademark infringement, that is, indicating that the target trademark image is highly similar to the candidate trademark image
  • the target trademark image is determined to be an infringing trademark image.
  • the target prediction label is 0, indicating that the target trademark image is irrelevant to the candidate trademark image, it is determined that the target trademark image is not an infringing trademark image.
  • the server inputs the obtained third image feature as an input feature into a trained infringement prediction model, performs prediction through the infringement prediction model, and obtains a corresponding target prediction label.
  • target prediction label is a label of trademark infringement
  • the server predicts to obtain the target prediction label, it is determined whether the target prediction label is a label representing a trademark infringement. When it is determined that the target prediction label is a label representing a trademark infringement, the server determines that the corresponding target trademark image is infringing, and determines the target trademark image as an infringing trademark image. When determining that the target prediction label is a label indicating that the trademark is not infringing, the server determines that the corresponding target trademark image is not infringing, and determines that the target trademark image is not an infringing trademark image, that is, a non-infringing trademark image.
  • the server when it is determined that the target trademark image is not an infringing trademark image, the server stores the target trademark image locally or at another server for storing the trademark image.
  • the server matches the obtained target predictive label with a preset label. When the match is successful, it indicates that the target predictive label is a label representing a trademark infringement, and the server determines the target trademark image as an infringing trademark image.
  • the default label is a label used to indicate trademark infringement, such as 1.
  • the server obtains the matching rate between the target predicted label and the preset label. When the obtained matching rate reaches the preset matching rate, the server determines the corresponding target trademark image as an infringing trademark image.
  • the foregoing trademark infringement recognition method queries a pre-stored candidate trademark image corresponding to the acquired target trademark image, and uses the trained feature extraction model and infringement prediction model to determine whether the target trademark image is based on the acquired candidate trademark image. Infringement has improved the efficiency and accuracy of infringement determination.
  • Use the feature extraction model to extract the first image feature from the target trademark image and the second image feature from the candidate trademark image, and then determine the corresponding input as the infringement prediction model according to the extracted first image feature and second image feature.
  • the third image feature of the feature to obtain the corresponding target prediction label through prediction by the infringement prediction model, which improves the efficiency and accuracy of the target prediction label acquisition, thereby improving the efficiency and accuracy of infringement determination.
  • the target prediction label matches the preset prediction label, the target trademark image is determined to be an infringing trademark image, which improves the accuracy of the trademark infringement determination result.
  • the feature extraction model includes a first feature extraction model and a second feature extraction model
  • the above-mentioned trademark infringement recognition method further includes: separately identifying the target trademark image with each candidate trademark The images are combined to obtain a plurality of trademark image pairs.
  • Step S206 includes: for each trademark image pair, inputting the target trademark image in the trademark image pair into a first feature extraction model, obtaining corresponding first image features, and combining the trademark image
  • the candidate trademark images in the input are predicted by a second feature extraction model to obtain corresponding second image features
  • step S212 includes: when a plurality of trademark images have corresponding trademark infringement prediction labels, the target is The trademark image was determined to be an infringing trademark image.
  • a trademark image pair refers to an image pair composed of two trademark images.
  • the trademark image pair may specifically be an image pair composed of a target trademark image and a candidate trademark image.
  • Both the first feature extraction model and the second feature extraction model are models obtained by performing model training according to a previously acquired training sample set, and can be used to extract corresponding image features from a trademark image.
  • the first feature extraction model and the second feature extraction model are feature extraction models with weight sharing, that is, twin feature extraction models.
  • a twin neural network framework is used in the training process of the first feature extraction model and the second feature extraction model, that is, model training is performed based on a neural network with two weights shared.
  • the server queries a plurality of pre-stored candidate trademark images correspondingly according to the target trademark image, and combines each of the queryed candidate trademark images with the target trademark image to obtain correspondingly the candidate trademark image and the target trademark image.
  • Multiple logo image pairs The number of trademark image pairs is the same as the number of query candidate trademark images, that is, the trademark image pairs correspond to the candidate trademark images.
  • the server extracts a corresponding first image feature from a target trademark image in the trademark image pair through a trained first feature extraction model, and passes the trained first feature
  • the two feature extraction model extracts the corresponding second image feature from the candidate trademark images in the trademark image pair, and stitches the first image feature with the second image feature to obtain the corresponding third image feature.
  • the infringement prediction model performs prediction according to the third image feature, and obtains the corresponding target prediction label.
  • the server performs the above-mentioned related steps of predicting a corresponding target prediction label based on the trademark image pair, respectively, to obtain a target prediction label corresponding to each trademark image pair.
  • the server obtains the target prediction labels corresponding to each of the plurality of trademark image pairs, it determines whether there is a label indicating a trademark infringement in the plurality of target prediction labels.
  • the server determines that the corresponding target trademark image is infringing, and determines the target trademark image as an infringing trademark image.
  • the server when the server obtains target prediction labels corresponding to each of the plurality of trademark image pairs, the server matches the plurality of target prediction labels with preset labels, respectively.
  • a matching result indicating successful matching exists in the matching result corresponding to each trademark image pair, it indicates that there is a label representing trademark infringement among the plurality of target prediction tags, and the server determines the target trademark image as an infringing trademark image.
  • the server determines that the target trademark image is not an infringing trademark image.
  • infringement recognition is performed on the target trademark image according to a plurality of pre-stored candidate trademark images.
  • the target trademark image is determined to be an infringing trademark image, which improves the trademark Identification accuracy of infringement.
  • step S206 is performed in accordance with the iterative execution in the specified order.
  • Step S212 includes: when the target predicted label corresponding to the currently predicted trademark image pair is a label of trademark infringement, stopping iteration to determine the target trademark image as an infringing trademark image.
  • the server queries a plurality of pre-stored candidate trademark images correspondingly according to the target trademark image, and combines each of the query candidate trademark images with the target trademark image to obtain the trademark image corresponding to each candidate trademark image correspondingly. Correct.
  • the plurality of trademark image pairs are composed of a target trademark image and a corresponding candidate trademark image.
  • the server prioritizes each trademark image pair in the plurality of trademark image pairs, iterates each trademark image pair in the plurality of trademark image pairs according to the priority order, and executes prediction of a corresponding target based on the trademark image pair. Related steps for predicting the label and determining whether the target trademark image is infringing according to the target predictive label.
  • the process of determining whether the target trademark image is infringing on the basis of the trademark image pair according to the priority order by the server specifically includes the following steps:
  • the server uses the trained feature extraction model from the target trademark image in the current priority trademark image pair. Extract the first image feature from the candidate image, and extract the second image feature from the candidate trademark image in the trademark image pair; the server combines the first image feature with the second image feature to obtain a third image feature, and passes the trained
  • the infringement prediction model makes predictions based on the characteristics of the third image to obtain corresponding target prediction tags; the server determines whether the target prediction tag is a tag indicating trademark infringement, and when it determines that it is a tag indicating trademark infringement, the server stops the iterative process, and Determining the corresponding target trademark image as an infringing trademark image; when it is determined that the label indicates that the trademark is not infringing, the server continues to perform the above-mentioned steps for determining whether the target trademark image is infringing based on the trademark image pair for
  • the server performs predictions according to the above method to obtain a target prediction label corresponding to the trademark image pair A.
  • the target prediction label is a trademark
  • the server stops iterating and determine the target trademark image as the infringing trademark image.
  • the target predictive label is a label indicating that the trademark is not infringing
  • the server continues to perform prediction to obtain the target predictive label corresponding to the trademark image pair B according to the above method.
  • the target image corresponding to the trademark image pair B predict whether the label is a label representing a trademark infringement, and continue to perform the corresponding steps.
  • the steps of determining whether the target trademark image is infringing according to the trademark image pair are sequentially and iteratively performed according to the obtained multiple trademark image pairs. When it is determined that the target trademark image is infringing, the iteration is stopped, improving the efficiency of trademark infringement And accuracy.
  • the feature extraction model training step includes: obtaining a preset trademark image; extracting a preset image feature from the preset trademark image according to a preset extraction method; The trademark image is used as the input feature, and the corresponding preset image feature is used as the desired output feature.
  • the initialized feature extraction model is trained to obtain the trained feature extraction model.
  • the server locally acquires a plurality of preset trademark images, and extracts corresponding preset image features from the plurality of preset trademark images respectively according to a preset extraction method.
  • the server obtains a corresponding training sample set according to the obtained preset trademark image and corresponding preset image features.
  • the preset brand image in the training sample set is used as the input feature, and the corresponding preset image feature is used as the desired output feature.
  • the server performs model training on the initialized feature extraction model according to the acquired training sample set to obtain a trained feature extraction model.
  • the server may also obtain the preset trademark image from another server for storing the trademark image, or obtain the preset trademark image from a designated webpage based on the network.
  • the server obtains a natural image data set used as a training sample set, where the natural image data set contains a large number of labeled natural images, and the server performs model training according to the acquired training sample set to obtain corresponding feature extraction model.
  • the feature extraction model trained by this training method is equivalent to a classifier, and the output features in the prediction process are categories corresponding to the input image as input features, that is, the category of the input image can be predicted by the feature extraction model.
  • the server inputs the target trademark image and the candidate trademark image into the trained feature extraction model, and obtains the target trademark image and the candidate trademark through the feature extraction layer of the feature extraction model.
  • the natural image data set may specifically be an ImageNet data set, which contains more than 14 million pictures, covering more than 20,000 categories.
  • model training is performed according to the obtained training sample set, so as to extract corresponding image features from the trademark image based on the feature extraction model obtained from the training, and then determine whether the target trademark image is infringing according to the image feature correspondence, thereby improving the trademark. Identification efficiency and accuracy of infringement.
  • the training steps of the infringement prediction model include: obtaining a training sample set; the training sample set includes a target trademark image pair and a corresponding target label; and according to the target image pair and the trained Feature extraction model to obtain corresponding target image features; using target image features as input features, and corresponding target labels as desired output features, perform model training on the convolutional neural network; stop training when preset conditions are reached, and obtain the Trained infringement prediction models.
  • the target trademark image pair is a trademark image pair used for model training.
  • the target trademark image pair includes a trademark image pair composed of two trademark images with high similarity, or a trademark image pair composed of two unrelated trademark images.
  • the target trademark image pair includes a trademark image pair composed of two mutually infringing trademark images, or a trademark image pair consisting of two non-infringing trademark images.
  • the target label is a label corresponding to the trademark image pair and used to indicate whether the two trademark images constituting the trademark image pair infringe each other. For example, when the target label is 1, it indicates that the two trademark images in the corresponding trademark image pair infringe each other. When the target label is 0, it means that the two trademark images in the corresponding trademark image pair do not infringe each other.
  • the preset condition is a condition for determining whether to stop the current model training process.
