WO2021017261A1 - 识别模型训练方法、图像识别方法、装置、设备及介质 - Google Patents

识别模型训练方法、图像识别方法、装置、设备及介质 Download PDF

Info

Publication number
WO2021017261A1
WO2021017261A1 PCT/CN2019/116489 CN2019116489W WO2021017261A1 WO 2021017261 A1 WO2021017261 A1 WO 2021017261A1 CN 2019116489 W CN2019116489 W CN 2019116489W WO 2021017261 A1 WO2021017261 A1 WO 2021017261A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
image
sample image
training
model
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/CN2019/116489
Other languages
English (en)
French (fr)
Inventor
王健宗
彭俊清
瞿晓阳
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 WO2021017261A1 publication Critical patent/WO2021017261A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/22Matching criteria, e.g. proximity measures
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/088Non-supervised learning, e.g. competitive learning
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172Classification, e.g. identification
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02TCLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00Road transport of goods or passengers
    • Y02T10/10Internal combustion engine [ICE] based vehicles
    • Y02T10/40Engine management systems

Definitions

  • This application relates to the field of image recognition technology, and in particular to a recognition model training method, image recognition method, device, equipment and medium.
  • face detectors With the rapid development of face detection technology, the accuracy of face detectors in identifying large and medium-sized face images has been rapidly improved. In the actual application process of face detection technology, it is usually necessary to recognize smaller or more obscure face images to identify the identity of the person corresponding to the face image. However, traditional face detectors treat smaller or more obscure persons. The recognition accuracy of face images is low, which affects the effect of practical applications. For example, in situations where criminal suspects are tracked, missing persons searched, or other persons need to be searched, a face detector can be used to identify the surveillance images taken by the surveillance equipment to identify criminal suspects, missing persons or other target persons.
  • the faces of suspects, missing persons, or other target persons in the surveillance images are small or blurry.
  • traditional face detectors are used for identification, The identities of criminal suspects, missing persons or other target persons cannot be accurately identified, which affects the efficiency of tracking target persons.
  • the embodiments of the present application provide a recognition model training method, an image recognition method, a device, equipment, and a medium to solve the problem of low accuracy when the current face detector recognizes a small or fuzzy face image.
  • a recognition model training method including:
  • the first annotation information includes a current image identifier, a face identifier, and a first fuzzy identifier
  • a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail to obtain the target positive sample image and the target negative sample image carrying second annotation information, and the second annotation information includes the source image identifier , Face identification and second fuzzy identification;
  • the target training data is input to the GAN model for model training, and the target GAN model is obtained.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology.
  • a recognition model training device includes:
  • An original sample image acquisition module configured to acquire an original positive sample image and an original negative sample image carrying first annotation information, where the first annotation information includes a current image identifier, a face identifier, and a first fuzzy identifier;
  • a sample thumbnail obtaining module configured to perform down-sampling processing on the original positive sample image and the original negative sample image, and obtain corresponding positive sample thumbnails and negative sample thumbnails respectively;
  • the target sample image acquisition module is configured to use a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail, to obtain the target positive sample image and the target negative sample image carrying the second annotation information, the
  • the second annotation information includes source image identification, face identification and second fuzzy identification;
  • the target detector acquisition module is configured to input the target positive sample image and the target negative sample image with different face identifiers into the MB-FCN model for model training, and obtain the target MB-FCN detector;
  • the target training data acquisition module is configured to use the original positive sample image and the target positive sample image matching the current image identifier in the first annotation information with the source image identifier in the second annotation information as one Group target training data;
  • the target GAN model acquisition module is configured to input the target training data into the GAN model for model training, and obtain the target GAN model.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology.
  • a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and capable of running on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the first annotation information includes a current image identifier, a face identifier, and a first fuzzy identifier
  • a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail to obtain the target positive sample image and the target negative sample image carrying second annotation information, and the second annotation information includes the source image identifier , Face identification and second fuzzy identification;
  • the target training data is input to the GAN model for model training, and the target GAN model is obtained.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology.
  • One or more readable storage media storing computer readable instructions
  • the computer readable storage medium storing computer readable instructions
  • the one Or multiple processors perform the following steps:
  • the first annotation information includes a current image identifier, a face identifier, and a first fuzzy identifier
  • a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail to obtain the target positive sample image and the target negative sample image carrying second annotation information, and the second annotation information includes the source image identifier , Face identification and second fuzzy identification;
  • the target training data is input to the GAN model for model training, and the target GAN model is obtained.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology.
  • An image recognition method including:
  • the detection result of determining the person corresponding to the target face image as the target person is obtained.
  • An image recognition device includes:
  • the target face image acquisition module is configured to acquire the image to be recognized, and use the above-mentioned target MB-FCN detector to perform face detection and image interception on the image to be recognized to obtain the target face image;
  • the target generation image acquisition module is configured to use the generation network formed based on the super-resolution reconstruction technology in the target GAN model to generate the target face image to acquire the target generation image;
  • the feature similarity acquisition module is configured to perform feature similarity calculation between the target generated image and the target person image in the target image library to obtain the feature similarity
  • the detection result obtaining module is configured to obtain the detection result of determining the person corresponding to the target face image as the target person if the feature similarity is greater than the similarity threshold.
  • a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and capable of running on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the detection result of determining the person corresponding to the target face image as the target person is obtained.
  • One or more readable storage media storing computer readable instructions
  • the computer readable storage medium storing computer readable instructions
  • the one Or multiple processors perform the following steps:
  • the detection result of determining the person corresponding to the target face image as the target person is obtained.
  • FIG. 1 is a schematic diagram of an application environment of the recognition model training method or the image recognition method in an embodiment of the present application
  • FIG. 2 is a flowchart of the recognition model training method in an embodiment of the present application
  • FIG. 3 is another flowchart of the recognition model training method in an embodiment of the present application.
  • FIG. 5 is another flowchart of the recognition model training method in an embodiment of the present application.
  • Fig. 6 is another flowchart of the recognition model training method in an embodiment of the present application.
  • FIG. 7 is another flowchart of the image recognition method in an embodiment of the present application.
  • FIG. 8 is a schematic diagram of the recognition model training device in an embodiment of the present application.
  • FIG. 9 is a schematic diagram of an image recognition device in an embodiment of the present application.
  • Fig. 10 is a schematic diagram of a computer device in an embodiment of the present application.
  • the recognition model training method provided by the embodiment of the present application can be applied in the application environment as shown in FIG. 1.
  • the recognition model training method is applied to a recognition model training system.
  • the recognition model training system includes a client and a server as shown in FIG. 1.
  • the client and the server communicate through a network and can be used for training to accurately identify small Or the recognition model of a more obscure face image to expand the application of the recognition model in the field of image recognition.
  • the client is also called the client, which refers to the program that corresponds to the server and provides local services to the client.
  • the client can be installed on, but not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a recognition model training method is provided. Taking the recognition model training method applied to the server shown in FIG. 1 as an example, the recognition model training method includes the following steps:
  • S201 Obtain an original positive sample image and an original negative sample image carrying first annotation information, where the first annotation information includes a current image identifier, a face identifier, and a first fuzzy identifier.
  • the original positive sample image is a clearer image containing a human face used for model training.
  • the original negative sample image is a clearer image without human face used for model training.
  • the first label information is label information corresponding to the original positive sample image and the original negative sample image, and is information used to identify the image attributes of the original positive sample image and the original negative sample image.
  • the server After the server obtains the original positive sample image and the original negative sample image, it needs to use a preset labeling rule to label the original positive sample image and the original negative sample image so that each original positive sample image and the original negative sample image
  • the images all carry corresponding first annotation information.
  • the preset labeling rule is a rule for labeling the current image ID, source image ID, face ID, and fuzzy ID corresponding to the image.
  • the current image identifier is a unique identifier used to identify the image to be annotated, which can be determined according to the serial number of the image acquisition, for example, T001 represents the first image, and T002 represents the second image.
  • the source image identifier is the identifier used to identify the source image corresponding to the image to be annotated. It can be understood as the identifier corresponding to the source image that intercepts the original positive sample image or the original negative sample image, such as P001 for the first source image, and P002 for the first image. 2 source images.
  • the face identifier refers to the identifier used to distinguish whether the image to be labeled contains a human face.
  • the face identifier includes a first face identifier used to indicate that the image contains a human face and a second person used to indicate that the image does not contain a human face Face logo.
  • the first face identifier may be represented by 1
  • the second face identifier may be represented by 0.
  • the fuzzy mark is a mark used to distinguish whether the image to be marked is a fuzzy image.
  • the fuzzy mark includes a first fuzzy mark used to indicate that the image is not blurred and a second fuzzy mark used to indicate that the image is fuzzy, that is, the first fuzzy mark It is used to indicate a clearer image.
  • the first fuzzy identification can be represented by 0, and the second fuzzy identification can be represented by 1.
  • the first annotation information carried by the original positive sample image includes the current image identifier, source image identifier, first face identifier, and first face identifier.
  • a fuzzy logo For example, ⁇ T001, P001, 1, 0 ⁇ indicates that the image with the current image identification of T001 is a clearer image containing the human face intercepted from the source image with the source identification of P001.
  • the first annotation information carried by the original negative sample image includes the current image identifier, the source image identifier, the second face identifier and The first fuzzy identification.
  • ⁇ T002, P002, 0, 0 ⁇ means that the image with the current image identification of T002 is a clearer image that does not contain a human face, which is intercepted from the source image with the source identification of P002.
  • the original positive sample image and the original negative sample image can be restricted to have the same size.
  • the original positive sample image containing the human face and the original negative sample image excluding the human face with the same image size are cut from the same clearer image.
  • S202 Perform down-sampling processing on the original positive sample image and the original negative sample image, and obtain corresponding positive sample thumbnails and negative sample thumbnails respectively.
  • subsampled also known as dowmsampled
  • dowmsampled is a process used to reduce an image so that the image fits the size of the display area or is used to generate a thumbnail of the image.
  • the down-sampling process if the size of an image I is M*N, perform s times down-sampling processing to obtain a resolution image of (M/s)*(N/s) size, at this time s Is the common divisor of M and N.
  • the image in the s*s window of the image needs to be turned into a pixel, and the value of this pixel is the average value of all pixels in the window.
  • methods such as nearest neighbor interpolation, bilinear interpolation, mean interpolation, and median interpolation can be used but not limited to.
  • the server performs different levels of down-sampling processing on the original positive sample image and the original negative sample image, so as to reduce the original positive sample image and the original negative sample image into positive sample thumbnails and negative sample thumbnails with different scaling ratios. Reduce its resolution to achieve the purpose of blurring. For example, the original positive sample image and the original negative sample image can be reduced into 20% positive sample thumbnails and negative sample thumbnails, respectively.
  • S203 Use a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail to obtain the target positive sample image and the target negative sample image carrying the second annotation information.
  • the second annotation information includes the source image identifier, the face identifier, and the first image. 2. Fuzzy identification.
  • the server first uses a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail, and obtains the corresponding target positive sample image and target negative sample image, so that the acquired target positive sample image and target negative sample image are Smaller or more obscure, used as training data for training a recognition model that can recognize smaller or more obscure images. Then annotate the target positive sample image and the target negative sample image to obtain the target positive sample image and the target negative sample image carrying the second annotation information.
  • the screenshot tool includes but is not limited to the OpenCV screenshot tool used in this embodiment.
  • OpenCV Open Source Computer Vision Library, Open Source Computer Vision Library
  • OpenCV Open Source Computer Vision Library
  • Open Source Computer Vision Library is a cross-platform computer vision library issued under a BSD license (open source), which can run on Linux , Windows, Android and Mac OS operating systems. It is lightweight and efficient—consisting of a series of C functions and a small number of C++ classes, it also provides interfaces to languages such as Python, Ruby, and MATLAB, and implements many common algorithms in image processing
  • the second annotation information carried by the target positive sample image includes the current image identifier, the source image identifier, the first face identifier, and the second fuzzy identifier.
  • the second label information of the target positive sample image acquired based on the original positive sample image with the first label information ⁇ T001, P001, 1, 0 ⁇ is (M001, T001, 1, 1 ⁇ , that is, the current image is identified as
  • the image of M001 is a blurry image containing a human face obtained from the source image with the source identification T001.
  • the source image with the source identification T001 here refers to the original positive sample image with the current image identification T001.
  • the current image identification T001 in the first annotation information of the original positive sample image matches the source image identification T001 in the second annotation information of the target positive sample image, indicating that the target positive sample image is based on the clearer original positive sample image The image acquired after blurring.
  • the second annotation information carried by the target negative sample image includes the current image identifier, the source image identifier, the second face identifier, and the second fuzzy identifier.
  • the second label information of the target negative sample image acquired based on the original negative sample image with the first label information ⁇ T002, P002, 0, 0 ⁇ is ⁇ M002, T002, 0, 1 ⁇ , that is, the current image is identified as
  • the image of M002 is a blurry image containing a human face obtained from a source image with a source ID of T002.
  • the source image with a source ID of T002 here refers to the original negative sample image with the current image ID of T002.
  • the current image identifier T002 in the first annotation information of the original negative sample image matches the source image identifier T002 in the second annotation information of the target negative sample image, indicating that the target negative sample image is based on the clearer original negative image.