  • the preset condition may specifically be a loss function convergence, a loss function value tending to be stable, or a weight parameter tending to be stable.
  • the server locally acquires multiple target trademark image pairs, determines a target label corresponding to each target trademark image pair in the multiple target trademark image pairs, and according to the obtained target trademark image pairs and corresponding
  • the target label obtains the training sample set.
  • the server uses the trained feature extraction model to extract the corresponding image features from each of the trademark images that make up the target trademark image pair, and extracts the two extracted images.
  • the features are stitched to obtain the corresponding target image features.
  • the server separately determines the corresponding target image features in the above manner.
  • the server performs model training on the convolutional neural network by using the target image features corresponding to the multiple target trademark images as input features and the corresponding target labels as desired output features.
  • the server stops training to obtain a trained infringement prediction model.
  • the server may also obtain a preset trademark image from another server for storing a trademark image, or obtain a preset trademark image from a designated webpage based on a network.
  • the server obtains a plurality of labeled target trademark image pairs used as a training sample set, and performs model training according to the training sample set and the trained feature extraction model to obtain a trained infringement prediction model.
  • the training sample set includes target trademark image pairs and corresponding target labels.
  • the server inputs one trademark image in the target trademark image pair, inputs the trained first feature extraction model to obtain the first image feature, and inputs another trademark image in the target trademark image pair into the trained second feature. Extract the model to obtain a second image feature, and stitch the first image feature and the second image feature to obtain a third image feature.
  • the server uses the third image feature as the input feature and the corresponding target label as the desired output feature, and performs model training on the convolutional neural network to obtain a tort prediction model.
  • the first feature extraction model and the second feature extraction model share weights.
  • the convolutional neural network includes at least one convolutional layer, a batch normalization layer, a maximum pooling layer, a fully connected layer, and a softmax layer.
  • Dropout can be performed for the neural network of each layer, that is, for each neural network unit that composes the neural network of each layer, it is temporarily discarded from the neural network according to a specified probability. Thereby obtaining a thinner neural network.
  • model training is performed according to the obtained training sample set, and the infringement prediction model obtained through training is correspondingly determined whether the target trademark image is infringing, which improves the recognition efficiency and accuracy of the trademark infringement.
  • the target image feature is used as the input feature
  • the corresponding target label is used as the desired output feature to perform model training on the convolutional neural network, including: inputting the target image feature as the input feature to the convolutional neural network for Predict and obtain corresponding prediction labels; calculate the comparison loss value between the prediction label and the corresponding target label as the desired output feature according to the preset calculation method; train the convolutional neural network weight according to the preset method based on the comparison loss value Parameters; stopping training when a preset condition is reached to obtain a trained infringement prediction model, including: stopping training when a weight parameter meets a preset condition to obtain a trained infringement prediction model.
  • the preset calculation method is a preset method for calculating a comparison loss value between the predicted label and the target label.
  • the predetermined calculation method is, for example, calculation based on a contrast loss function.
  • the contrast loss value is the difference between the predicted label and the corresponding target label.
  • the comparison loss value may specifically be a difference between the predicted label and the corresponding target label calculated based on the corresponding comparison loss function.
  • the preset method refers to a preset method for training weight parameters of a convolutional neural network according to a contrast loss value.
  • the preset method may specifically correspond to the weight parameters of the training convolutional neural network through a back propagation algorithm.
  • each of the target image features obtained as input features is sequentially input to the convolutional neural network for processing. Prediction, to obtain prediction labels corresponding to each target image feature.
  • the server For each target image feature in the plurality of target image features, the server sequentially calculates a contrast loss value between the predicted label and the target label corresponding to each target image feature based on the contrast loss function.
  • a target label is a desired output feature corresponding to a target image as an input feature.
  • the server sequentially trains the weight parameters of the convolutional neural network through the back propagation algorithm according to the calculated contrast loss value. When the weight parameters tend to be stable, that is, when the weight parameters converge, the server stops training to obtain a trained infringement prediction model.
  • the server iteratively executes to obtain a corresponding prediction label according to the target image feature prediction, and then calculates a corresponding contrast loss value, and according to the contrast loss value Relevant steps for training the weight parameters of the convolutional neural network, and stop the iteration when the weight parameters meet the preset conditions to obtain the trained infringement prediction parameters.
  • the current iteration process is completed, if the weight parameter does not meet the preset condition, the above iterative process is continued for another target image feature until the weight parameter meets the preset condition, or for each target in the multiple target image features When all the image features have performed the above iterative process, the iteration is stopped.
  • the iteration is stopped to obtain a trained infringement prediction model.
  • the weight parameters of the training convolutional neural network are correspondingly calculated according to the comparison loss value between the prediction label and the corresponding target label to obtain a trained infringement prediction model, which improves the model training efficiency and accuracy.
  • a method for identifying a trademark infringement is provided.
  • the method specifically includes the following steps:
  • S304 Query a pre-stored candidate trademark image according to the target trademark image; there are multiple candidate trademark images.
  • S310 Merge the first image feature and the second image feature to obtain a third image feature.
  • the third image feature is input into a trained infringement prediction model for prediction, and a target prediction label is obtained.
  • the target trademark image is combined with multiple candidate trademark images, and based on the multiple trademark image pairs obtained through the combination, the target trademark image is subjected to infringement recognition with the help of a feature extraction model and an infringement prediction model, thereby improving recognition. Efficiency and accuracy.
  • a training step of an infringement prediction model in a method for identifying a trademark infringement is provided, which specifically includes:
  • the training sample set includes a target trademark image pair and a corresponding target label.
  • the target image features are input as input features to the convolutional neural network for prediction, and corresponding prediction labels are obtained.
  • a training step of an infringement prediction model is provided.
  • the infringement prediction model trained based on the training step can be used to identify trademark infringement, which improves the recognition efficiency and accuracy.
  • steps in the flowcharts of FIGS. 2-4 are sequentially displayed in accordance with the directions of the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless explicitly stated herein, the steps are performed in a non-strict order, and the steps may be performed in other orders. Moreover, at least a part of the steps in Figure 2-4 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily performed at the same time, but may be performed at different times. These sub-steps or stages The execution order of is not necessarily performed sequentially, but may be performed in turn or alternately with at least a part of another step or a sub-step or stage of another step.
  • a trademark infringement identification device 500 which includes: an acquisition module 502, an inquiry module 504, an extraction module 506, a stitching module 508, a prediction module 510, and a determination module 512. among them:
  • the obtaining module 502 is configured to obtain a target trademark image.
  • the query module 504 is configured to query a pre-stored candidate trademark image according to the target trademark image.
  • An extraction module 506 is configured to input the target trademark image and the candidate trademark image into a trained feature extraction model for prediction, and obtain a first image feature corresponding to the target trademark image and a second image feature corresponding to the candidate trademark image.
  • a stitching module 508 is configured to stitch the first image feature and the second image feature to obtain a third image feature.
  • a prediction module 510 is configured to input a third image feature into a trained infringement prediction model for prediction, and obtain a target prediction label.
  • a determination module 512 is configured to determine the target trademark image as an infringing trademark image when the target predicted label is a label of a trademark infringement.
  • the feature extraction model includes a first feature extraction model and a second feature extraction model; an extraction module 506 is further configured to combine the target trademark image with each candidate trademark image separately To obtain multiple trademark image pairs; for each trademark image pair, input the target trademark image in the trademark image pair into the first feature extraction model, obtain the corresponding first image feature, and input the candidate trademark image in the trademark image pair
  • the second feature extraction model performs prediction to obtain corresponding second image features;
  • the determination module 512 is further configured to determine the target trademark image as a trademark infringement tag when a plurality of trademark image pairs have corresponding target prediction tags. Infringing trademark image.
  • the extraction module 506 is further configured to combine the target trademark image with each candidate trademark image separately to obtain multiple trademark image pairs; for multiple trademark image pairs, according to The execution of the specified sequence iteration is a step of inputting the target trademark image and the candidate trademark image into the trained feature extraction model respectively for prediction, obtaining a first image feature corresponding to the target trademark image and a second image feature corresponding to the candidate trademark image;
  • the determination module 512 is further configured to stop iteration and determine the target trademark image as an infringing trademark image when the target predicted label corresponding to the currently predicted trademark image pair is a trademark infringement label.
  • the device 500 for identifying trademark infringement further includes: a model training module 514;
  • a model training module 514 configured to obtain a preset trademark image; extract a preset image feature from the preset trademark image according to a preset extraction method; use the preset trademark image as an input feature, and use the corresponding preset image feature as a desired output Feature, and training the initialized feature extraction model to obtain a trained feature extraction model.
  • the model training module 514 is further configured to obtain a training sample set; the training sample set includes a target trademark image pair and a corresponding target label; and a corresponding target image is obtained according to the target image pair and the trained feature extraction model Feature; using the target image feature as the input feature and the corresponding target label as the desired output feature, perform model training on the convolutional neural network; stop training when the preset conditions are reached to obtain a trained infringement prediction model.
  • the model training module 514 is further configured to input the target image feature as an input feature to the convolutional neural network for prediction and obtain the corresponding prediction label; calculate the prediction label and the corresponding, as expected, according to a preset calculation method. Contrast loss values between the target labels of the output features; weight parameters of the convolutional neural network are trained in a preset manner according to the contrast loss values; training is stopped when the weight parameters meet the preset conditions to obtain a trained infringement prediction model.
  • Each module in the above-mentioned trademark infringement identification device may be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the hardware form or independent of the processor in the computer device, or may be stored in the memory of the computer device in the form of software, so that the processor calls and performs the operations corresponding to the above modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 7.
  • the computer device includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the processor of the computer device 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 operating the operating system and computer-readable instructions in a non-volatile storage medium.
  • the computer equipment database is used to store candidate trademark images.
  • 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 for identifying a trademark infringement.
  • FIG. 7 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied. Include more or fewer parts than shown in the figure, or combine certain parts, or have a different arrangement of parts.
  • a computer device includes a memory and one or more processors.
  • Computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by one or more processors, the one or more processors implement any one of the present application. The steps of the method for identifying a trademark infringement provided in the embodiment.
  • One or more non-volatile computer-readable storage media storing computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors implement one of the embodiments of the present application Provide steps for identifying trademark infringement methods.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
  • SRAM static RAM
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM dual data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain Synchlink DRAM
  • Rambus direct RAM
  • DRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • General Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Technology Law (AREA)
  • Marketing (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种商标侵权的识别方法包括：获取目标商标图像；根据所述目标商标图像查询预存储的候选商标图像；将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。

Description

商标侵权的识别方法、装置、计算机设备和存储介质
本申请要求于2018年09月10日提交中国专利局，申请号为2018110516166，申请名称为“商标侵权的识别方法、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
本申请涉及一种商标侵权的识别方法、装置、计算机设备和存储介质。
背景技术
随着人工智能技术的发展，基于机器学习的智能识别方法逐渐发展起来，提高了识别效率和准确性，有效解决了人工识别所存在的问题。基于机器学习的智能识别应用于各行各业，比如商标侵权的识别。商标侵权是指未经商标权人许可，在相同或类似商品上使用与其注册商标相同或相近的商标。商标侵权会给商标权人造成经济损失，并带来不良后果，因而如何有效判断商标是否侵权是值得关注的问题。以图形商标为例，常用的方法是将商标图像对分别输入预先训练好的两个特征提取模型进行特征提取，获得相应的特征向量，通过计算两个特征向量之间的欧式距离即可获得相应商标图像对之间的相似程度，从而判定是否侵权。
然而，发明人意识到，目前的商标侵权判定方法中，需要获取大量已标注的商标图像作为训练样本进行模型训练，但是现有的已标注商标图像数量较少，因而导致模型训练不充分，降低了模型预测准确性，从而降低了商标侵权判定结果的准确性。
发明内容
根据本申请公开的各种实施例，提供一种商标侵权的识别方法、装置、计算机设备和存储介质。
一种商标侵权的识别方法包括：
获取目标商标图像；
根据所述目标商标图像查询预存储的候选商标图像；
将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
一种商标侵权的识别装置包括：
获取模块，用于获取目标商标图像；
查询模块，用于根据所述目标商标图像查询预存储的候选商标图像；
提取模块，用于将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
拼接模块，用于将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
预测模块，用于将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
判定模块，用于当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
一种计算机设备，包括存储器和一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述一个或多个处理器执行时，使得所述一个或多个处理器实现本申请任意一个实施例中提供的商标侵权的识别方法的步骤。
一个或多个存储有计算机可读指令的非易失性计算机可读存储介质，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器实现本申请任意一个实施例中提供的商标侵权的识别方法的步骤。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1为根据一个或多个实施例中商标侵权的识别方法的应用场景图。
图2为根据一个或多个实施例中商标侵权的识别方法的流程示意图。
图3为另一个实施例中商标侵权的识别方法的流程示意图。
图4为根据一个或多个实施例中特征提取模型的训练步骤的流程示意图。
图5为根据一个或多个实施例中商标侵权的识别装置的框图。
图6为另一个实施例中商标侵权的识别装置的框图。
图7为根据一个或多个实施例中计算机设备的框图。
具体实施方式
为了使本申请的技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请提供的商标侵权的识别方法，可以应用于如图1所示的应用环境中。终端102通过网络与服务器104通过网络进行通信。服务器104根据所获取到的目标商标图像查询预存储的候选商标图像，通过已训练的特征提取模型从目标商标图像中提取第一图像特征，从候选商标图像中提取第二图像特征，对第一图像特征和第二图像特征进行拼接获得相应的第三图像特征，并将所获取的第三图像特征输入已训练的侵权预测模型进行预测获得目标预测标签，当目标预测标签为表示商标侵权的标签时，则判定目标图像商标为侵权商标图像，并将判定结果发送至终端102。终端102可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备，服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在其中一个实施例中，如图2所示，提供了一种商标侵权的识别方法，以该方法应用于图1中的服务器为例进行说明，包括以下步骤：
S202，获取目标商标图像。
目标商标图像是指待判定是否侵权的商标图像。目标商标图像具体可以是新录入的商标图像。商标图像是包含商标的图像，即以商标为图像内容的图像。商标图像具体可以是商标本身构成的图像，即商标图像本身。商标是用于区别品牌或服务的标记。商标具体可以包括图形商标和文字商标，图形商标是指以图形的形式来表示的商标，文字商标是指以文字的形式来表示的商标。
具体地，服务器实时检测新录入的商标图像，当检测到新录入的商标图像时，将所检测到的商标图像确定为目标商标图像。
在其中一个实施例中，服务器检测指定触发操作，当检测到指定触发操作时，根据所检测到的指定触发操作获取相应的目标商标图像。指定触发操作是预先指定的用于触发商标侵权识别流程的操作，比如预设触发控件的触发操作。在本实施例中，指定触发操作具体可以是商标图像的录入操作，或者对用于触发商标侵权识别流程的预设控件的触发操作。
在其中一个实施例中，服务器根据所检测到的指定触发操作生成相应的查询指令，根据所生成的查询指令从本地或其他计算机设备获取相应的目标商标图像。其他计算机设备比如终端或其他用于存储商标图像的服务器。服务器根据查询指令也可以基于网络从指定网页获取相应的目标商标图像。
S204，根据目标商标图像查询预存储的候选商标图像。
候选商标图像是预先存储的、且能够用于与目标商标图像进行比较，以判定目标商标图像是否侵权的商标图像。候选商标图像具体可以是已注册的商标图像，即已预先录入本地数据库的已有商标图像。
具体地，服务器获取到目标商标图像时，根据所获取到的目标商标图像从本地或其他用于存储商标图像的服务器，对应查询预存储的候选商标图像。服务器获取到目标商标图像时，对应获取预存储的一个或多个候选商标图像。
在其中一个实施例中，服务器获取到目标商标图像时，在本地或其他用于存储商标图像的服务器查询预存储的商标图像，将所查询到的商标图像确定为对应于目标商标图像所查询到的候选商标图像。
在其中一个实施例中，服务器在本地或其他用于存储商标图像的服务器查询到预存储的多个商标图像时，根据目标商标图像按照预设筛选方式，从所查询到的多个商标图像中筛选出一个或多个候选商标图像。预设筛选方式具体可以是按照商标图像中的商标所对应的商标类型进行筛选，比如从所查询到的多个商标图像中，筛选出商标类型与目标商标图像中的商标所对应的商标类型相匹配的商标所对应的商标图像。商标类型包括图形商标和文字商标。预设筛选方式具体还可以是按照商标图像的关键特征进行筛选，比如筛选出关键特征最接近的商标图像作为候选商标图像。关键特征比如商标图像的标志性特征。
S206，将目标商标图像和候选商标图像分别输入已训练的特征提取模型进行预测，获得与目标商标图像对应的第一图像特征，以及与候选商标图像对应的第二图像特征。
特征提取模型是根据预先获取的训练样本集进行模型训练获得的、能够用于从商标图像中提取图像特征的模型。图像特征是指图像所具有的特征，具体可以是商标图像中的商标所具有的特征。在本实施例中，图像特征可以是用于表征图像所具有的特征的指定维数特征向量，具体可以是用于表征商标图像中的商标所具有的特征的指定维数特征向量。指定维数比如4096。
具体地，服务器将目标商标图像输入已训练的特征提取模型进行预测，获得相应的第一图像特征，并将候选商标图像输入已训练的特征提取模型进行预测，获得相应的第二图像特征。
在其中一个实施例中，用于从目标商标图像中提取第一图像特征的特征提取模型，与用于从候选商标图像中提取第二图像特征的特征提取模型可以是相同模型，也可以是权重共享的不同模型。换而言之，服务器可以将目标商标图像和候选商标图像依次输入已训练的特征提取模型进行预测，分别获得与目标商标图像对应的第一图像特征，以及与候选商标图像对应的第二图像特征。服务器也可以将目标商标图像输入已训练的第一特征提取模型进行预测，获得相应的第一图像特征，将候选商标图像输入已训练的第二特征提取模型进行预测，获得相应的第二图像特征。第一特征提取模型与第二特征提取模型可以是权重共享的特征提取模型，即可以是孪生特征提取模型。