  • the image obtained after the sample image is blurred.
  • S204 Input the target positive sample image and the target negative sample image with different face identifiers into the MB-FCN model for model training, and obtain the target MB-FCN detector.
  • the target MB-FCN (Multi-branch fully convolutional network, multi-branch fully convolutional network) detector is a detector used to identify whether a blurry image is a face image, so that the target MB-FCN detector can be used later. Recognize small or fuzzy faces in the image to be recognized. Since the target MB-FCN detector is a detector used to identify whether a blurry image is a face image, in order to train the detector, it is necessary to use blurry images that contain and do not contain faces for training.
  • the positive sample image of the target carrying the first face identifier and the second fuzzy identifier (that is, the blurry image containing the human face) is used as the positive sample of the target MB-FCN detector, and the second face identifier and the second fuzzy identifier will be carried
  • the target negative sample image (that is, the blurry image that does not contain the human face) is used as the negative sample of the target MB-FCN detector.
  • FCN fully convolutional network, fully convolutional network
  • FCN turns the fully connected layer of CNN into a convolutional layer, making the entire network a network with only a convolutional layer and a pooling layer, so the network is called a fully convolutional network.
  • FCN is generally used for semantic segmentation of images, so each pixel on the image needs to be classified. This requires an upsampling to upsample the final output to the size of the original image, thereby supporting semantic segmentation, making it based on Compared with CNN detector, FCN detector has higher recognition accuracy and faster recognition efficiency.
  • the MB-FCN Multi-branch fully convolutional network
  • FCN Fast convolutional network
  • MB-FCN can use feature extractors of different scales to extract richer features. Compared with the features extracted by FCN, the characterization effect is better, which helps the model to have a higher recognition accuracy.
  • MB-FCN considers the efficiency and effectiveness in the design process, and can process facial features of all proportions through the backbone network only once. Compared with the method where FCN requires multiple passes, MB-FCN saves calculations. Therefore, it is more effective, making the recognition efficiency of the model faster.
  • S205 Use the original positive sample image and the target positive sample image that match the current image identifier in the first annotation information with the source image identifier in the second annotation information as a set of target training data.
  • the target training data is the training data used to train the GAN model.
  • the first annotation information carried by the original positive sample image includes the current image identifier, the source image identifier, the first face identifier, and the first fuzzy identifier, that is, the original positive sample image clearly contains the human face. image.
  • the second label information carried by the target positive sample image includes the current image identifier, the source image identifier, the first face identifier, and the second fuzzy identifier, that is, the target positive sample image is a relatively fuzzy image containing a human face.
  • the target positive sample image is based on the clearer original positive image.
  • the image obtained after the sample image is blurred.
  • the original positive sample image and the target positive sample image whose current image identifier in the first annotation information matches the source image identifier in the second annotation information are used as a set of target training data, so that each set of target training data
  • the original positive sample image and the target positive sample image in are based on two images with different definitions of the same face image, so that in the subsequent GAN model training process, the clearer original positive sample image can be used to compare the blurry target positive sample
  • the image is verified to update the model parameters in the GAN model.
  • the generation network in the target GAN model is a generation network formed based on the super-resolution reconstruction technology.
  • Generative Adversarial Networks is a deep learning model and an unsupervised learning network.
  • the Generative Adversarial Network (GAN) framework includes two sub-networks, one is a Generative and the other is a Discriminative Network, which produces a fairly good output through mutual game learning between the Generative Network and the Discriminative Network.
  • super resolution reconstruction Super Resolution Reconstruction, hereinafter referred to as SPN
  • SPN Super Resolution Reconstruction
  • the loss function is used, and the loss function is used to force the network to distinguish whether the image contains human faces, that is, to distinguish human faces from non-human faces.
  • the server inputs each set of target training data to the GAN model for model training, that is, the clearer original positive sample data and the fuzzy target positive sample image of the same content are input to the GAN model for training.
  • a generation network based on super-resolution reconstruction technology is used to reconstruct the fuzzy target positive sample image, and then the generated image and the original positive sample image are input to the GAN discriminant network for processing
  • Use the output value of the discriminant network to iteratively update, thereby updating the generation network of the GAN model and the model parameters corresponding to the discriminant network, so as to obtain the target GAN model to ensure that the target GAN model will be a smaller or more obscure low-resolution image.
  • a clearer super-resolution image is formed so that the super-resolution image can be used for face recognition, thereby ensuring the accuracy and efficiency of face recognition.
  • the server counts the number of targets corresponding to the target training data, and if the number of targets is greater than the number threshold, the target training data is divided into at least two target batches of target training data, so that the target training data of each target batch The target quantity of data is less than the quantity threshold. Then, the target training data is input into the GAN model in batches for model training, and a small batch of gradient descent algorithm is used to update the model parameters in the GAN model to obtain the target GAN model.
  • This process of dividing the target training data into batches for GAN model training can avoid that the sample set of the target training data is too large and the system GPU memory is limited, which affects the efficiency and accuracy of GAN model training.
  • the Mini-batch Gradient Descent (MBGD) algorithm is to update the model parameters, accumulate the errors generated in the training process in batches according to preset batches, and obtain the accumulated errors corresponding to several batches.
  • a backward propagation algorithm can be used to calculate the minimum value of the loss function.
  • Back propagation algorithm (Back Propagation, BP algorithm for short) is a training and learning method in neural network learning, which is used to adjust the weights and offsets between each node in the neural network.
  • the original positive sample image and the original negative sample image carrying the first annotation information are down-sampling and image interception operations are performed to obtain the target positive sample image and the original negative sample image carrying the second annotation information.
  • the target negative sample image so that the original positive sample image and the target positive sample image have an association relationship, help to ensure the realization of subsequent model training.
  • the original positive sample image and the target positive sample image that match the current image identifier in the first annotation information with the source image identifier in the second annotation information are used as a set of target training data, and the target training data is used to train the GAN model to Update the model parameters in the generation network and the discriminant network formed based on super-resolution reconstruction technology, so that the generated target GAN model can reconstruct the smaller or blurry low-resolution image into a clearer super-resolution image, so that Use super-resolution images for face recognition to ensure the accuracy and efficiency of face recognition.
  • step S201 namely obtaining the original positive sample image and the original negative sample image carrying the first annotation information, specifically includes the following steps:
  • S301 Obtain a training sample image including a human face from an image database.
  • the image database is a database used to store training sample images containing human faces.
  • the image database may be a system local database, that is, a local database connected to the server, so that the server can quickly obtain corresponding training sample images containing human faces.
  • the image database may also be a network image database, that is, a database used to store image information on the Internet, such as a Baidu image database.
  • the training sample image is an unprocessed image used for model training.
  • the server needs to use a crawler tool to grab training sample images containing human faces from the network image database.
  • crawler tools refer to web crawlers (also known as web spiders or web robots), which are programs or scripts that automatically crawl information on the World Wide Web according to certain rules.
  • the crawler tool includes but is not limited to the Python crawler tool. That is, a crawler tool is a tool used to crawl images that meet a specific condition from the Internet, and the specific condition can be set to include a human face.
  • the server uses a crawler tool to execute the crawler file, and crawls training sample images that meet the data crawling conditions set by the crawler file from the network image library.
  • the crawler file includes, but is not limited to, two data crawling conditions of target URL and search keywords.
  • the target URL is the URL of the target website in the crawler file that is used to limit the network image library where the image to be crawled is located.
  • the URL (short for Uniform Resource Locator, that is, uniform resource locator) is available on the Internet
  • a concise representation of the location and access method of the resource is the address of a standard resource on the Internet, that is, the URL address of the network image library.
  • the search keywords refer to keywords in the crawler file that are used to limit the common characteristics of the images that the crawler file needs to crawl, and can be specifically set as human faces.
  • S302 Use a fuzzy detection algorithm to perform fuzzy detection on the training sample image, and obtain the fuzzy degree corresponding to the training sample.
  • the recognition model training process requires clearer images to verify the fuzzy images in the recognition model training process, and the original positive sample images and original negative sample images are both clearer sample images, the training samples need to be tested in advance The blurriness of the image, so as to filter the training sample images based on the blurriness to obtain valid sample images that can be trained for the model.
  • the blur detection algorithm is an algorithm used to detect the blur degree of an image, and the blur detection algorithm may use a detection algorithm commonly used in the prior art.
  • S303 Determine a valid sample image based on the blur degree of the training sample image.
  • the server compares the fuzziness of the training sample image with the blur threshold preset by the system; if the fuzziness of the training sample image is less than the blur threshold, the training sample image is determined as a valid sample image; if the training sample image is blurred If the degree is not less than the blur threshold, the training sample image is not used as the effective sample image, so as to achieve the purpose of filtering the fuzzy training sample image, so that the final definition of the effective sample image reaches the preset standard, which is helpful Ensure the accuracy of subsequent recognition model training.
  • the blur threshold is a preset threshold used to evaluate whether the blur of the image reaches a preset standard.
  • the effective sample image is a training sample image with a blur degree less than the blur threshold, and can be used as a source image for subsequent image capture.
  • S304 Use a face detection algorithm to perform face detection on the effective sample image to obtain the size of the face area.
  • the face detection algorithm is an algorithm used to identify whether an image contains a face, and in this embodiment, a face detection algorithm commonly used in the industry can be used.
  • the server uses a face detection algorithm to perform face detection on valid sample images to identify the face area in a rectangular frame circumscribed by the face, and determine the size of the face area according to the rectangular frame circumscribed by the face.
  • the rectangular frame circumscribed by the face is a rectangular frame that uses a face detection algorithm to detect a face area in the image.
  • the circumscribed rectangular frame of the face is a virtual face frame, which is used to select the face area recognized by the face detection algorithm, and the size of the circumscribed rectangular frame of the face is adjusted autonomously with the size of the face area.
  • the size of the face area can be determined by the length and height of the rectangular frame circumscribing the face.
  • the preset area size is a preset minimum area size that can be intercepted as the original positive sample image.
  • the standard area size is the area size used to intercept the original positive sample image and the original negative sample image.
  • the size of the preset area may be the same as or smaller than the size of the standard area.
  • step S304 Since the face detection algorithm is used in step S304 to detect valid sample images, all face regions containing faces are selected by using a rectangular frame circumscribed by the face. If the face circumscribes the rectangular frame, the selected face region corresponds to the face If the area size is small, it means that it is a small face image. If the small face image is captured as the original positive sample image for subsequent recognition model training, the original positive sample image captured based on the small face image is not sufficiently clear. As the verification image of the target positive sample image after downsampling and thumbnail interception, it affects the accuracy of the recognition model training.
  • the server uses the face detection algorithm to perform face detection on the valid sample image, and obtains the size of the face area corresponding to each face area, it needs to compare the size of the face area with the preset area size; if the face area is If the size is greater than the preset area size, the original positive sample image and original negative sample image corresponding to the standard area size are intercepted from the valid sample image to ensure the clarity of the intercepted original positive sample image and original negative sample image, thereby ensuring The accuracy of the trained recognition model; if the size of the face area is not larger than the preset area size, there is no need to intercept the original positive sample image and the original negative sample image from the valid sample image.
  • the size of the original positive sample image and the original negative sample image obtained finally can be the same to avoid the recognition model training process
  • the interference caused by the inconsistent size of the sample images ensures the accuracy and training efficiency of the recognition model training.
  • the interception of the original positive sample image and the original negative sample image corresponding to the standard area size from the effective sample image in this embodiment specifically includes: the server uses the OpenCV tool to take screenshots of the face area and the non-face area in the effective sample image. Obtain the original positive sample image and the original negative sample image corresponding to the size of the standard area.
  • the server uses the OpenCV tool to perform screenshot processing on the coordinates of the four corners of the face area and the non-face area in the effective sample image to obtain the corresponding original positive sample image and the original negative sample image through the OpenCV tool
  • the calculation is simple, the calculation efficiency is high, and the performance is relatively stable.
  • S306 Annotate the original positive sample image and the original negative sample image, and obtain the original positive sample image and the original negative sample image carrying the first annotation information.
  • the server After the server obtains the original positive sample image and the original negative sample image, it needs to use a preset labeling rule to label the original positive sample image and the original negative sample image to obtain the original positive sample image and the original Negative sample image.
  • the first annotation information carried by the original positive sample image includes the current image identifier, the source image identifier, the first face identifier, and the first fuzzy identifier.
  • the second annotation information carried by the original negative sample image includes the current image identifier, the source image identifier, the second face identifier, and the first fuzzy identifier.
  • the training sample image is first subjected to blur detection, so as to filter based on the blur degree of the training sample image to obtain a valid sample image whose blur degree reaches the standard, so as to ensure the acquisition based on the valid sample image
  • the clarity of the original positive sample image and the original negative sample image to ensure the recognition accuracy of the trained recognition model.
  • the size of the face area in the valid sample image is greater than the preset area size
  • the original positive sample image and the original negative sample image are intercepted from the valid sample image to ensure that the original positive sample image and the original negative sample image are intercepted
  • the clarity of the training to ensure the accuracy of the training model.