在其中一个实施例中，特征提取模型是基于VGG网络(Visual Geometry Group，深度神经网络)进行训练获得的模型。特征提取模型具体可以是基于ImageNet数据集和VGG网络进行训练获得的模型。ImageNet是自然图像数据集。
在其中一个实施例中，通过基于VGG网络训练获得的特征提取模型从目标商标图像 中提取第一图像特征的过程中，服务器将目标商标图像输入该特征提取模型后，从该特征提取模型中的特征提取层获取与目标商标图像对应的第一图像特征。特征提取层比如特征提取模型的倒数第三层。换而言之，服务器基于VGG网络进行模型训练获得相应的VGG网络模型，该VGG网络模型即为各个参数已训练完成的VGG网络，也即为上述各个实施例中的特征提取模型。服务器将目标商标图像输入该VGG网络模型进行预测时，从该VGG网络模型的倒数第三层提取第一图像特征。类似地，服务器通过基于VGG网络训练获得的特征提取模型从候选商标图像中提取第二图像特征。
举例说明，服务器将尺寸大小为224*224的目标商标图像输入基于VGG网络训练获得的特征提取模型进行预测，从该特征提取模型的倒数第三层提取4096维的特征向量，该4096维的特征向量即为第一图像特征。
S208，将第一图像特征和第二图像特征进行拼接获得第三图像特征。
具体地，服务器将所提取到的第一图像特征和相应的第二图像特征，按照预设拼接方式进行拼接，获得相应的第三图像特征。预设拼接方式是预先设定的拼接方式，具体可以是将低维图像特征拼接成高维图像特征，或者将向量形式的图像特征拼接成矩阵形式的图像特征。
在其中一个实施例中，第一图像特征为指定维数的第一特征向量，第二图像特征为与第一图像特征维数相同的第二特征向量。服务器按照预设拼接方式将第一特征向量与第二特征向量进行拼接，获得相应的矩阵。预设拼接方式具体可以是将第一特征向量中的各个元素，分别与第二特征向量中的相应元素进行对齐，将元素对齐后的第一特征向量与第二特征向量进行拼接组合，获得第三特征矩阵。还第三特征矩阵即为拼接获得的第三图像特征。
举例说明，假设第一图像特征为4096维的第一特征向量，第二图像特征为4096维的第二特征向量，服务器按照特征向量对应元素分别对齐的方式，将第一特征向量与第二特征向量进行拼接获得2*4096的矩阵。
S210，将第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签。
侵权预测模型是根据预先获取的训练样本集进行训练获得的、能够用于根据已知的第三图像特征对应确定未知的目标预测标签的模型。目标预测标签是通过侵权预测模型根据已知的第三图像特征预测获得的标签。目标预测标签是用于判定目标商标图像是否侵权的判定依据。目标预测标签具体可以是有数字、字母和符号等字符中的至少一种组成的字符或字符串。目标预测标签可以是表示商标侵权的标签，也可以是表示商标不侵权的标签，比如用1表示商标侵权的标签，用0表示商标不侵权的标签。
举例说明，当目标预测标签为1时，表示该目标预测标签为表示商标侵权的标签，即表示目标商标图像与候选商标图像相似度很高，则判定目标商标图像为侵权商标图像。当目标预测标签为0时，表示目标商标图像与候选商标图像是无关的，则判定目标商标图像不是侵权商标图像。
具体地，服务器将所获得的第三图像特征作为输入特征输入已训练好的侵权预测模型，通过该侵权预测模型进行预测，获得相应的目标预测标签。
S212，当目标预测标签为商标侵权的标签时，将目标商标图像确定为侵权商标图像。
具体地，服务器预测获得目标预测标签时，判断该目标预测标签是否为表示商标侵权的标签。当判定该目标预测标签为表示商标侵权的标签时，服务器则判定相应的目标商标图像侵权，将该目标商标图像确定为侵权商标图像。当判定该目标预测标签为表示商标不侵权的标签时，服务器则判定相应的目标商标图像不侵权，则确定该目标商标图像不是侵权商标图像，即非侵权商标图像。
在其中一个实施例中，当判定目标商标图像不是侵权商标图像时，服务器将该目标商标图像存储在本地或其他用于存储商标图像的服务器。
在其中一个实施例中，服务器将所获得的目标预测标签与预设标签进行匹配，当匹配成功时，表明目标预测标签为表示商标侵权的标签，服务器将该目标商标图像确定为侵权商标图像。预设标签是用于表示商标侵权的标签，比如1。在其中一个实施例中，服务器获取目标预测标签与预设标签之间的匹配率，当所获取的匹配率达到预设匹配率时，服务器将相应目标商标图像确定为侵权商标图像。
上述商标侵权的识别方法，根据所获取到的目标商标图像对应查询预存储的候选商标图像，通过已训练好的特征提取模型和侵权预测模型，基于所获取到的候选商标图像判定目标商标图像是否侵权，提高了侵权判定效率和准确性。通过特征提取模型从目标商标图像中提取第一图像特征，并从候选商标图像中提取第二图像特征，进而根据所提取的第一图像特征和第二图像特征，对应确定作为侵权预测模型的输入特征的第三图像特征，以通过侵权预测模型进行预测获得相应的目标预测标签，提高了目标预测标签的获取效率和准确性，从而提高了侵权判定效率和准确性。当目标预测标签与预设预测标签相匹配时，则判定目标商标图像为侵权商标图像，提高了商标侵权判定结果的准确性。
在其中一个实施例中，候选商标图像有多个；特征提取模型包括第一特征提取模型和第二特征提取模型；上述商标侵权的识别方法，还包括：将目标商标图像分别与每个候选商标图像进行组合，获得多个商标图像对；步骤S206包括：对于每个商标图像对，将商标图像对中的目标商标图像输入第一特征提取模型，获得相应的第一图像特征，并将商标图像对中的候选商标图像输入第二特征提取模型进行预测，获得相应的第二图像特征；步骤S212包括：当多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将目标商标图像确定为侵权商标图像。
商标图像对是指由两个商标图像组合而成的图像对。商标图像对具体可以是由目标商标图像与候选商标图像构成的图像对。第一特征提取模型和第二特征提取模型均是根据预先获取到的训练样本集进行模型训练获得的、能够用于从商标图像中提取相应图像特征的模型。第一特征提取模型与第二特征提取模型为权重共享的特征提取模型，即为孪生特征提取模型。
在其中一个实施例中，第一特征提取模型与第二特征提取模型的训练过程中采用孪生神经网络框架，即基于两个权重共享的神经网络进行模型训练。
具体地，服务器根据目标商标图像对应查询预存储的多个候选商标图像，并将所查询到的每个候选商标图像分别与目标商标图像进行组合，对应获得由候选商标图像和目标商标图像组成的多个商标图像对。商标图像对的数量与所查询到的候选商标图像的数量相同，即商标图像对与候选商标图像相对应。对于该多个商标图像对中的每个商标图像对，服务器通过已训练的第一特征提取模型从该商标图像对中的目标商标图像中提取相应的第一图像特征，并通过已训练的第二特征提取模型从该商标图像对中的候选商标图像中提取相应的第二图像特征，并将该第一图像特征与第二图像特征进行拼接获得相应的第三图像特征，进而通过已训练的侵权预测模型根据该第三图像特征进行预测，获得相应的目标预测标签。
进一步地，服务器针对该多个商标图像对中的每个商标图像对，分别执行上述根据商标图像对预测相应的目标预测标签的相关步骤，获得每个商标图像对所对应的目标预测标签。服务器获得该多个商标图像对中的每个商标图像对所对应的目标预测标签时，判定该多个目标预测标签中是否存在表示商标侵权的标签。当判定存在表示商标侵权的标签时，服务器则判定相应的目标商标图像侵权，将该目标商标图像确定为侵权商标图像。
在其中一个实施例中，服务器获得该多个商标图像对中的每个商标图像对所对应的目标预测标签时，将该多个目标预测标签分别与预设标签进行匹配。当各个商标图像对所对应的匹配结果中存在表示匹配成功的匹配结果时，表示该多个目标预测标签中存在表示商标侵权的标签，服务器则将目标商标图像确定为侵权商标图像。当该多个匹配结果均为表示匹配失败的匹配结果时，服务器则判定目标商标图像不是侵权商标图像。
上述实施例中，根据预存储的多个候选商标图像分别对目标商标图像进行侵权识别，当识别结果中存在至少一个表示商标侵权的结果时，则判定目标商标图像为侵权商标图像，提高了商标侵权的识别准确性。
在其中一个实施例中，候选商标图像有多个；上述商标侵权的识别方法，还包括：将目标商标图像分别与每个候选商标图像进行组合，获得多个商标图像对；对于多个商标图像对，按照指定顺序迭代的执行分别执行步骤S206；步骤S212包括：在当前预测的商标图像对所对应的目标预测标签为商标侵权的标签时，停止迭代，将目标商标图像确定为侵权商标图像。
具体地，服务器根据目标商标图像对应查询预存储的多个候选商标图像，并将所查询到的每个候选商标图像分别与目标商标图像进行组合，对应获得每个候选商标图像所对应的商标图像对。该多个商标图像对由目标商标图像和相应的候选商标图像组成。服务器对该多个商标图像对中的各个商标图像对进行优先级排序，按照优先级顺序迭代的对该多个商标图像对中的每个商标图像对，执行根据该商标图像对预测相应的目标预测标签，并根据目标预测标签判定目标商标图像是否侵权的相关步骤。
进一步地，服务器按照优先级顺序迭代的根据商标图像对判定目标商标图像是否侵权的过程中，具体包括以下步骤：服务器通过已训练的特征提取模型从当前优先级的商标图像对中的目标商标图像中提取第一图像特征，以及从该商标图像对中的候选商标图像中提取第二图像特征；服务器将该第一图像特征与第二图像特征进行拼接获得第三图像特征，并通过已训练的侵权预测模型根据该第三图像特征进行预测，获得相应的目标预测标签；服务器对应判定该目标预测标签是否为表示商标侵权的标签，当判定为表示商标侵权的标签时，服务器停止迭代过程，并将相应的目标商标图像确定为侵权商标图像；当判定为表示商标不侵权的标签时，服务器针对下一优先级的商标图像对继续执行上述根据商标图像对判定目标商标图像是否侵权的相关步骤，直至当前预测的商标图像对所对应的目标预测标签为表示商标侵权的标签时，或者针对该多个商标图像对，均执行完上述根据商标图像对判定目标商标图像是否侵权的相关步骤时，停止迭代。
举例说明，假设存在已进行优先级排序的4个商标图像对A、B、C和D，服务器按照上述方法进行预测获得与商标图像对A对应的目标预测标签，当该目标预测标签为表示商标侵权的标签时，停止迭代，将目标商标图像确定为侵权商标图像；当该目标预测标签为表示商标不侵权的标签时，服务器按照上述方法继续进行预测获得与商标图像对B对应的目标预测标签，并根据商标图像对B所对应的目标预测标签是否为表示商标侵权的标签，继续执行相应的步骤。
上述实施例中，根据所获取到的多个商标图像对，依次迭代的执行根据商标图像对判定目标商标图像是否侵权的步骤，当判定目标商标图像侵权时则停止迭代，提高了商标侵权的效率和准确性。
在其中一个实施例中，上述商标侵权的识别方法中，特征提取模型的训练步骤，包括：获取预设商标图像；按照预设提取方式从预设商标图像中提取预设图像特征；将预设商标图像作为输入特征，将相应的预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
具体地，服务器在本地获取多个预设商标图像，并按照预设提取方式分别从该多个预设商标图像中提取相应的预设图像特征。服务器根据所获取到的预设商标图像和相应的预设图像特征，获得相应的训练样本集。训练样本集中的预设商标图像作为输入特征，相应的预设图像特征作为期望的输出特征。服务器根据所获取到的训练样本集，对初始化的特征提取模型进行模型训练获得已训练的特征提取模型。服务器也可以从其他用于存储商标图像的服务器获取预设商标图像，或者基于网络从指定网页获取预设商标图像。
在其中一个实施例中，服务器获取用作训练样本集的自然图像数据集，该自然图像数据集中包含大量已标注的自然图像，服务器根据所获取的训练样本集进行模型训练，获得相应的特征提取模型。通过该种训练方式训练获得的特征提取模型相当于分类器，预测过程中的输出特征为与作为输入特征的输入图像相对应的类别，即通过该特征提取模型可预测获得输入图像的类别。在上述各个实施例中的商标侵权的识别方法中，服务器将目标商 标图像与候选商标图像分别输入已训练的特征提取模型中，通过该特征提取模型的特征提取层分别获得目标商标图像与候选商标图像各自对应的图像特征。自然图像数据集具体可以是ImageNet数据集，该ImageNet数据集包含有1400多万幅图片，涵盖2万多个类别。
上述实施例中，根据所获取到的训练样本集进行模型训练，以根据训练获得的特征提取模型从商标图像中提取相应的图像特征，进而根据图像特征对应判定目标商标图像是否侵权，提高了商标侵权的识别效率和准确性。
在其中一个实施例中，上述商标侵权的识别方法中，侵权预测模型的训练步骤，包括：获取训练样本集；训练样本集包括目标商标图像对和相应的目标标签；根据目标图像对和已训练的特征提取模型获得相应的目标图像特征；将目标图像特征作为输入特征，将相应的目标标签作为期望的输出特征，对卷积神经网络进行模型训练；当达到预设条件时停止训练，获得已训练的侵权预测模型。
目标商标图像对是用于进行模型训练的商标图像对。目标商标图像对中包括由相似度很高的两个商标图像组成的商标图像对，或由两个无关的商标图像组成的商标图像对。换而言之，目标商标图像对中包括由相互侵权的两个商标图像组成的商标图像对，或者，由两个互不侵权的商标图像组成的商标图像对。目标标签是与商标图像对所对应的、用于表示组成该商标图像对的两个商标图像是否相互侵权的标签，比如目标标签为1时表示相应商标图像对中的两个商标图像相互侵权，目标标签为0时表示相应商标图像对中的两个商标图像互不侵权。预设条件是用于判定是否停止当前的模型训练过程的条件。预设条件具体可以是损失函数收敛、损失函数值趋于稳定或权重参数趋于稳定等。
具体地，服务器在本地获取多个目标商标图像对，分别确定该多个目标商标图像对中的每个目标商标图像对所对应的目标标签，并根据所获取到的目标商标图像对和相应的目标标签获得训练样本集。对于训练样本集中的每个目标商标图像对，服务器通过已训练的特征提取模型，分别从组成该目标商标图像对的每个商标图像中提取相应的图像特征，并将所提取到的两个图像特征进行拼接，获得相应的目标图像特征。服务器针对训练样本集中的每个目标商标图像对，按照上述方式分别确定相应的目标图像特征。
进一步地，服务器将该多个目标商标图像对各自对应的目标图像特征作为输入特征，将相应的目标标签作为期望的输出特征，对卷积神经网络进行模型训练。当达到预设条件时，服务器停止训练，获得已训练的侵权预测模型。
在其中一个实施例中，服务器也可以从其他用于存储商标图像的服务器获取预设商标图像，或者基于网络从指定网页获取预设商标图像。
在其中一个实施例中，服务器获取用作训练样本集的多个已标注的目标商标图像对，并根据该训练样本集和已训练的特征提取模型进行模型训练，获得已训练的侵权预测模型。训练样本集包括目标商标图像对和相应的目标标签。具体地，服务器将目标商标图像对中的一个商标图像，输入已训练的第一特征提取模型获得第一图像特征，并将该目标商标图像对中的另一个商标图像输入已训练的第二特征提取模型获得第二图像特征，并将该 第一图像特征和第二图像特征进行拼接获得第三图像特征。服务器将该第三图像特征作为输入特征，相应的目标标签作为期望的输出特征，对卷积神经网络进行模型训练获得侵权预测模型。第一特征提取模型和第二特征提取模型权重共享。
在其中一个实施例中，卷积神经网络包括至少一个卷积层、批量归一化层、最大池化层、全连接层和softmax层。为了防止过拟合，在上述模型训练过程中，对于每层的神经网络可执行Dropout，即对于组成每层神经网络的每个神经网络单元，按照指定概率将其从该神经网络中暂时丢弃，从而获得更瘦的神经网络。
上述实施例中，根据所获取到的训练样本集进行模型训练，以通过训练获得的侵权预测模型对应判定目标商标图像是否侵权，提高了商标侵权的识别效率和准确性。
在其中一个实施例中，将目标图像特征作为输入特征，将相应的目标标签作为期望的输出特征，对卷积神经网络进行模型训练，包括：将目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签；按照预设计算方式计算预测标签与相应的、作为期望的输出特征的目标标签之间的对比损失值；根据对比损失值按照预设方式训练卷积神经网络的权重参数；当达到预设条件时停止训练，获得已训练的侵权预测模型，包括：当权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
预设计算方式是预先设定的用于计算预测标签与目标标签之间的对比损失值的方式。预设计算方式比如基于对比损失函数进行计算。对比损失值是指预测标签相对于相应目标标签的差值。对比损失值具体可以是基于对比损失函数对应计算的预测标签与相应目标标签之间的差值。预设方式是指预先设定的用于根据对比损失值对卷积神经网络的权重参数进行训练的方式。预设方式具体可以是通过反向传播算法来对应训练卷积神经网络的权重参数。
具体地，服务器基于已训练的特征提取模型，获得与训练样本集中的每个目标图像对所对应的目标图像特征时，将所获得的每个目标图像特征作为输入特征依次输入卷积神经网络进行预测，获得每个目标图像特征所对应的预测标签。对于该多个目标图像特征中的每个目标图像特征，服务器基于对比损失函数依次计算每个目标图像特征所对应的预测标签与目标标签之间的对比损失值。目标标签是与作为输入特征的目标图像相对应的期望的输出特征。服务器根据计算所得的对比损失值通过反向传播算法依次训练卷积神经网络的权重参数。当权重参数趋于稳定时，即当权重参数收敛时，服务器停止训练，获得已训练的侵权预测模型。