  • step S302 which uses a blur detection algorithm to perform blur detection on the training sample image to obtain the blur degree corresponding to the training sample image, specifically includes the following steps:
  • S401 Use the Laplacian to perform sharpening processing on the training sample image to obtain the sharpened image and the pixel gray value of the sharpened image.
  • the Laplacian operator is a second-order differential operator, which is suitable for improving image blur caused by diffuse reflection of light.
  • the principle is that in the process of taking and recording an image, the light spot diffusely reflects the light to its surrounding area. This kind of diffuse reflection causes the image to be blurred to a certain extent, and the degree of blur is relatively compared with the image taken under normal circumstances. , Is often a constant multiple of the Laplace operator.
  • the Laplacian is used to sharpen the training sample image to obtain the sharpened image, which specifically includes: the Laplacian is used to process the training sample image to obtain a pull that describes the gray-scale mutation. Plus image, and then superimpose the Laplacian image with the training sample image to obtain a sharpened image. After obtaining the sharpened image, the RGB value of each pixel in the sharpened image is obtained, and the RGB value is processed to obtain the pixel gray value corresponding to the sharpened image.
  • S402 Perform variance calculation on the pixel gray value of the sharpened image, obtain a target variance value corresponding to the sharpened image, and determine the target variance value as the blur degree corresponding to the training sample image.
  • the variance calculation is performed on the pixel gray value of the sharpened image to obtain its target variance value, which can be understood as the blur degree of the sharpened image.
  • calculating the variance of the pixel gray value of the sharpened image specifically includes: calculating the pixel gray value of each pixel in the sharpened image minus the square sum of the average gray value of the sharpened image, and then the square The sum is divided by the number of pixels to obtain the target variance value that can reflect the blur of the sharpened image.
  • the smaller the target variance value the closer the pixel gray value of each pixel in the sharpened image is to the average gray value, and the difference between the pixel gray values of each pixel in the sharpened image is not obvious , Making the edges of the image unclear, therefore, the smaller the target variance value, the more blurred the sharpened image.
  • the Laplacian operator is first used to sharpen the original image to obtain a sharpened image with clearer details than the training sample image, thereby improving the clarity of the image. Then, the target variance value of the sharpened image is calculated to reflect the difference between the pixel gray values of each pixel of the sharpened image. Use the target variance value of the sharpened image as the fuzziness of the training sample image, so as to achieve the fuzzy filtering of the training sample image according to the comparison result of the fuzziness and the preset threshold to obtain a clearer training sample image. purpose.
  • step S204 that is, the target positive sample image and the target negative sample image with different face identifiers are input into the MB-FCN model for model training, and the target MB-FCN detector is obtained. Including the following steps:
  • the target positive sample image carrying the first face identifier and the target negative sample image carrying the second face identifier are used as model training data.
  • the distribution ratio of positive and negative samples is the ratio preset by the system for the distribution of positive samples and negative samples.
  • the setting of the ratio of positive and negative samples is to avoid overfitting during model training.
  • the scale range threshold is the scale range that may cause the model training to overfit.
  • the server obtains the corresponding target positive sample image carrying the first face identifier and the target negative sample image carrying the second face identifier based on the positive and negative sample allocation ratio, and uses the target positive sample image and the target negative sample image as the model Training data for subsequent MB-FCN model training.
  • the model training data is a relatively fuzzy target positive sample image containing a human face and a target negative sample image not containing a human face.
  • S502 Divide the model training data into a training set and a test set based on the training test allocation ratio.
  • the training test allocation ratio is the ratio preset by the system for dividing the training set and the test set.
  • the training set is a collection of model training data used to train the MB-FCN model
  • the test set is a collection of model training data used to test whether the accuracy of the MB-FCN model obtained by training meets the standard.
  • S503 Input the model training data in the training set to the MB-FCN model for model training, and obtain the original MB-FCN detector.
  • the server inputs model training data such as the target positive sample image carrying the first face identifier and the target negative sample image carrying the second face identifier in the training set into the MB-FCN model for model training.
  • model training data such as the target positive sample image carrying the first face identifier and the target negative sample image carrying the second face identifier in the training set into the MB-FCN model for model training.
  • the recognition result corresponding to each model training data is calculated with the pre-carried first face identifier or second face identifier, and the loss function of the model is updated based on the error, so as to update the model in the MB-FCN model based on the loss function Parameters to obtain the original MB-FCN detector.
  • the server inputs each model training data in the test set to the original MB-FCN detector for testing, and obtains the test result corresponding to the model training data; if the test result matches the face identifier corresponding to the model training data, then The test of the model training data is successful; on the contrary, if the test result does not match the face identifier corresponding to the model training data, the test of the model test data fails. Then, the server counts the number of model training data corresponding to all successful tests in the test set as the first number, and counts the number corresponding to all model training data in the test set as the second number; based on the quotient of the first number and the second number, obtain The test accuracy of the original MB-FCN detector.
  • test accuracy rate is greater than the accuracy rate threshold, determine the original MB-FCN detector as the target MB-FCN detector.
  • the accuracy threshold is a threshold preset by the system for evaluating whether the accuracy of the model meets the standard. Specifically, the server compares the test accuracy of the original MB-FCN detector with the accuracy threshold; if the test accuracy is greater than the accuracy threshold, the original MB-FCN detector is determined as the target MB-FCN detector; if If the test accuracy is not greater than the accuracy threshold, the model training data needs to be re-acquired for model training.
  • corresponding model training data is obtained based on the positive and negative sample allocation ratio, so as to avoid overfitting of the target MB-FCN detector obtained in the final training to affect the recognition accuracy.
  • Train the original MB-FCN detector based on the model training data in the training set, and then use the model training data in the test set for testing to determine its test accuracy.
  • the test accuracy is greater than the accuracy threshold, the original MB-FCN is detected
  • the target MB-FCN detector is determined as the target MB-FCN detector, so that the target MB-FCN detector can be used to identify a smaller or more obscure face image from the image to be recognized, so as to ensure that the smaller or more obscure face image Recognition accuracy rate.
  • step S206 that is, inputting the target training data to the GAN model for model training, and obtaining the target GAN model, specifically includes the following steps:
  • S601 Input the target positive sample image in the target training data into the generation network formed based on the super-resolution reconstruction technology in the GAN model for processing, and obtain the training generated image.
  • the generation network in the GAN model specifically adopts a generation network formed based on super-resolution reconstruction technology.
  • the generation network can use the DenseNet framework, including a convolutional layer, N DenseBlock (density blocks) and a fully connected layer.
  • Each DenseBlock (density block) is provided with an upsampling subnet that can achieve super-resolution reconstruction and Refine the subnet.
  • the up-sampling subnet can enlarge the resolution of the image several times in a coarse-grained manner, that is, enlarge the resolution of the image several times through the up-sampling operation.
  • the refined subnet is to process the details of several times the resolution through fine-grained processing, in which the deconvolution method is specifically used for processing in the refined subnet. Since in the traditional convolutional neural network, if there is an L layer, it needs to have L connections. In the DenseNet framework, there will be L(L+1)/2 connections. That is, in the DenseNet framework, the input of each layer comes from the output of all the previous layers. This connection method makes the network narrower, the parameters are smaller, and the transfer of features and gradients is more effective, and the network is easier to train.
  • the server inputs the fuzzy target positive sample images with lower resolution in the target training data into the generation network of the GAN model.
  • the super-resolution reconstruction technology is used to correct the target positive samples.
  • the sample image is reconstructed, that is, iterative processing is performed through the up-sampling subnet and the refined subnet in the N DenseBlock (density block) to obtain the reconstructed super-resolution training generated image, so as to generate based on the training
  • the image updates the model parameters of the GAN model.
  • S602 Input the original positive sample image and the training generated image in the target training data into the discriminant network in the GAN model for processing, and obtain the target loss value.
  • the server inputs the clearer original positive sample image in a set of target training data and the training generated image formed by super-resolution reconstruction based on the fuzzy target positive sample image into the discriminant network in the GAN model for processing. , Using the original positive sample image as a reference for the reconstructed super-resolution training generated image to obtain the target loss value between the original positive sample image and the training generated image.
  • the preset loss value is the minimum loss value preset by the system for evaluating model convergence.
  • the server compares the target loss value with the preset loss value; if the target loss value is greater than the preset loss value, it means that the GAN model has not converged to the preset standard.
  • the loss function needs to be updated based on the target loss value .
  • the model parameters of the generating network include the weights and offsets of the generating network
  • the model parameters in the discriminating network include the weights and offsets of the discriminating network.
  • a relatively fuzzy target positive sample image is input to a generation network formed based on super-resolution reconstruction technology for processing, and a clearer super-resolution training generated image is obtained;
  • the original positive sample image corresponding to the image generated by the training and the same face is input to the discriminant network for processing, and the model parameters are updated according to the target loss value until the model converges, and the target GAN model is obtained to make the super-resolution reconstructed by the target GAN model
  • the accuracy of the rate image is higher and the efficiency is faster.
  • the original negative sample image and the target negative sample image that match the current image identifier in the first annotation information and the source image identifier in the second annotation information may be used as a set of target training data. It is understandable that the process of obtaining the target training data corresponding to the negative sample image is the same as the process of obtaining the target training data corresponding to the positive sample image, and to avoid repetition, it will not be repeated here.
  • the target negative sample image in the target training data can also be input into the generation network formed based on the super-resolution reconstruction technology in the GAN model for processing, and the training generated image is obtained.
  • the discriminant network of the GAN model uses two parallel fully connected layers (Fully Connected Layer, referred to as FC layer).
  • the first fully connected layer is used to distinguish natural and real images (ie original positive sample images or original negative images).
  • Sample image) and the generated super-resolution image ie training generated image
  • the GAN model can not only reconstruct the super-resolution image, but also recognize the face in the image, ensuring the efficiency of the target GAN model for super-resolution image reconstruction and face recognition.
  • the image recognition method provided by the embodiment of the present application can be applied to the application environment shown in FIG. 1.
  • the recognition model training method is applied to an image recognition system.
  • the image recognition system includes a client and a server as shown in FIG. 1.
  • the client and the server communicate through a network and are used for the recognition obtained by training in the foregoing embodiment.
  • the model recognizes the image to be processed, so as to accurately identify a small or fuzzy face image, so as to determine whether the face image is an image of the target face.
  • this embodiment provides an image recognition method.
  • the image recognition method includes the following steps:
  • S701 Obtain an image to be recognized, and use a target MB-FCN detector to perform face detection and image interception on the image to be recognized, to obtain a target face image.
  • the target MB-FCN detector is the detector obtained by training based on the target positive sample image and the target negative sample image in the above embodiment.
  • the image to be recognized is an image that needs to be recognized to determine whether it is the target person.
  • the target person is a person who needs to be searched, including but not limited to criminal suspects, missing persons, or other persons who need to be searched.
  • the target face image is a face image obtained by using a target MB-FCN detector to detect the image to be recognized to detect at least one face area in the image to be recognized, and using a screenshot tool to take a screenshot of the at least one face area. Since the MB-FCN detector adopts the fuzzy target positive sample image and the fuzzy target negative sample image for model training, the detector is used to identify whether it is a face image, it can identify the fuzzy image to be recognized To determine whether the fuzzy image to be recognized contains the face area corresponding to the face, so that the server can use a screenshot tool to intercept the target face image containing the face from the face area.
  • S702 Use the generation network formed based on the super-resolution reconstruction technology in the target GAN model to generate the target face image to obtain the target generated image.
  • the target GAN model is the model obtained by using target training data for model training in the above-mentioned embodiment.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology. High-speed image reconstruction forms a clearer super-resolution image, and the reconstruction process is more accurate and efficient.
  • the server inputs the target face image into the target GAN model, and uses the generation network based on the super-resolution reconstruction technology in the target GAN model to generate the target face image to obtain the super-resolution target generation Image, in order to use the target to generate an image for subsequent facial image recognition to improve recognition accuracy and recognition efficiency.
  • the target generated image is to input a lower resolution target face image into the target GAN model, and use a generation network based on super-resolution reconstruction technology to reconstruct a clearer super-resolution image.
  • S703 Perform feature similarity calculation on the target generated image and the target person image in the target image library to obtain the feature similarity.
  • the target image database is a database corresponding to the image of the target person.
  • the target person image refers to the image of the target person to be tracked.
  • the server uses a face similarity detection algorithm to calculate the feature similarity between the target generated image and each target person image in the target image library, and obtain the calculated feature similarity.
  • the server can perform feature extraction on the target generated image and each target person image, obtain the corresponding generated image feature and the person image feature, and then use feature distance algorithms (including but not limited to Euclidean distance algorithm or cosine similarity algorithm) to generate The image feature and the person image feature are calculated to obtain the corresponding feature similarity.
  • feature distance algorithms including but not limited to Euclidean distance algorithm or cosine similarity algorithm
  • the similarity threshold is a threshold preset by the system for evaluating the similarity of the same person. After obtaining the feature similarity between the target generated image and the image of each target person, the server compares the feature similarity with the similarity threshold. If the feature similarity is greater than the similarity threshold, it means that the target face image in the image to be recognized and the target person image in the target image library are more likely to be the same person. Therefore, the person corresponding to the target person image can be obtained as the target person The test results. Correspondingly, if the feature similarity is not greater than the similarity threshold, it means that the target face image is not an image of the target person.