在其中一个实施例中，服务器针对多个目标图像特征中的每个目标图像特征，迭代的执行根据目标图像特征预测获得相应的预测标签，进而计算相应的对比损失值，并根据该对比损失值对卷积神经网络的权重参数进行训练的相关步骤，且当权重参数符合预设条件时停止迭代，获得已训练的侵权预测参数。在当前迭代过程完成时，若权重参数不符合预设条件，则针对另一目标图像特征继续执行上述迭代过程，直至权重参数符合预设条件，或者针对该多个目标图像特征中的每个目标图像特征均执行完上述迭代过程时，停止迭 代。
在其中一个实施例中，在上述迭代进行的训练过程中，当对比损失函数收敛或者对比损失值趋于稳定时，停止迭代，获得已训练的侵权预测模型。
上述实施例中，在模型训练过程中，根据预测标签和相应的目标标签之间的对比损失值对应训练卷积神经网络的权重参数，以获得已训练的侵权预测模型，提高了模型训练效率和准确性。
如图3所示，在其中一个实施例中，提供了一种商标侵权的识别方法，该方法具体包括以下步骤：
S302，获取目标商标图像。
S304，根据目标商标图像查询预存储的候选商标图像；候选商标图像有多个。
S306，将目标商标图像分别与每个候选商标图像进行组合，获得多个商标图像对。
S308，对于每个商标图像对，将商标图像对中的目标商标图像输入第一特征提取模型，获得相应的第一图像特征，并将商标图像对中的候选商标图像输入第二特征提取模型进行预测，获得相应的第二图像特征。
S310，将第一图像特征和第二图像特征进行拼接获得第三图像特征。
S312，将第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签。
S314，当多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将目标商标图像确定为侵权商标图像。
上述实施例中，将目标商标图像分别与多个候选商标图像进行组合，基于组合获得的多个商标图像对，借助于特征提取模型和侵权预测模型，对目标商标图像进行侵权识别，提高了识别效率和准确性。
如图4所示，在其中一个实施例中，提供了一种商标侵权的识别方法中侵权预测模型的训练步骤，具体包括：
S402，获取训练样本集；训练样本集包括目标商标图像对和相应的目标标签。
S404，根据目标图像对和已训练的特征提取模型获得相应的目标图像特征。
S406，将目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签。
S408，按照预设计算方式计算预测标签与相应的、作为期望的输出特征的目标标签之间的对比损失值。
S410，根据对比损失值按照预设方式训练卷积神经网络的权重参数。
S412，当权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
上述实施例中，提供了侵权预测模型的训练步骤，基于该训练步骤训练获得的侵权预测模型能够用于商标侵权的识别，提高了识别效率和准确性。
应该理解的是，虽然图2-4的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的 执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，图2-4中的至少一部分步骤可以包括多个子步骤或者多个阶段，这些子步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些子步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
在其中一个实施例中，如图5所示，提供了一种商标侵权的识别装置500，包括：获取模块502、查询模块504、提取模块506、拼接模块508、预测模块510和判定模块512，其中：
获取模块502，用于获取目标商标图像。
查询模块504，用于根据目标商标图像查询预存储的候选商标图像。
提取模块506，用于将目标商标图像和候选商标图像分别输入已训练的特征提取模型进行预测，获得与目标商标图像对应的第一图像特征，以及与候选商标图像对应的第二图像特征。
拼接模块508，用于将第一图像特征和第二图像特征进行拼接获得第三图像特征。
预测模块510，用于将第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签。
判定模块512，用于当目标预测标签为商标侵权的标签时，将目标商标图像确定为侵权商标图像。
在其中一个实施例中，候选商标图像有多个；特征提取模型包括第一特征提取模型和第二特征提取模型；提取模块506，还用于将目标商标图像分别与每个候选商标图像进行组合，获得多个商标图像对；对于每个商标图像对，将商标图像对中的目标商标图像输入第一特征提取模型，获得相应的第一图像特征，并将商标图像对中的候选商标图像输入第二特征提取模型进行预测，获得相应的第二图像特征；判定模块512，还用于当多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将目标商标图像确定为侵权商标图像。
在其中一个实施例中，候选商标图像有多个；提取模块506，还用于将目标商标图像分别与每个候选商标图像进行组合，获得多个商标图像对；对于多个商标图像对，按照指定顺序迭代的执行将目标商标图像和候选商标图像分别输入已训练的特征提取模型进行预测，获得与目标商标图像对应的第一图像特征，以及与候选商标图像对应的第二图像特征的步骤；判定模块512，还用于在当前预测的商标图像对所对应的目标预测标签为商标侵权的标签时，停止迭代，将目标商标图像确定为侵权商标图像。
如图6所示，在其中一个实施例中，商标侵权的识别装置500，还包括：模型训练模块514；
模型训练模块514，用于获取预设商标图像；按照预设提取方式从预设商标图像中提 取预设图像特征；将预设商标图像作为输入特征，将相应的预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
在其中一个实施例中，模型训练模块514，还用于获取训练样本集；训练样本集包括目标商标图像对和相应的目标标签；根据目标图像对和已训练的特征提取模型获得相应的目标图像特征；将目标图像特征作为输入特征，将相应的目标标签作为期望的输出特征，对卷积神经网络进行模型训练；当达到预设条件时停止训练，获得已训练的侵权预测模型。
在其中一个实施例中，模型训练模块514，还用于将目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签；按照预设计算方式计算预测标签与相应的、作为期望的输出特征的目标标签之间的对比损失值；根据对比损失值按照预设方式训练卷积神经网络的权重参数；当权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
关于商标侵权的识别装置的具体限定可以参见上文中对于商标侵权的识别方法的限定，在此不再赘述。上述商标侵权的识别装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在其中一个实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图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。该计算机设备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候选商标图像。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商标侵权的识别方法。
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，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
一种计算机设备，包括存储器和一个或多个处理器，存储器中储存有计算机可读指令，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器实现本申请任意一个实施例中提供的商标侵权的识别方法的步骤。
一个或多个存储有计算机可读指令的非易失性计算机可读存储介质，计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器实现本申请任意一个实施例中提供的商标侵权的识别方法的步骤。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (20)