  • the image recognition method provided in this embodiment can be applied to criminal suspect tracking, missing person search, or other scenes where a person search is required.
  • This method can first use the target MB-FCN model for face detection and image interception from the low-resolution images to be processed randomly shot by monitoring equipment or individuals, so as to obtain the target face image containing the face, which is helpful Reduce the amount of data for subsequent image reconstruction and image recognition, and improve recognition efficiency.
  • the generation network based on the super-resolution reconstruction technology of the target GAN model is used to process the target face image to obtain a clearer super-resolution target generated image, so that subsequent recognition based on the target generated image , The recognition accuracy is higher and the recognition efficiency is faster.
  • the feature similarity and similarity threshold calculated by the target generated image and the target person image are compared to obtain the detection result of whether the person corresponding to the target face image is the target person, so as to realize the fast tracking of the target person.
  • a recognition model training device is provided, and the recognition model training device corresponds to the recognition model training method in the foregoing embodiment in a one-to-one correspondence.
  • the recognition model training device includes an original sample image acquisition module 801, a sample thumbnail acquisition module 802, a target sample image acquisition module 803, a target detector acquisition module 804, a target training data acquisition module 805, and a target GAN model Obtaining module 806.
  • the detailed description of each functional module is as follows:
  • the original sample image acquisition module 801 is configured to acquire the original positive sample image and the original negative sample image carrying the first annotation information.
  • the first annotation information includes the current image identifier, the face identifier and the first fuzzy identifier.
  • the sample thumbnail acquisition module 802 is configured to perform down-sampling processing on the original positive sample image and the original negative sample image, and obtain corresponding positive sample thumbnails and negative sample thumbnails respectively.
  • the target sample image acquisition module 803 is configured to use a screenshot tool to perform screenshot processing on the positive sample thumbnail and the negative sample thumbnail to obtain the target positive sample image and the target negative sample image carrying the second annotation information, the second annotation information includes the source image Logo, face logo and second fuzzy logo.
  • the target detector acquisition module 804 is configured to input target positive sample images and target negative sample images with different face identifiers into the MB-FCN model for model training, and obtain the target MB-FCN detector.
  • the target training data acquisition module 805 is configured to use the original positive sample image and target positive sample image matching the current image identifier in the first annotation information with the source image identifier in the second annotation information as a set of target training data.
  • the target GAN model acquisition module 806 is configured to input target training data into the GAN model for model training, and acquire the target GAN model.
  • the generation network in the target GAN model is a generation network formed based on super-resolution reconstruction technology.
  • the original sample image acquisition module 801 includes:
  • the training sample image obtaining unit is used to obtain training sample images including human faces from the image database.
  • the fuzzy degree acquisition unit is used to perform fuzzy detection on the training sample image by using a fuzzy detection algorithm to obtain the fuzzy degree corresponding to the training sample.
  • the effective sample image determining unit is used to determine the effective sample image based on the blur degree of the training sample image.
  • the face area size obtaining unit is used to detect the face of the valid sample image using a face detection algorithm to obtain the face area size.
  • the original sample image interception unit is used to intercept the original positive sample image and the original negative sample image corresponding to the standard area size from the valid sample image if the size of the face area is larger than the preset area size.
  • the original sample image labeling unit is used to label the original positive sample image and the original negative sample image, and obtain the original positive sample image and the original negative sample image carrying the first label information.
  • the ambiguity acquisition unit includes:
  • the gray gray value obtaining subunit is used to sharpen the training sample image by using the Laplacian operator to obtain the sharpened image and the pixel gray value of the sharpened image.
  • the fuzzy degree determination subunit is used to calculate the variance of the pixel gray value of the sharpened image, obtain the target variance value corresponding to the sharpened image, and determine the target variance value as the fuzzy degree corresponding to the training sample image.
  • the target detector acquisition module 804 includes:
  • the model training data acquisition unit is configured to use the target positive sample image carrying the first face identifier and the target negative sample image carrying the second face identifier as model training data based on the positive and negative sample allocation ratio.
  • the model training data dividing unit is used to divide the model training data into a training set and a test set based on the training test allocation ratio.
  • the original detector acquisition unit is used to input the model training data in the training set to the MB-FCN model for model training, and obtain the original MB-FCN detector.
  • the test accuracy rate acquisition unit is used to test the original MB-FCN detector with the model training data in the test set to obtain the test accuracy rate.
  • the target detector determining unit is configured to determine the original MB-FCN detector as the target MB-FCN detector if the test accuracy rate is greater than the accuracy rate threshold.
  • the target GAN model acquisition module 806 includes:
  • the training generated image acquisition unit is used to input the target positive sample image in the target training data into the generation network formed based on the super-resolution reconstruction technology in the GAN model for processing, and obtain the training generated image.
  • the target loss value obtaining unit is used to input the original positive sample image and the training generated image in the target training data into the discriminant network in the GAN model for processing, and obtain the target loss value.
  • the model parameter update unit is used to update the loss function based on the target loss value if the target loss value is greater than the preset loss value, update the model parameters in the generating network and the discrimination network based on the updated loss function, and repeat the execution of the target training data
  • the target positive sample image is input into the generation network based on super-resolution reconstruction technology in the GAN model for processing, and the training generated image is obtained.
  • the target GAN model acquisition unit is used to if the target loss value is not greater than the preset loss value, the model converges to acquire the target GAN model.
  • an image recognition device is provided, and the image recognition device corresponds to the image recognition method in the above-mentioned embodiment one-to-one.
  • the image recognition device includes a target face image acquisition module 901, a target generated image acquisition module 902, a feature similarity acquisition module 903, and a detection result acquisition module 904.
  • the detailed description of each functional module is as follows:
  • the target face image acquisition module 901 is configured to acquire the image to be recognized, and use the above-mentioned target MB-FCN detector to perform face detection and image interception on the image to be recognized to obtain the target face image.
  • the target generation image acquisition module 902 is configured to use the generation network formed based on the super-resolution reconstruction technology in the target GAN model to generate the target face image to obtain the target generation image.
  • the feature similarity acquisition module 903 is configured to perform feature similarity calculations between the target generated image and the target person image in the target image library to obtain the feature similarity.
  • the detection result obtaining module 904 is configured to obtain the detection result of determining the person corresponding to the target face image as the target person if the feature similarity is greater than the similarity threshold.
  • each module in the above-mentioned recognition model training device or image recognition device may be implemented in whole or in part by software, hardware, and a combination thereof.
  • the foregoing modules may be embedded in the form of hardware 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 can call and execute the operations corresponding to the foregoing modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 10.
  • the computer equipment includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide calculation 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 the non-volatile storage medium.
  • the database of the computer equipment is used to store data used or generated in the process of executing the above-mentioned recognition model training method or image recognition method.
  • 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 the processor to realize a recognition model training method or an image recognition method.
  • a computer device including a memory, a processor, and computer-readable instructions stored in the memory and running on the processor.
  • the processor executes the computer-readable instructions to realize the identification in the above-mentioned embodiments.
  • the steps of the model training method are, for example, the steps shown in Figs. 2-6.
  • the functions of the modules/units in this embodiment of the recognition model training device are realized, such as the original sample image acquisition module 801, the sample thumbnail acquisition module 802, and the target sample image shown in FIG.
  • a computer device including a memory, a processor, and computer readable instructions stored in the memory and running on the processor.
  • the processor executes the computer readable instructions to implement the image in the above embodiment.
  • the steps of the identification method are as shown in Figure 7.
  • the processor implements the functions of the modules/units in this embodiment of the image recognition device when the processor executes the computer-readable instructions, for example, the target face image acquisition module 901, the target generated image acquisition module 902, and similar features shown in FIG.
  • one or more readable storage media storing computer readable instructions are provided.
  • the computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by one or more processors. When executed, the one or more processors are executed to implement the steps of the recognition model training method in the foregoing embodiment, for example, the steps shown in FIGS. 2 to 6. Or, when the computer-readable instruction is executed by the processor, the function of each module/unit in the embodiment of the recognition model training device is realized, for example, the original sample image acquisition module 801, the sample thumbnail acquisition module 802, The functions of the target sample image acquisition module 803, the target detector acquisition module 804, the target training data acquisition module 805, and the target GAN model acquisition module 806.
  • the readable storage medium in this embodiment includes a nonvolatile readable storage medium and a volatile readable storage medium.
  • one or more readable storage media storing computer readable instructions are provided.
  • the computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by one or more processors.
  • the steps of the image recognition method in the foregoing embodiment are realized when the one or more processors are executed, and the steps are shown in FIG. 7.
  • the computer-readable instruction is executed by the processor, the function of each module/unit in this embodiment of the image recognition device is realized, for example, the target face image acquisition module 901 and the target generation image acquisition module 902 shown in FIG. 9
  • the readable storage medium in this embodiment includes a nonvolatile readable storage medium and a volatile readable storage medium.
  • 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 may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开一种识别模型训练方法、图像识别方法、装置、设备及介质。该识别模型训练方法包括:获取携带第一标注信息的原始正样本图像和原始负样本图像;对原始正样本图像和原始负样本图像进行下采样处理,并采用截图工具进行截图处理,获取携带第二标注信息的目标正样本图像和目标负样本图像;将目标正样本图像和目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;将第一标注信息中的当前图像标识与第二标注信息中的来源图像标识相匹配的原始正样本图像和目标正样本图像作为一组目标训练数据;将目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型。该目标GAN模型可有助于提高模糊图像人脸识别的准确率。

Description

识别模型训练方法、图像识别方法、装置、设备及介质
本申请以2019年8月1日提交的申请号为201910706615.9,名称为“识别模型训练方法、图像识别方法、装置、设备及介质”的中国发明申请为基础,并要求其优先权。
技术领域
本申请涉及图像识别技术领域,尤其涉及一种识别模型训练方法、图像识别方法、装置、设备及介质。
背景技术
随着人脸检测技术的快速发展,使得人脸检测器识别大中型人脸图像的准确性快速提高。而人脸检测技术在实际应用过程中,通常需要对较小或者较模糊的人脸图像进行识别,以识别人脸图像对应的人员身份,但传统人脸检测器对较小或较模糊的人脸图像的识别准确率较低,影响实际应用的效果。例如,在犯罪嫌疑人追踪、失踪人员搜索或者其他需要进行人员搜索的场景下,可采用人脸检测器对监控设备拍摄的监控图像进行识别,以确定犯罪嫌疑人、失踪人员或者其他目标人员的行踪,由于监控设备的拍摄像素或者拍摄角度等因素的限制,使得监控图像中的犯罪嫌疑人、失踪人员或者其他目标人员的人脸较小或者较模糊,采用传统人脸检测器进行识别时,无法准确识别出犯罪嫌疑人、失踪人员或者其他目标人员的身份,从而影响目标人员追踪的效率。
发明内容
本申请实施例提供一种识别模型训练方法、图像识别方法、装置、设备及介质,以解决当前人脸检测器对较小或者较模糊的人脸图像进行识别时准确率较低的问题。
一种识别模型训练方法,包括:
获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
一种识别模型训练装置,包括:
原始样本图像获取模块,用于获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
样本缩略图获取模块,用于对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
目标样本图像获取模块,用于采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
目标检测器获取模块,用于将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
目标训练数据获取模块,用于将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
目标GAN模型获取模块,用于将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质存储有计算机可读指令,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如下步骤:
获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
一种图像识别方法,包括:
获取待识别图像,采用上述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
一种图像识别装置,包括:
目标人脸图像获取模块,用于获取待识别图像,采用上述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
目标生成图像获取模块,用于采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
特征相似度获取模块,用于将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
检测结果获取模块,用于若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
获取待识别图像,采用上述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质存储有计算机可读指令,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如下步骤:
获取待识别图像,采用上述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
本申请的一个或多个实施例的细节在下面的附图及描述中提出。本申请的其他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的技术方案,下面将对本申请实施例的描述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性的前提下,还可以根据这些附图获得其他的附图。
图1是本申请一实施例中识别模型训练方法或者图像识别方法的一应用环境示意图;图2是本申请一实施例中识别模型训练方法的一流程图;
图3是本申请一实施例中识别模型训练方法的另一流程图;
图4是本申请一实施例中识别模型训练方法的另一流程图;
图5是本申请一实施例中识别模型训练方法的另一流程图;
图6是本申请一实施例中识别模型训练方法的另一流程图;
图7是本申请一实施例中图像识别方法的另一流程图;
图8是本申请一实施例中识别模型训练装置的一示意图;
图9是本申请一实施例中图像识别装置的一示意图;
图10是本申请一实施例中计算机设备的一示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然, 所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
本申请实施例提供的识别模型训练方法,该识别模型训练方法可应用如图1所示的应用环境中。具体地,该识别模型训练方法应用在识别模型训练系统中,该识别模型训练系统包括如图1所示的客户端和服务器,客户端与服务器通过网络进行通信,用于训练可准确识别较小或者较模糊的人脸图像的识别模型,以扩大识别模型在图像识别领域的应用。其中,客户端又称为用户端,是指与服务器相对应,为客户提供本地服务的程序。客户端可安装在但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备上。服务器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。
在一实施例中,如图2所示,提供一种识别模型训练方法,以该识别模型训练方法应用在图1所示的服务器为例,该识别模型训练方法包括如下步骤:
S201:获取携带第一标注信息的原始正样本图像和原始负样本图像,第一标注信息包括当前图像标识、人脸标识和第一模糊标识。
其中,原始正样本图像是用于进行模型训练的较清晰的包含人脸的图像。原始负样本图像是用于进行模型训练的较清晰的不包含人脸的图像。
其中,第一标注信息是原始正样本图像和原始负样本图像对应的标注信息,是用于标识原始正样本图像和原始负样本图像的图像属性的信息。本实施例中,服务器在获取原始正样本图像和原始负样本图像之后,需采用预设标注规则对原始正样本图像和原始负样本图像进行标注,以使每一原始正样本图像和原始负样本图像均携带相应的第一标注信息。该预设标注规则是用于标注图像对应的当前图像标识、来源图像标识、人脸标识和模糊标识的规则。
其中,当前图像标识是用于识别所要标注的图像的唯一标识,可以依据图像获取的序号确定,如T001表示第1张图像,T002代表第2张图像。来源图像标识是用于识别所要标注的图像对应的来源图像的标识,可以理解为截取原始正样本图像或者原始负样本图像的来源图像对应的标识,如P001表示第1张来源图像,P002表示第2张来源图像。人脸标识是指用于区别所要标注的图像是否包含人脸的标识,该人脸标识包括用于表示图像包含人脸的第一人脸标识和用于表示图像不包含人脸的第二人脸标识。例如,第一人脸标识可采用1来表示,第二人脸标识可采用0来表示。模糊标识是用于区别所要标注的图像是否为模糊图像的标识,该模糊标识包括用于表示图像不模糊的第一模糊标识和用于表示图像较模糊的第二模糊标识,即第一模糊标识是用于表示图像较清晰的标识。例如,第一模糊标识可采用0来表示,第二模糊标识可采用1来表示。
由于原始正样本图像是用于模型训练的较清晰的包含人脸的图像,因此,该原始正样本图像所携带的第一标注信息包含当前图像标识、来源图像标识、第一人脸标识和第一模糊标识。如{T001,P001,1,0}表示当前图像标识为T001的图像是从来源标识为P001的来源图像中截取的包含人脸的较清晰的图像。
由于原始负样本图像是用于模型训练的较清晰的不包含人脸的图像,因此,该原始负样本图像所携带的第一标注信息包含当前图像标识、来源图像标识、第二人脸标识和第一模糊标识。如{T002,P002,0,0}表示当前图像标识为T002的图像是从来源标识为P002的来源图像中截取的不包含人脸的较清晰的图像。