  1. 一种商标侵权的识别方法，包括：
    获取目标商标图像；
    根据所述目标商标图像查询预存储的候选商标图像；
    将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
    将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
    将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
    当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  2. 根据权利要求1所述的方法，其特征在于，所述候选商标图像有多个；所述特征提取模型包括第一特征提取模型和第二特征提取模型；所述方法还包括：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；
    所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征，包括：
    对于每个所述商标图像对，将所述商标图像对中的目标商标图像输入所述第一特征提取模型，获得相应的第一图像特征，并将所述商标图像对中的候选商标图像输入所述第二特征提取模型进行预测，获得相应的第二图像特征；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    当所述多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  3. 根据权利要求1所述的方法，其特征在于，所述候选商标图像有多个；所述方法还包括：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；及
    对于所述多个商标图像对，按照指定顺序迭代的执行所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征的步骤；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    在当前预测的商标图像对所对应的目标预测标签为商标侵权的标签时，停止迭代，将所述目标商标图像确定为侵权商标图像。
  4. 根据权利要求1至3任意一项所述的方法，其特征在于，所述特征提取模型的训练步骤，包括：
    获取预设商标图像；
    按照预设提取方式从所述预设商标图像中提取预设图像特征；及
    将所述预设商标图像作为输入特征，将相应的所述预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
  5. 根据权利要求1至3任意一项所述的方法，其特征在于，所述侵权预测模型的训练步骤，包括：
    获取训练样本集；所述训练样本集包括目标商标图像对和相应的目标标签；
    根据所述目标图像对和已训练的特征提取模型获得相应的目标图像特征；
    将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练；及
    当达到预设条件时停止训练，获得已训练的侵权预测模型。
  6. 根据权利要求5所述的方法，其特征在于，所述将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练，包括：
    将所述目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签；
    按照预设计算方式计算所述预测标签与相应的、作为期望的输出特征的所述目标标签之间的对比损失值；及
    根据所述对比损失值按照预设方式训练所述卷积神经网络的权重参数；
    所述当达到预设条件时停止训练，获得已训练的侵权预测模型，包括：
    当所述权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
  7. 一种商标侵权的识别装置，包括：
    获取模块，用于获取目标商标图像；
    查询模块，用于根据所述目标商标图像查询预存储的候选商标图像；
    提取模块，用于将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
    拼接模块，用于将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
    预测模块，用于将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
    判定模块，用于当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  8. 根据权利要求7所述的装置，其特征在于，还包括：模型训练模块；
    所述模型训练模块，用于获取预设商标图像；按照预设提取方式从所述预设商标图像中提取预设图像特征；将所述预设商标图像作为输入特征，将相应的所述预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
  9. 一种计算机设备，包括存储器和一个或多个处理器，所述存储器中储存有计算机可读指令，所述计算机可读指令被所述一个或多个处理器执行时，使得所述一个或多个处理器执行以下步骤：
    获取目标商标图像；
    根据所述目标商标图像查询预存储的候选商标图像；
    将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
    将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
    将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
    当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  10. 根据权利要求9所述的计算机设备，其特征在于，所述候选商标图像有多个；所述特征提取模型包括第一特征提取模型和第二特征提取模型；所述处理器执行所述计算机可读指令时还执行以下步骤：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；
    所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征，包括：
    对于每个所述商标图像对，将所述商标图像对中的目标商标图像输入所述第一特征提取模型，获得相应的第一图像特征，并将所述商标图像对中的候选商标图像输入所述第二特征提取模型进行预测，获得相应的第二图像特征；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    当所述多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  11. 根据权利要求9所述的计算机设备，其特征在于，所述候选商标图像有多个；所述处理器执行所述计算机可读指令时还执行以下步骤：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；及
    对于所述多个商标图像对，按照指定顺序迭代的执行所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的 第一图像特征，以及与所述候选商标图像对应的第二图像特征的步骤；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    在当前预测的商标图像对所对应的目标预测标签为商标侵权的标签时，停止迭代，将所述目标商标图像确定为侵权商标图像。
  12. 根据权利要求9至11任意一项所述的计算机设备，其特征在于，所述处理器执行所述计算机可读指令时还执行特征提取模型的训练步骤，包括：
    获取预设商标图像；
    按照预设提取方式从所述预设商标图像中提取预设图像特征；及
    将所述预设商标图像作为输入特征，将相应的所述预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
  13. 根据权利要求9至11任意一项所述的计算机设备，其特征在于，所述处理器执行所述计算机可读指令时还执行侵权预测模型的训练步骤，包括：
    获取训练样本集；所述训练样本集包括目标商标图像对和相应的目标标签；
    根据所述目标图像对和已训练的特征提取模型获得相应的目标图像特征；
    将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练；及
    当达到预设条件时停止训练，获得已训练的侵权预测模型。
  14. 根据权利要求13所述的计算机设备，其特征在于，所述将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练，包括：
    将所述目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签；
    按照预设计算方式计算所述预测标签与相应的、作为期望的输出特征的所述目标标签之间的对比损失值；及
    根据所述对比损失值按照预设方式训练所述卷积神经网络的权重参数；
    所述当达到预设条件时停止训练，获得已训练的侵权预测模型，包括：
    当所述权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
  15. 一个或多个存储有计算机可读指令的非易失性计算机可读存储介质，所述计算机可读指令被一个或多个处理器执行时，使得所述一个或多个处理器执行以下步骤：
    获取目标商标图像；
    根据所述目标商标图像查询预存储的候选商标图像；
    将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征；
    将所述第一图像特征和所述第二图像特征进行拼接获得第三图像特征；
    将所述第三图像特征输入已训练的侵权预测模型进行预测，获得目标预测标签；及
    当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  16. 根据权利要求15所述的存储介质，其特征在于，所述候选商标图像有多个；所述特征提取模型包括第一特征提取模型和第二特征提取模型；所述计算机可读指令被所述处理器执行时还执行以下步骤：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；
    所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征，包括：
    对于每个所述商标图像对，将所述商标图像对中的目标商标图像输入所述第一特征提取模型，获得相应的第一图像特征，并将所述商标图像对中的候选商标图像输入所述第二特征提取模型进行预测，获得相应的第二图像特征；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    当所述多个商标图像对各自对应的目标预测标签中，存在商标侵权的标签时，将所述目标商标图像确定为侵权商标图像。
  17. 根据权利要求15所述的存储介质，其特征在于，所述候选商标图像有多个；所述计算机可读指令被所述处理器执行时还执行以下步骤：
    将所述目标商标图像分别与每个所述候选商标图像进行组合，获得多个商标图像对；及
    对于所述多个商标图像对，按照指定顺序迭代的执行所述将所述目标商标图像和所述候选商标图像分别输入已训练的特征提取模型进行预测，获得与所述目标商标图像对应的第一图像特征，以及与所述候选商标图像对应的第二图像特征的步骤；
    所述当所述目标预测标签为商标侵权的标签时，将所述目标商标图像确定为侵权商标图像，包括：
    在当前预测的商标图像对所对应的目标预测标签为商标侵权的标签时，停止迭代，将所述目标商标图像确定为侵权商标图像。
  18. 根据权利要求15至17任意一项所述的存储介质，其特征在于，所述计算机可读指令被所述处理器执行时还执行特征提取模型的训练步骤，包括：
    获取预设商标图像；
    按照预设提取方式从所述预设商标图像中提取预设图像特征；及
    将所述预设商标图像作为输入特征，将相应的所述预设图像特征作为期望的输出特征，对初始化的特征提取模型进行训练获得已训练的特征提取模型。
  19. 根据权利要求15至17任意一项所述的存储介质，其特征在于，所述计算机可读指令被所述处理器执行时还执行侵权预测模型的训练步骤，包括：
    获取训练样本集；所述训练样本集包括目标商标图像对和相应的目标标签；
    根据所述目标图像对和已训练的特征提取模型获得相应的目标图像特征；
    将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练；及
    当达到预设条件时停止训练，获得已训练的侵权预测模型。
  20. 根据权利要求19所述的存储介质，其特征在于，所述将所述目标图像特征作为输入特征，将相应的所述目标标签作为期望的输出特征，对卷积神经网络进行模型训练，包括：
    将所述目标图像特征作为输入特征输入卷积神经网络进行预测，获得相应的预测标签；
    按照预设计算方式计算所述预测标签与相应的、作为期望的输出特征的所述目标标签之间的对比损失值；及
    根据所述对比损失值按照预设方式训练所述卷积神经网络的权重参数；
    所述当达到预设条件时停止训练，获得已训练的侵权预测模型，包括：
    当所述权重参数符合预设条件时停止训练，获得已训练的侵权预测模型。
PCT/CN2019/071363 2018-09-10 2019-01-11 商标侵权的识别方法、装置、计算机设备和存储介质 Ceased WO2020052183A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811051616.6 2018-09-10
CN201811051616.6A CN109376741A (zh) 2018-09-10 2018-09-10 商标侵权的识别方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2020052183A1 true WO2020052183A1 (zh) 2020-03-19