进一步地,为了避免后续识别模型训练过程中因样本图像的大小不一致而存在干扰,影响最终形成的识别模型训练的准确率和效率,可限定原始正样本图像和原始负样本图像的大小一致。例如,从同一较清晰的图像中截取图像大小相同的包含人脸的原始正样本图像和不包括人脸的原始负样本图像。
S202:对原始正样本图像和原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图。
其中,下采样(Subsampled),又称为降采样(dowmsampled),是用于缩小图像,以使图像符合显示区域的大小或者用于生成图像的缩略图的过程。在下采样处理过程中,若一幅图像I的尺寸为M*N,对其进行s倍下采样处理,即可得到(M/s)*(N/s)尺寸的分辨率图像,此时s是M和N的公约数。对于矩阵形式的图像,需要将该图像s*s窗口内的图像变成一个像素,这个像素点的值就是窗口内所有像 素的均值。本实施例中下采样处理过程中可以采用但不限于最近邻插值、双线性插值、均值插值和中值插值等方法。
具体地,服务器对原始正样本图像和原始负样本图像进行不同程度的下采样处理,以将原始正样本图像和原始负样本图像缩小成不同缩放比例的正样本缩略图和负样本缩略图,通过降低其分辨率达到模糊化处理的目的。例如,可将原始正样本图像和原始负样本图像分别缩小成20%的正样本缩略图和负样本缩略图。
S203:采用截图工具对正样本缩略图和负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和目标负样本图像,第二标注信息包括来源图像标识、人脸标识和第二模糊标识。
具体地,服务器先采用截图工具对正样本缩略图和负样本缩略图进行截图处理,获取相应的目标正样本图像和目标负样本图像,以使获取的目标正样本图像和目标负样本图像的图像较小或较模糊,用于作为训练可识别较小或者较模糊图像的识别模型的训练数据。再对目标正样本图像和目标负样本图像进行标注,以获取携带第二标注信息的目标正样本图像和目标负样本图像。该截图工具包括但不限于本实施例所采用的OpenCV截图工具,OpenCV(Open Source Computer Vision Library,开源计算机视觉库)是一个基于BSD许可(开源)发行的跨平台计算机视觉库,可以运行在Linux、Windows、Android和Mac OS操作系统上。它轻量级而且高效——由一系列C函数和少量C++类构成,同时提供了Python、Ruby、MATLAB等语言的接口,实现了图像处理和计算机视觉方面的很多通用算法。
本实施例中,目标正样本图像所携带的第二标注信息包括当前图像标识、来源图像标识、第一人脸标识和第二模糊标识。例如,基于第一标注信息为{T001,P001,1,0}的原始正样本图像所获取的目标正样本图像的第二标注信息为(M001,T001,1,1},即当前图像标识为M001的图像是从来源标识为T001的来源图像中获取的包含人脸的较模糊的图像,此处的来源标识为T001的来源图像是指当前图像标识为T001的原始正样本图像。可以理解地,原始正样本图像的第一标注信息中的当前图像标识T001与目标正样本图像的第二标注信息中的来源图像标识T001相匹配,说明目标正样本图像是基于较清晰的原始正样本图像进行模糊处理之后获取的图像。
相应地,目标负样本图像所携带的第二标注信息包括当前图像标识、来源图像标识、第二人脸标识和第二模糊标识。例如,基于第一标注信息为{T002,P002,0,0}的原始负样本图像所获取的目标负样本图像的第二标注信息为{M002,T002,0,1},即当前图像标识为M002的图像是从来源标识为T002的来源图像中获取的包含人脸的较模糊的图像,此处的来源标识为T002的来源图像是指当前图像标识为T002的原始负样本图像。可以理解地,原始负样本图像的第一标注信息中的当前图像标识T002与目标负样本图像的第二标注信息中的来源图像标识T002相匹配,说明目标负样本图像是基于较清晰的原始负样本图像进行模糊处理之后获取的图像。
S204:将人脸标识不相同的目标正样本图像和目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器。
其中,目标MB-FCN(Multi-branch fully convolutional network,多分支全卷积网络)检测器是用于识别较模糊的图像是否为人脸图像的检测器,以便后续利用该目标MB-FCN检测器从待识别图像中识别较小或较模糊的人脸。由于目标MB-FCN检测器是用于识别较模糊的图像是否为人脸图像的检测器,为了训练该检测器,需要采用较模糊的包含人脸和不包含人脸的图像进行训练,所以需将携带第一人脸标识和第二模糊标识的目标正样本图像(即包含人脸的较模糊的图像)作为目标MB-FCN检测器的正样本,将携带第二人脸标识和第二模糊标识的目标负样本图像(即不包含人脸的较模糊的图像)作为目标MB-FCN检测器的负样本。
与常用CNN分类模型所使用的网络在卷积层之后连接全连接层,将原来的二维的矩阵压缩成维的输出向量(即分类标签)相比,FCN(fully convolutional network,全卷积网络)可以接受任意尺寸的输入图像,采用反卷积层对最后一个卷积层的feature map进行上采样,使它恢复到输入图像相同的尺寸,从而可以对每个像素都产生了一个预测,同时保留了原始输入图像中的空间信息,最后在上采样的特征图上进行逐像素分类。FCN将CNN的全连接层变成卷积层,使得整个网络变成了只有卷积层和池化层的网络,于是网络就称之为全卷积网络。FCN一般是用来对图像进行语义分割的,于是就需要对图 像上的各个像素进行分类,这就需要一个上采样将最后得到的输出上采样到原图的大小,从而支持语义分割,使得基于FCN检测器相比于CNN检测器,其识别准确率较高且识别效率较快。
为了进一步提高检测器的识别准确率和识别效率,本实施例中采用MB-FCN(Multi-branch fully convolutional network,多分支全卷积网络)模型代替FCN模型进行训练,以获取目标MB-FCN检测器。MB-FCN可以使用不同尺度的特征提取器提取更丰富的特征,相比FCN提取的特征的表征效果更好,有助于模型的识别准确率较高。而且,MB-FCN考虑设计过程中的效率和有效性,只需一次通过骨干网络即可处理所有比例范围的面部特征,与FCN需经过多次传递的方法相比,MB-FCN节省了计算,因此更有效,使得模型的识别效率较快。
S205:将第一标注信息中的当前图像标识与第二标注信息中的来源图像标识相匹配的原始正样本图像和目标正样本图像作为一组目标训练数据。
目标训练数据是用于训练GAN模型的训练数据。本实施例中,原始正样本图像所携带的第一标注信息包括当前图像标识、来源图像标识、第一人脸标识和第一模糊标识,即该原始正样本图像是较清晰的包含人脸的图像。而目标正样本图像所携带的第二标注信息包括当前图像标识、来源图像标识、第一人脸标识和第二模糊标识,即该目标正样本图像是较模糊的包含人脸的图像。若原始正样本图像所携带的第一标注信息中的当前图像标识与目标正样本图像所携带的第二标注信息中的来源图像标识相匹配,则说明目标正样本图像是基于较清晰的原始正样本图像进行模糊处理之后获取的图像。
本实施例中将第一标注信息中的当前图像标识与第二标注信息中的来源图像标识相匹配的原始正样本图像和目标正样本图像作为一组目标训练数据,使得每一组目标训练数据中的原始正样本图像和目标正样本图像是基于同一人脸图像的清晰度不同的两张图像,以便后续GAN模型训练过程中,可以利用较清晰的原始正样本图像对较模糊的目标正样本图像进行验证,以更新GAN模型中的模型参数。
S206:将目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
其中,生成式对抗网络(Generative Adversarial Networks,简称为GAN)是一种深度学习模型,是一种无监督学习的网络。生成对抗网络(GAN)框架包括两个子网络,一个是生成网络(Generative),另一个是判别网络(Discriminative),通过生成网络和判别网络互相博弈学习产生相当好的输出。本实施例中,在GAN模型的生成网络中,使用超分辨率重构(Super Resolution Reconstruction,以下简称SPN)技术,将小面目采样到精细尺度以找到更微小的面;相应地,在判别网络中,采用损失函数,通过该损失函数强制判别网络区分图像中是否包含人脸,即区分人脸或非人脸。
具体地,服务器将每一组目标训练数据输入到GAN模型进行模型训练,即将相同内容的较清晰的原始正样本数据和较模糊的目标正样本图像输入到GAN模型进行训练。在GAN模型训练过程中,采用基于超分辨率重构技术形成的生成网络对较模糊的目标正样本图像进行重构,再将所生成的图像与原始正样本图像输入到GAN的判别网络进行处理,利用判别网络的输出值进行迭代更新,从而更新GAN模型的生成网络和判别网络对应的模型参数,从而获取目标GAN模型,以保证目标GAN模型将较小或较模糊的低分辨率图像,重构成较清晰的超分辨率图像,以便利用超分辨率图像进行人脸识别,从而保障人脸识别的准确率和效率。
在一具体实施方式中,服务器统计目标训练数据对应的目标数量,若目标数量大于数量阈值,则将目标训练数据划分至少两个目标批次的目标训练数据,使得每一目标批次的目标训练数据的目标数量小于数量阈值。然后,再将目标训练数据分批输入GAN模型进行模型训练,并采用小批梯度下降算法更新GAN模型中的模型参数,以获取目标GAN模型。这种将目标训练数据分批次进行GAN模型训练的过程,可以避免目标训练数据的样本集过大,且系统GPU的显存有限时,影响GAN模型训练的效率和准确率。
其中,小批量梯度下降(Mini-batch Gradient Descent,简称MBGD)算法是在更新模型参数时,将训练过程中产生的误差按预设批次分批累加,得到若干批次对应的累加误差,并采用该若干批次对应的累加误差进行模型参数更新的处理方法。在对神经网络中的权值和偏移量等模型参数进行调整时需要求出损失函数的极小值,本实施例中可采用后向传播算法计算损失函数的极小值。后向传播算法(Back Propagation,简称BP算法)是神经网络学习中一种训练与学习方法,用来调整神经网络中各个节点之 间的权值和偏移量。
本实施例所提供的识别模型训练方法中,对携带第一标注信息的原始正样本图像和原始负样本图像进行下采样处理和图像截取操作,以获取携带第二标注信息的目标正样本图像和目标负样本图像,以使原始正样本图像与目标正样本图像之间具有关联关系,有助于保障后续模型训练的实现。采用人脸标识不相同且模糊标识均为第二模糊标识的目标正样本图像和目标负样本图像进行MB-FCN模型训练,以使训练所得的目标MB-FCN模型可快速且准确地较模糊的图像中的人脸图像。将第一标注信息中的当前图像标识与第二标注信息中的来源图像标识相匹配的原始正样本图像和目标正样本图像作为一组目标训练数据,利用该目标训练数据进行GAN模型训练,以更新基于超分辨率重构技术形成的生成网络和判别网络中的模型参数,使得生成的目标GAN模型可以将较小或较模糊的低分辨率图像,重构成较清晰的超分辨率图像,以便利用超分辨率图像进行人脸识别,从而保障人脸识别的准确率和效率。
在一实施例中,如图3所示,步骤S201,即获取携带第一标注信息的原始正样本图像和原始负样本图像,具体包括如下步骤:
S301:从图像数据库中获取包括人脸的训练样本图像。
其中,图像数据库是用于存储包含人脸的训练样本图像的数据库。该图像数据库可以是系统本地数据库,即与服务器相连的本地数据库,以使服务器可快速获取到相应的包含人脸的训练样本图像。或者,该图像数据库还可以是网络图像库,即互联网上用于存储图像信息的数据库,如百度图像库。训练样本图像是用于进行模型训练的未经处理的图像。
在一具体实施方式中,若图像数据库为网络图像库时,服务器需采用爬虫工具从网络图像库中抓取包含人脸的训练样本图像。其中,爬虫工具是指网络爬虫(又被称为网页蜘蛛或网络机器人),是一种按照一定的规则,自动地抓取万维网信息的程序或者脚本。该爬虫工具包括但不限于Python爬虫工具。即爬虫工具是用于从互联网中爬取满足特定条件的图像的工具,该特定条件可以设定为包含人脸。
具体地,服务器采用爬虫工具执行爬虫文件,从网络图像库中爬取符合爬虫文件设置的数据爬取条件的训练样本图像。其中,该爬虫文件包括但不限于目标URL和搜索关键词这两个数据爬取条件。该目标URL是该爬虫文件中用于限定所需爬取图像所在的网络图像库的目标网站的URL,该URL(Uniform Resource Locator的简称,即统一资源定位符)是对可以从互联网上得到的资源的位置和访问方法的一种简洁的表示,是互联网上标准资源的地址,即网络图像库的URL地址。搜索关键词是指该爬虫文件中用于限定该爬虫文件所需爬取图像的共同特性的关键词,具体可以设置为人脸。
S302:采用模糊检测算法对训练样本图像进行模糊检测,获取训练样本对应的模糊度。
由于识别模型训练过程中,需采用较清晰图像对识别模型训练过程中的模糊图像进行校验,而原始正样本图像和原始负样本图像均为较清晰的样本图像,因此,需要预先检测训练样本图像的模糊度,以便基于基于该模糊度对训练样本图像进行过滤,以获取可进行模型训练的有效样本图像。可以理解地,模糊检测算法是用于检测图像的模糊度的算法,该模糊检测算法可以采用现有技术中常用的检测算法。
S303:基于训练样本图像的模糊度,确定有效样本图像。
具体地,服务器基于训练样本图像的模糊度与系统预先设置的模糊阈值进行比较;若训练样本图像的模糊度小于模糊阈值,则将该训练样本图像确定为有效样本图像;若训练样本图像的模糊度不小于模糊阈值,则不将训练样本图像作为有效样本图像,从而达到过滤较模糊的训练样本图像的目的,以使最终确定的有效样本图像的清晰度均达到预先设置的标准,有助于保障后续识别模型训练的准确率。其中,模糊阈值是预先设置的用于评估图像的模糊度是否达到预设标准的阈值。有效样本图像是模糊度小于模糊阈值的训练样本图像,是可以作为后续图像截取的来源图像。
S304:采用人脸检测算法对有效样本图像进行人脸检测,获取人脸区域大小。
其中,人脸检测算法是用于识别图像中是否包含人脸的算法,本实施例中可以采用行业通用的人脸检测算法。具体地,服务器采用人脸检测算法对有效样本图像进行人脸检测,以识别人脸外接矩形框中的人脸区域,并根据该人脸外接矩形框确定人脸区域大小。该人脸外接矩形框是采用人脸检测算法检测出图像中包含人脸区域的矩形框。可以理解地,该人脸外接矩形框是虚拟人脸框,用于框选出人脸检测算法识别出的人脸区域,该人脸外接矩形框的大小随着人脸区域的大小自主调整。可以理解地,人脸 区域大小可以由人脸外接矩形框的长和高确定。
S305:若人脸区域大小大于预设区域大小,则从有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像。
其中,预设区域大小是预先设置的可以截取作为原始正样本图像的最小的区域大小。标准区域大小是用于截取原始正样本图像和原始负样本图像的的区域大小。该预设区域大小可以与标准区域大小相同,也可以小于标准区域大小。
由于步骤S304中采用人脸检测算法对有效样本图像进行检测时,采用人脸外接矩形框选出所有包含人脸的人脸区域,若人脸外接矩形框选出的人脸区域对应的人脸区域大小较小,则说明其为小脸图像,若将该小脸图像截取作为原始正样本图像进行后续的识别模型训练时,由于基于小脸图像截取的原始正样本图像的清晰度不够,无法作为经过下采样和缩略图截取之后的目标正样本图像的验证图像,影响识别模型训练的准确率。因此,服务器在采用人脸检测算法对有效样本图像进行人脸检测,获取每一人脸区域对应的人脸区域大小之后,需将该人脸区域大小与预设区域大小进行比较;若人脸区域大小大于预设区域大小,则从有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像,以保证所截取出的原始正样本图像和原始负样本图像的清晰度,从而确保训练所得的识别模型的准确率;若人脸区域大小不大于预设区域大小,则无需有效样本图像中截取原始正样本图像和原始负样本图像。
可以理解地,从有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像,可使最终获取的原始正样本图像和原始负样本图像的大小一致,以避免识别模型训练过程中因样本图像的大小不一致所引起的干扰,从而保障识别模型训练的准确率和训练效率。
本实施例中的从有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像,具体包括:服务器采用OpenCV工具对有效样本图像中的人脸区域和非人脸区域进行截图,获取与标准区域大小相对应的原始正样本图像和原始负样本图像。本实施例中,服务器通过OpenCV工具对有效样本图像中的人脸区域和非人脸区域的4个角的坐标进行截图处理,以获取相应的原始正样本图像和原始负样本图像,通过OpenCV工具进行截图处理,其计算简单、运算效率较高且性能较稳定。
S306:对原始正样本图像和原始负样本图像进行标注,获取携带第一标注信息的原始正样本图像和原始负样本图像。
具体地,服务器在获取原始正样本图像和原始负样本图像之后,需采用预设标注规则对原始正样本图像和原始负样本图像进行标注,以获取携带第一标注信息的原始正样本图像和原始负样本图像。其中,原始正样本图像所携带的第一标注信息包括当前图像标识、来源图像标识、第一人脸标识和第一模糊标识。原始负样本图像所携带的第二标注信息包括当前图像标识、来源图像标识、第二人脸标识和第一模糊标识。
本实施例所提供的识别模型训练方法中,先对训练样本图像进行模糊检测,以便基于训练样本图像的模糊度进行过滤,以获取模糊度达到标准的有效样本图像,从而保障基于有效样本图像获取原始正样本图像和原始负样本图像的清晰度,以保障训练所得的识别模型的识别准确度。然后,在有效样本图像中的人脸区域大小大于预设区域大小时,从有效样本图像中截取原始正样本图像和原始负样本图像,以保证所截取出的原始正样本图像和原始负样本图像的清晰度,从而确保训练所得的识别模型的准确。
在一实施例中,如图4所示,步骤S302,即采用模糊检测算法对训练样本图像进行模糊检测,获取训练样本图像对应的模糊度,具体包括如下步骤:
S401:采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像和锐化图像的像素灰度值。
其中,拉普拉斯算子(Laplacian operator)是一种二阶微分算子,适用于改善因为光线的漫反射造成的图像模糊。其原理是,在摄像记录图像的过程中,光点将光漫反射到其周围区域,这种由于光漫反射造成了图像一定程度的模糊,其模糊程度相对与正常情形下拍摄的图像来说,往往是拉普拉斯算子的常数倍。本实施例中,采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像,具体包括:采用拉普拉斯算子对训练样本图像进行处理,获取描述灰度突变的拉普拉斯图像,再将拉普拉斯图像与训练样本图像叠加从而获取锐化图像。在获取锐化图像之后,获取锐化图像中每一像素点的RGB值,并对RGB值进行处理,以获取锐化图像对应的像素灰度值。
S402:对锐化图像的像素灰度值进行方差计算,获取锐化图像对应的目标方差值,将目标方差值确定为训练样本图像对应的模糊度。
本实施例中,对锐化图像的像素灰度值进行方差计算,以获取其目标方差值,该目标方差值可以理解为锐化图像的模糊度。具体地,对锐化图像的像素灰度值进行方差计算具体包括:计算锐化图像中每个像素点的像素灰度值减去锐化图像的平均灰度值的平方和,再将该平方和除以像素点数量,即可获取可反映锐化图像模糊度的目标方差值。本实施例中,目标方差值越小,说明锐化图像中各像素点的像素灰度值越接近平均灰度值,则锐化图像中各像素点的像素灰度值之间差距不明显,使得图像边缘不清楚,因此,目标方差值越小,其锐化图像越模糊。
本实施例所提供的识别模型训练方法中,先采用拉普拉斯算子对原始图像进行锐化处理,以获取细节比训练样本图像更清晰的锐化图像,从而提高图像的清晰度。然后,通过计算锐化图像的目标方差值,以体现该锐化图像的各像素点的像素灰度值之间的差异性。将锐化图像的目标方差值作为训练样本图像的模糊度,以便依据该模糊度与预设阈值的比较结果,从而达到对训练样本图像进行模糊过滤,以达到获取较清晰的训练样本图像的目的。
在一实施例中,如图5所示,步骤S204,即将人脸标识不相同的目标正样本图像和目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器,具体包括如下步骤:
S501:基于正负样本分配比例,将携带第一人脸标识的目标正样本图像和携带第二人脸标识的目标负样本图像作为模型训练数据。
其中,正负样本分配比例是系统预先设置的用于分配正样本和负样本的比例。该正负样本比例的设置,是为了避免模型训练过程中出现过拟合现象。例如,本实施例中的正负样本比例可以设置为等比例,即正样本数量:负样本数量=1:1,也可以设置为其他不超过比例范围阈值的值。该比例范围阈值是可能引起模型训练过拟合的比例范围。
具体地,服务器基于正负样本分配比例,获取相应的携带第一人脸标识的目标正样本图像和携带第二人脸标识的目标负样本图像,将目标正样本图像和目标负样本图像作为模型训练数据,以便进行后续的MB-FCN模型训练。可以理解地,该模型训练数据是较模糊的包含人脸的目标正样本图像和不包含人脸的目标负样本图像。
S502:基于训练测试分配比例,将模型训练数据划分为训练集和测试集。
其中,训练测试分配比例是系统预先设置的用于划分训练集和测试集的比例。该训练集是用于训练MB-FCN模型的模型训练数据的集合,该测试集是用于测试训练所得的MB-FCN模型的准确率是否达标的模型训练数据的集合。例如,该训练测试分配比例可以设置为训练集中的图像数量:测试集中的图像数量=9:1。
S503:将训练集中的模型训练数据输入到MB-FCN模型进行模型训练,获取原始MB-FCN检测器。
具体地,服务器将训练集中携带第一人脸标识的目标正样本图像和携带第二人脸标识的目标负样本图像等模型训练数据输入MB-FCN模型进行模型训练,在模型训练过程中,根据每一模型训练数据对应的识别结果与预先携带的第一人脸标识或者第二人脸标识进行误差计算,基于该误差更新模型的损失函数,以便基于该损失函数更新MB-FCN模型中的模型参数,从而获取原始MB-FCN检测器。
S504:采用测试集中的模型训练数据对原始MB-FCN检测器进行测试,获取测试准确率。
具体地,服务器将测试集中的每一模型训练数据输入到原始MB-FCN检测器进行测试,获取该模型训练数据对应的测试结果;若测试结果与模型训练数据对应的人脸标识相匹配,则该模型训练数据测试成功;反之,若测试结果与模型训练数据对应的人脸标识不相匹配,则该模型测试数据测试失败。接着,服务器统计测试集中所有测试成功对应的模型训练数据对应的数量为第一数量,并统计测试集中所有模型训练数据对应的数量为第二数量;基于第一数量与第二数量的商,获取原始MB-FCN检测器的测试准确率。
S505:若测试准确率大于准确率阈值,则将原始MB-FCN检测器确定为目标MB-FCN检测器。
其中,准确率阈值是系统预先设置的用于评估模型准确率是否达标的阈值。具体地,服务器将原始MB-FCN检测器的测试准确率与准确率阈值进行比较;若该测试准确率大于准确率阈值,则将原始 MB-FCN检测器确定为目标MB-FCN检测器;若该测试准确率不大于准确率阈值,则需要重新获取模型训练数据,以进行模型训练。
本实施例所提供的识别模型训练方法中,基于正负样本分配比例获取相应的模型训练数据,避免最终训练所得的目标MB-FCN检测器过拟合而影响识别准确性。在基于训练集中的模型训练数据训练原始MB-FCN检测器,再采用测试集中的模型训练数据进行测试,以确定其测试准确率,在测试准确率大于准确率阈值时,将原始MB-FCN检测器确定为目标MB-FCN检测器,以使后续利用该目标MB-FCN检测器从待识别图像中识别出较小或较模糊的人脸图像,以保证较小或较模糊的人脸图像的识别准确率。
在一实施例中,如图6所示,步骤S206,即将目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,具体包括如下步骤:
S601:将目标训练数据中的目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像。
本实施例中,GAN模型中的生成网络具体采用基于超分辨率重构技术形成的生成网络。具体地,该生成网络可以采用DenseNet框架,包括卷积层、N个DenseBlock(密度块)和全连接层,每个DenseBlock(密度块)中设置可实现超分辨率重构的上采样子网和细化子网。其中,上采样子网可是粗粒度地把图像的分辨率放大几倍,即通过上采样操作将图像的分辨率放大几倍。细化子网是通过细粒度处理分辨率放大几倍的细节,其中,细化子网中具体采用反卷积方法进行处理。由于传统的卷积神经网络中,若存在L层,则需要有L个连接,由在DenseNet框架中,会有L(L+1)/2个连接。即在DenseNet框架,每一层的输入均来自前面所有层的输出,这种连接方式使得网络更窄、参数更小,且特征和梯度的传递更加有效,网络也就更加容易训练。
具体地,服务器将目标训练数据中分辨率较低的较模糊的目标正样本图像输入到GAN模型的生成网络中,在该GAN模型的生成网络中,采用基于超分辨率重构技术对目标正样本图像进行重构处理,即通过N个DenseBlock(密度块)中的上采样子网和细化子网进行迭代处理,以获取重构后的超分辨率的训练生成图像,以便基于该训练生成图像更新GAN模型的模型参数。
S602:将目标训练数据中的原始正样本图像和训练生成图像输入GAN模型中的判别网络进行处理,获取目标损失值。
具体地,服务器将一组目标训练数据中较清晰的原始正样本图像和基于较模糊的目标正样本图像进行超分辨率重构形成的训练生成图像一起输入到GAN模型中的判别网络中进行处理,以将该原始正样本图像作为重构的超分辨率的训练生成图像的参照物,获取原始正样本图像与训练生成图像之间的目标损失值。
S603:若目标损失值大于预设损失值,则基于目标损失值更新损失函数,基于更新后的损失函数更新生成网络和判别网络中的模型参数,重复执行将目标训练数据中的目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像。
S604:若目标损失值不大于预设损失值,则模型收敛,获取目标GAN模型。
其中,预设损失值是系统预先设置的用于评估模型收敛时对应的最小的损失值。本实施例中,服务器将目标损失值与预设损失值进行比较;若目标损失值大于预设损失值,则说明GAN模型未收敛达到预设标准,此时,需基于目标损失值更新损失函数,基于更新后的损失函数更新生成网络和判别网络中的模型参数,重复执行上述步骤S601-S602,直至迭代计算之后的目标损失值不大于预设损失值,获取模型收敛达到预设标准的目标GAN模型,以使基于目标GAN模型中的生成网络重构出的超分辨率图像的准确率较高,识别效率较快。本实施例中,生成网络的模型参数包括生成网络的权值和偏移值,而判别网络中的模型参数包括判别网络的权值和偏移值。
本实施例所提供的识别模型训练方法中,将较模糊的目标正样本图像输入到基于超分辨率重构技术形成的生成网络进行处理,获取较清晰的超分辨率的训练生成图像;再将该训练生成图像与同一人脸对应的原始正样本图像输入到判别网络进行处理,根据目标损失值更新模型参数,直至模型收敛,获取目标GAN模型,以使该目标GAN模型重构出的超分辨率图像的准确率较高且效率较快。
进一步地,步骤S205中还可以将第一标注信息中的当前图像标识与第二标注信息中的来源图像标 识相匹配的原始负样本图像和目标负样本图像作为一组目标训练数据。可以理解地,负样本图像对应的目标训练数据的获取过程与正样本图像对应的目标训练数据的获取过程相同,为避免重复,此处不一一赘述。相应地,步骤S601中,还可以将目标训练数据中的目标负样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像。
此时,该GAN模型的判别网络采用两个平行的全连接层(Fully connected layer,简称FC层),在第一个全连接层中用于区分自然真实图像(即原始正样本图像或原始负样本图像)和生成的超分辨率图像(即训练生成图像);在第二个全连接层中用于对包含人脸的图像和不包含人脸的图像进行识别,以使最终训练所得的目标GAN模型不仅可以重构出超分辨率图像,还可以识别图像中的人脸,保障目标GAN模型进行超分辨率图像重构和人脸识别的效率。
本申请实施例还提供的图像识别方法,该图像识别方法可应用如图1所示的应用环境中。具体地,该识别模型训练方法应用在图像识别系统中,该图像识别系统包括如图1所示的客户端和服务器,客户端与服务器通过网络进行通信,用于利用上述实施例训练所得的识别模型,对待处理图像进行识别,以便准确识别出较小或者较模糊的人脸图像,从而确定该人脸图像是否为目标人脸的图像。
如图7所示,本实施例提供一种图像识别方法,以该图像识别方法应用在图1所示的服务器为例,该图像识别方法包括如下步骤:
S701:获取待识别图像,采用目标MB-FCN检测器对待识别图像进行人脸检测和图像截取,获取目标人脸图像。
其中,目标MB-FCN检测器为上述实施例中基于目标正样本图像和目标负样本图像训练所得的检测器。待识别图像是需要进行识别,以确定是否为目标人员的图像。该目标人员为需要进行搜索的人员,包括但不限于犯罪嫌疑人员、失踪人员或者其他需要进行搜索的人员。
目标人脸图像是采用目标MB-FCN检测器检测待识别图像,以检测出待识别图像中的至少一个人脸区域,并采用截图工具对至少一个人脸区域进行截图后获取的人脸图像。由于MB-FCN检测器采用较模糊的目标正样本图像和较模糊的目标负样本图像进行模型训练后获取的用于识别是否为人脸图像的检测器,其可以对较模糊的待识别图像进行识别,以确定较模糊的待识别图像中是否包含人脸对应的人脸区域,以使服务器可采用截图工具从人脸区域中截取包含人脸的目标人脸图像。
S702:采用目标GAN模型中的基于超分辨率重构技术形成的生成网络,对目标人脸图像进行生成处理,获取目标生成图像。
其中,目标GAN模型是上述实施例中采用目标训练数据进行模型训练所获得的模型,该目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络,可将较模糊的低分辨率图像重构形成较清晰的超分辨率图像,重构过程准确率较高且效率较快。
具体地,服务器将目标人脸图像输入到目标GAN模型中,利用目标GAN模型中的基于超分辨率重构技术形成的生成网络对目标人脸图像进行生成处理,以获取超分辨率的目标生成图像,以便利用该目标生成图像进行后续人脸图像识别,以提高识别准确率和识别效率。即目标生成图像是将较低分辨率的目标人脸图像输入到目标GAN模型中,利用基于超分辨率重构技术形成的生成网络,重构出较清晰的超分辨率图像。
S703:将目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度。
其中,目标图像库是用于存储目标人员图像对应的数据库。目标人员图像是指所需追踪的目标人员的图像。具体地,服务器采用人脸相似度检测算法对目标生成图像与目标图像库中的每一目标人员图像进行特征相似度计算,获取计算所得的特征相似度。例如,服务器可以对目标生成图像与每一目标人员图像进行特征提取,获取相应的生成图像特征和人员图像特征,再采用特征距离算法(包括但不限于欧式距离算法或者余弦相似度算法)对生成图像特征和人员图像特征进行计算,以获取相应的特征相似度。
S704:若特征相似度大于相似度阈值,则获取将目标人脸图像对应的人员确定为目标人员的检测结果。
其中,相似度阈值是系统预先设置的用于评估为同一人员时相似度所需达到的阈值。服务器在获 取目标生成图像与每一目标人员图像的特征相似度之后,将该特征相似度与相似度阈值进行比较。若特征相似度大于相似度阈值,说明该待识别图像中的目标人脸图像与目标图像库中的目标人员图像有较大可能为同一人员,因此,可获取目标人员图像对应的人员为目标人员的检测结果。相应地,若特征相似度不大于相似度阈值,则说明该目标人脸图像不为目标人员的图像。
本实施例所提供的图像识别方法,可应用在犯罪嫌疑人追踪、失踪人员搜索或者其他需要进行人员搜索的场景。该方法可从监控设备或者个人随机拍摄的分辨率较低的待处理图像中,先采用目标MB-FCN模型进行人脸检测和图像截取,以获取包含人脸的目标人脸图像,有助于减少后续图像重构和图像识别的数据量,提高识别效率。接着,采用目标GAN模型的基于超分辨率重构技术形成的生成网络对目标人脸图像进行处理,以获取超分辨率的较清晰的目标生成图像,以使后续基于该目标生成图像进行识别时,识别准确率较高且识别效率较快。最后,利用目标生成图像与目标人员图像计算出的特征相似度和相似度阈值比较,以获取目标人脸图像对应的人员是否目标人员的检测结果,从而实现对目标人员的快速追踪。
应理解,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本申请实施例的实施过程构成任何限定。
在一实施例中,提供一种识别模型训练装置,该识别模型训练装置与上述实施例中识别模型训练方法一一对应。如图8所示,该识别模型训练装置包括原始样本图像获取模块801、样本缩略图获取模块802、目标样本图像获取模块803、目标检测器获取模块804、目标训练数据获取模块805和目标GAN模型获取模块806。各功能模块详细说明如下:
原始样本图像获取模块801,用于获取携带第一标注信息的原始正样本图像和原始负样本图像,第一标注信息包括当前图像标识、人脸标识和第一模糊标识。
样本缩略图获取模块802,用于对原始正样本图像和原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图。
目标样本图像获取模块803,用于采用截图工具对正样本缩略图和负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和目标负样本图像,第二标注信息包括来源图像标识、人脸标识和第二模糊标识。
目标检测器获取模块804,用于将人脸标识不相同的目标正样本图像和目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器。
目标训练数据获取模块805,用于将第一标注信息中的当前图像标识与第二标注信息中的来源图像标识相匹配的原始正样本图像和目标正样本图像作为一组目标训练数据。
目标GAN模型获取模块806,用于将目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
优选地,原始样本图像获取模块801,包括:
训练样本图像获取单元,用于从图像数据库中获取包括人脸的训练样本图像。
模糊度获取单元,用于采用模糊检测算法对训练样本图像进行模糊检测,获取训练样本对应的模糊度。
有效样本图像确定单元,用于基于训练样本图像的模糊度,确定有效样本图像。
人脸区域大小获取单元,用于采用人脸检测算法对有效样本图像进行人脸检测,获取人脸区域大小。
原始样本图像截取单元,用于若人脸区域大小大于预设区域大小,则从有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像。
原始样本图像标注单元,用于对原始正样本图像和原始负样本图像进行标注,获取携带第一标注信息的原始正样本图像和原始负样本图像。
优选地,模糊度获取单元,包括:
灰度灰度值获取子单元,用于采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像和锐化图像的像素灰度值。
模糊度确定子单元,用于对锐化图像的像素灰度值进行方差计算,获取锐化图像对应的目标方差值,将目标方差值确定为训练样本图像对应的模糊度。
优选地,目标检测器获取模块804,包括:
模型训练数据获取单元,用于基于正负样本分配比例,将携带第一人脸标识的目标正样本图像和携带第二人脸标识的目标负样本图像作为模型训练数据。
模型训练数据划分单元,用于基于训练测试分配比例,将模型训练数据划分为训练集和测试集。
原始检测器获取单元,用于将训练集中的模型训练数据输入到MB-FCN模型进行模型训练,获取原始MB-FCN检测器。
测试准确率获取单元,用于采用测试集中的模型训练数据对原始MB-FCN检测器进行测试,获取测试准确率。
目标检测器确定单元,用于若测试准确率大于准确率阈值,则将原始MB-FCN检测器确定为目标MB-FCN检测器。
优选地,目标GAN模型获取模块806,包括:
训练生成图像获取单元,用于将目标训练数据中的目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像。
目标损失值获取单元,用于将目标训练数据中的原始正样本图像和训练生成图像输入GAN模型中的判别网络进行处理,获取目标损失值。
模型参数更新单元,用于若目标损失值大于预设损失值,则基于目标损失值更新损失函数,基于更新后的损失函数更新生成网络和判别网络中的模型参数,重复执行将目标训练数据中的目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像。
目标GAN模型获取单元,用于若目标损失值不大于预设损失值,则模型收敛,获取目标GAN模型。
在一实施例中,提供一种图像识别装置,该图像识别装置与上述实施例中图像识别方法一一对应。如图9所示,该图像识别装置包括目标人脸图像获取模块901、目标生成图像获取模块902、特征相似度获取模块903和检测结果获取模块904。各功能模块详细说明如下:
目标人脸图像获取模块901,用于获取待识别图像,采用上述目标MB-FCN检测器对待识别图像进行人脸检测和图像截取,获取目标人脸图像。
目标生成图像获取模块902,用于采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对目标人脸图像进行生成处理,获取目标生成图像。
特征相似度获取模块903,用于将目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度。
检测结果获取模块904,用于若特征相似度大于相似度阈值,则获取将目标人脸图像对应的人员确定为目标人员的检测结果。
关于识别模型训练装置的具体限定可以参见上文中对于识别模型训练方法的限定,在此不再赘述。关于图像识别装置的具体限定可以参见上文中对于图像识别方法的限定,在此不再赘述。上述识别模型训练装置或者图像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10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执行上述识别模型训练方法或者图像识别方法过程中采用或生成的数据。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识别模型训练方法或者图像识别方法。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器 上运行的计算机可读指令,处理器执行计算机可读指令时实现上述实施例中识别模型训练方法的步骤,例如图2至图6所示的步骤。或者,处理器执行计算机可读指令时实现识别模型训练装置这一实施例中的各模块/单元的功能,例如图8所示原始样本图像获取模块801、样本缩略图获取模块802、目标样本图像获取模块803、目标检测器获取模块804、目标训练数据获取模块805和目标GAN模型获取模块806的功能。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令,处理器执行计算机可读指令时实现上述实施例中图像识别方法的步骤,如图7所示的步骤。或者,处理器执行计算机可读指令时实现图像识别装置这一实施例中的各模块/单元的功能,例如图9所示的目标人脸图像获取模块901、目标生成图像获取模块902、特征相似度获取模块903和检测结果获取模块904的功能。为避免重复,这里不再赘述。
在一实施例中,提供一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质存储有计算机可读指令,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时实现上述实施例中识别模型训练方法的步骤,例如图2至图6所示的步骤。或者,该计算机可读指令被处理器执行时实现上述识别模型训练装置这一实施例中的各模块/单元的功能,例如图8所示原始样本图像获取模块801、样本缩略图获取模块802、目标样本图像获取模块803、目标检测器获取模块804、目标训练数据获取模块805和目标GAN模型获取模块806的功能。本实施例中的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