Family

ID=65404760

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/071363 Ceased WO2020052183A1 (zh) 2018-09-10 2019-01-11 商标侵权的识别方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN109376741A (zh)
WO (1) WO2020052183A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814820A (zh) * 2020-05-18 2020-10-23 北京迈格威科技有限公司 图像处理方法及装置
CN112288670A (zh) * 2020-11-18 2021-01-29 苏州臻迪智能科技有限公司 一种图像处理方法、装置、设备及存储介质
CN114428878A (zh) * 2022-04-06 2022-05-03 广东知得失网络科技有限公司 一种商标图像检索方法及系统
CN118051637A (zh) * 2024-02-05 2024-05-17 北京卓佳国际知识产权代理有限公司 一种商标查询方法、装置及计算机设备
CN118691857A (zh) * 2023-12-29 2024-09-24 中国计量大学 一种商标侵权检测方法和系统
US12406144B2 (en) 2023-05-19 2025-09-02 Raj Abhyanker Linguistic analysis to automatically generate a hypothetical likelihood of confusion office action using Dupont factors

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110033018B (zh) * 2019-03-06 2023-10-31 平安科技（深圳）有限公司 图形相似度判断方法、装置及计算机可读存储介质
CN112163140A (zh) * 2020-10-27 2021-01-01 北京梦知网科技有限公司 一种检测商标侵权的方法、装置及系统
CN114529784B (zh) * 2022-02-18 2022-11-18 广东数源智汇科技有限公司 一种面向电商数据的商标侵权分析方法及系统
CN115880484A (zh) * 2022-04-28 2023-03-31 中国计量大学 一种基于分类匹配的商标侵权识别方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103258037A (zh) * 2013-05-16 2013-08-21 西安工业大学 一种针对多组合内容的商标识别检索方法
CN105701501A (zh) * 2016-01-04 2016-06-22 北京大学 一种商标图像识别方法
CN107273535A (zh) * 2017-06-29 2017-10-20 朱峰 一种商标智能分析系统
CN108038122A (zh) * 2017-11-03 2018-05-15 福建师范大学 一种商标图像检索的方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1304993C (zh) * 2003-08-04 2007-03-14 中国科学院自动化研究所 商标检索方法
CN102799653A (zh) * 2012-06-29 2012-11-28 中国科学院自动化研究所 一种基于空间连通域预定位的商标检测方法
CN102982165B (zh) * 2012-12-10 2015-05-13 南京大学 一种大规模人脸图像检索方法
CN106530194B (zh) * 2015-09-09 2020-02-07 阿里巴巴集团控股有限公司 一种疑似侵权产品图片的检测方法及装置
CN105654056A (zh) * 2015-12-31 2016-06-08 中国科学院深圳先进技术研究院 人脸识别的方法及装置
CN108009560B (zh) * 2016-11-02 2021-05-11 广州图普网络科技有限公司 商品图像相似类别判定方法及装置
CN106682127A (zh) * 2016-12-13 2017-05-17 上海联影医疗科技有限公司 图像搜索系统及方法
CN108154182A (zh) * 2017-12-25 2018-06-12 合肥阿巴赛信息科技有限公司 基于成对比较网络的珠宝相似性度量方法
CN108446612A (zh) * 2018-03-07 2018-08-24 腾讯科技（深圳）有限公司 车辆识别方法、装置及存储介质
CN108427927B (zh) * 2018-03-16 2020-11-27 深圳市商汤科技有限公司 目标再识别方法和装置、电子设备、程序和存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103258037A (zh) * 2013-05-16 2013-08-21 西安工业大学 一种针对多组合内容的商标识别检索方法
CN105701501A (zh) * 2016-01-04 2016-06-22 北京大学 一种商标图像识别方法
CN107273535A (zh) * 2017-06-29 2017-10-20 朱峰 一种商标智能分析系统
CN108038122A (zh) * 2017-11-03 2018-05-15 福建师范大学 一种商标图像检索的方法

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814820A (zh) * 2020-05-18 2020-10-23 北京迈格威科技有限公司 图像处理方法及装置
CN112288670A (zh) * 2020-11-18 2021-01-29 苏州臻迪智能科技有限公司 一种图像处理方法、装置、设备及存储介质
CN114428878A (zh) * 2022-04-06 2022-05-03 广东知得失网络科技有限公司 一种商标图像检索方法及系统
US12406144B2 (en) 2023-05-19 2025-09-02 Raj Abhyanker Linguistic analysis to automatically generate a hypothetical likelihood of confusion office action using Dupont factors
CN118691857A (zh) * 2023-12-29 2024-09-24 中国计量大学 一种商标侵权检测方法和系统
CN118051637A (zh) * 2024-02-05 2024-05-17 北京卓佳国际知识产权代理有限公司 一种商标查询方法、装置及计算机设备

Also Published As

Publication number Publication date
CN109376741A (zh) 2019-02-22

Similar Documents

Publication Publication Date Title
WO2020052183A1 (zh) 商标侵权的识别方法、装置、计算机设备和存储介质
CN113157863B (zh) 问答数据处理方法、装置、计算机设备及存储介质
CN109447105B (zh) 合同审核方法、装置、计算机设备和存储介质
CN108830277B (zh) 语义分割模型的训练方法、装置、计算机设备和存储介质
CN109285105B (zh) 水印检测方法、装置、计算机设备和存储介质
WO2020140665A1 (zh) 双录视频质量检测方法、装置、计算机设备和存储介质
WO2020082569A1 (zh) 文本分类方法、装置、计算机设备和存储介质
WO2021169111A1 (zh) 简历筛选方法、装置、计算机设备和存储介质
WO2021043076A1 (zh) 网络发布数据处理方法、装置、计算机设备和存储介质
WO2020057022A1 (zh) 关联推荐方法、装置、计算机设备和存储介质
CN112699923A (zh) 文档分类预测方法、装置、计算机设备及存储介质
CN116108363B (zh) 基于标签引导的不完备多视图多标签分类方法和系统
CN113254687B (zh) 图像检索、图像量化模型训练方法、装置和存储介质
JP6971514B1 (ja) 情報処理装置、情報処理方法及びプログラム
CN111507285A (zh) 人脸属性识别方法、装置、计算机设备和存储介质
WO2020048048A1 (zh) 不平衡样本数据预处理方法、装置和计算机设备
CN110750523A (zh) 数据标注方法、系统、计算机设备和存储介质
CN114168780B (zh) 多模态数据处理方法、电子设备及存储介质
CN110046577B (zh) 行人属性预测方法、装置、计算机设备和存储介质
CN115374277B (zh) 评论处理模型训练、评论排序方法、装置、设备及介质
CN117093682A (zh) 意图识别方法、装置、计算机设备及存储介质
CN114691897B (zh) 一种深度自适应多模态哈希检索方法及相关设备
CN110162689B (zh) 信息推送方法、装置、计算机设备和存储介质
CN117610666A (zh) 问答模型的训练、答案确定方法、装置、设备及介质
CN115424001A (zh) 场景相似度估计方法、装置、计算机设备和存储介质

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: 19859387

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 16.06.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19859387

Country of ref document: EP

Kind code of ref document: A1