在一实施例中,提供一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质存储有计算机可读指令,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时实现上述实施例中图像识别方法的步骤,如图7所示的步骤,为避免重复,这里不再赘述。或者,该计算机可读指令被处理器执行时实现上述图像识别装置这一实施例中的各模块/单元的功能,例如图9所示的目标人脸图像获取模块901、目标生成图像获取模块902、特征相似度获取模块903和检测结果获取模块904的功能。为避免重复,这里不再赘述。本实施例中的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
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,的计算机可读指令可存储于非易失性可读存储介质也可以存储在一易失性可读存储介质中,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。其中,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。
所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,仅以上述各功能单元、模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能单元、模块完成,即将所述装置的内部结构划分成不同的功能单元或模块,以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案,而非对其限制;尽管参照前述实施例对本申请进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围,均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种识别模型训练方法,其特征在于,包括:
    获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
    对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
    采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
    将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
    将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
    将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
  2. 如权利要求1所述的识别模型训练方法,其特征在于,所述获取携带第一标注信息的原始正样本图像和原始负样本图像,包括:
    从图像数据库中获取包括人脸的训练样本图像;
    采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本对应的模糊度;
    基于所述训练样本图像的模糊度,确定有效样本图像;
    采用人脸检测算法对所述有效样本图像进行人脸检测,获取人脸区域大小;
    若所述人脸区域大小大于预设区域大小,则从所述有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像;
    对所述原始正样本图像和所述原始负样本图像进行标注,获取携带第一标注信息的所述原始正样本图像和所述原始负样本图像。
  3. 如权利要求2所述的识别模型训练方法,其特征在于,所述采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本图像对应的模糊度,包括:
    采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像和所述锐化图像的像素灰度值;
    对所述锐化图像的像素灰度值进行方差计算,获取所述锐化图像对应的目标方差值,将所述目标方差值确定为所述训练样本图像对应的模糊度。
  4. 如权利要求1所述的识别模型训练方法,其特征在于,所述将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器,包括:
    基于正负样本分配比例,将携带第一人脸标识的所述目标正样本图像和携带第二人脸标识的所述目标负样本图像作为模型训练数据;
    基于训练测试分配比例,将所述模型训练数据划分为训练集和测试集;
    将所述训练集中的模型训练数据输入到MB-FCN模型进行模型训练,获取原始MB-FCN检测器;
    采用所述测试集中的模型训练数据对所述原始MB-FCN检测器进行测试,获取测试准确率;
    若所述测试准确率大于准确率阈值,则将所述原始MB-FCN检测器确定为目标MB-FCN检测器。
  5. 如权利要求1所述的识别模型训练方法,其特征在于,所述将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,包括:
    将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    将所述目标训练数据中的所述原始正样本图像和所述训练生成图像输入GAN模型中的判别网络进行处理,获取目标损失值;
    若所述目标损失值大于预设损失值,则基于所述目标损失值更新损失函数,基于更新后的损失函数更新所述生成网络和所述判别网络中的模型参数,重复执行所述将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    若所述目标损失值不大于预设损失值,则模型收敛,获取目标GAN模型。
  6. 一种图像识别方法,其特征在于,包括:
    获取待识别图像,采用权利要求1-5任一项所述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
    采用权利要求1-5任一项所述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
    将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
    若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
  7. 一种识别模型训练装置,其特征在于,包括:
    原始样本图像获取模块,用于获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
    样本缩略图获取模块,用于对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
    目标样本图像获取模块,用于采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
    目标检测器获取模块,用于将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
    目标训练数据获取模块,用于将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
    目标GAN模型获取模块,用于将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
  8. 一种图像识别装置,其特征在于,包括:
    目标人脸图像获取模块,用于获取待识别图像,采用上述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
    目标生成图像获取模块,用于采用上述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
    特征相似度获取模块,用于将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
    检测结果获取模块,用于若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
  9. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其特征在于,所述处理器执行所述计算机可读指令时实现如下步骤:
    获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
    对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
    采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
    将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型 训练,获取目标MB-FCN检测器;
    将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
    将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
  10. 如权利要求9所述的计算机设备,其特征在于,所述获取携带第一标注信息的原始正样本图像和原始负样本图像,包括:
    从图像数据库中获取包括人脸的训练样本图像;
    采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本对应的模糊度;
    基于所述训练样本图像的模糊度,确定有效样本图像;
    采用人脸检测算法对所述有效样本图像进行人脸检测,获取人脸区域大小;
    若所述人脸区域大小大于预设区域大小,则从所述有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像;
    对所述原始正样本图像和所述原始负样本图像进行标注,获取携带第一标注信息的所述原始正样本图像和所述原始负样本图像。
  11. 如权利要求10所述的计算机设备,其特征在于,所述采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本图像对应的模糊度,包括:
    采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像和所述锐化图像的像素灰度值;
    对所述锐化图像的像素灰度值进行方差计算,获取所述锐化图像对应的目标方差值,将所述目标方差值确定为所述训练样本图像对应的模糊度。
  12. 如权利要求9所述的计算机设备,其特征在于,所述将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器,包括:
    基于正负样本分配比例,将携带第一人脸标识的所述目标正样本图像和携带第二人脸标识的所述目标负样本图像作为模型训练数据;
    基于训练测试分配比例,将所述模型训练数据划分为训练集和测试集;
    将所述训练集中的模型训练数据输入到MB-FCN模型进行模型训练,获取原始MB-FCN检测器;
    采用所述测试集中的模型训练数据对所述原始MB-FCN检测器进行测试,获取测试准确率;
    若所述测试准确率大于准确率阈值,则将所述原始MB-FCN检测器确定为目标MB-FCN检测器。
  13. 如权利要求9所述的计算机设备,其特征在于,所述将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,包括:
    将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    将所述目标训练数据中的所述原始正样本图像和所述训练生成图像输入GAN模型中的判别网络进行处理,获取目标损失值;
    若所述目标损失值大于预设损失值,则基于所述目标损失值更新损失函数,基于更新后的损失函数更新所述生成网络和所述判别网络中的模型参数,重复执行所述将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    若所述目标损失值不大于预设损失值,则模型收敛,获取目标GAN模型。
  14. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其特征在于,所述处理器执行所述计算机可读指令时实现如下步骤:
    获取待识别图像,采用权利要求1-5任一项所述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
    采用权利要求1-5任一项所述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
    将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
    若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
  15. 一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质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如下步骤:
    获取携带第一标注信息的原始正样本图像和原始负样本图像,所述第一标注信息包括当前图像标识、人脸标识和第一模糊标识;
    对所述原始正样本图像和所述原始负样本图像进行进行下采样处理,分别获取对应的正样本缩略图和负样本缩略图;
    采用截图工具对所述正样本缩略图和所述负样本缩略图进行截图处理,获取携带第二标注信息的目标正样本图像和所述目标负样本图像,所述第二标注信息包括来源图像标识、人脸标识和第二模糊标识;
    将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器;
    将所述第一标注信息中的所述当前图像标识与第二标注信息中的来源图像标识相匹配的所述原始正样本图像和所述目标正样本图像作为一组目标训练数据;
    将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,所述目标GAN模型中的生成网络为基于超分辨率重构技术形成的生成网络。
  16. 如权利要求15所述的可读存储介质,其特征在于,所述获取携带第一标注信息的原始正样本图像和原始负样本图像,包括:
    从图像数据库中获取包括人脸的训练样本图像;
    采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本对应的模糊度;
    基于所述训练样本图像的模糊度,确定有效样本图像;
    采用人脸检测算法对所述有效样本图像进行人脸检测,获取人脸区域大小;
    若所述人脸区域大小大于预设区域大小,则从所述有效样本图像中截取标准区域大小对应的原始正样本图像和原始负样本图像;
    对所述原始正样本图像和所述原始负样本图像进行标注,获取携带第一标注信息的所述原始正样本图像和所述原始负样本图像。
  17. 如权利要求16所述的可读存储介质,其特征在于,所述采用模糊检测算法对所述训练样本图像进行模糊检测,获取所述训练样本图像对应的模糊度,包括:
    采用拉普拉斯算子对训练样本图像进行锐化处理,获取锐化图像和所述锐化图像的像素灰度值;
    对所述锐化图像的像素灰度值进行方差计算,获取所述锐化图像对应的目标方差值,将所述目标方差值确定为所述训练样本图像对应的模糊度。
  18. 如权利要求15所述的可读存储介质,其特征在于,所述将所述人脸标识不相同的所述目标正样本图像和所述目标负样本图像输入到MB-FCN模型进行模型训练,获取目标MB-FCN检测器,包括:
    基于正负样本分配比例,将携带第一人脸标识的所述目标正样本图像和携带第二人脸标识的所述目标负样本图像作为模型训练数据;
    基于训练测试分配比例,将所述模型训练数据划分为训练集和测试集;
    将所述训练集中的模型训练数据输入到MB-FCN模型进行模型训练,获取原始MB-FCN检测器;
    采用所述测试集中的模型训练数据对所述原始MB-FCN检测器进行测试,获取测试准确率;
    若所述测试准确率大于准确率阈值,则将所述原始MB-FCN检测器确定为目标MB-FCN检测器。
  19. 如权利要求15所述的可读存储介质,其特征在于,所述将所述目标训练数据输入到GAN模型进行模型训练,获取目标GAN模型,包括:
    将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    将所述目标训练数据中的所述原始正样本图像和所述训练生成图像输入GAN模型中的判别网络进行处理,获取目标损失值;
    若所述目标损失值大于预设损失值,则基于所述目标损失值更新损失函数,基于更新后的损失函数更新所述生成网络和所述判别网络中的模型参数,重复执行所述将所述目标训练数据中的所述目标正样本图像输入GAN模型中的基于超分辨率重构技术形成的生成网络进行处理,获取训练生成图像;
    若所述目标损失值不大于预设损失值,则模型收敛,获取目标GAN模型。
  20. 一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读存储介质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如下步骤:
    获取待识别图像,采用权利要求1-5任一项所述目标MB-FCN检测器对所述待识别图像进行人脸检测和图像截取,获取目标人脸图像;
    采用权利要求1-5任一项所述目标GAN模型中的基于超分辨率重构技术形成的生成网络,对所述目标人脸图像进行生成处理,获取目标生成图像;
    将所述目标生成图像与目标图像库中的目标人员图像进行特征相似度计算,获取特征相似度;
    若所述特征相似度大于相似度阈值,则获取将所述目标人脸图像对应的人员确定为目标人员的检测结果。
PCT/CN2019/116489 2019-08-01 2019-11-08 识别模型训练方法、图像识别方法、装置、设备及介质 Ceased WO2021017261A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910706615.9 2019-08-01
CN201910706615.9A CN110569721B (zh) 2019-08-01 2019-08-01 识别模型训练方法、图像识别方法、装置、设备及介质

Publications (1)

Publication Number Publication Date
WO2021017261A1 true WO2021017261A1 (zh) 2021-02-04

Family

ID=68774279

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/116489 Ceased WO2021017261A1 (zh) 2019-08-01 2019-11-08 识别模型训练方法、图像识别方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN110569721B (zh)
WO (1) WO2021017261A1 (zh)

Cited By (73)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112883925A (zh) * 2021-03-23 2021-06-01 杭州海康威视数字技术股份有限公司 一种人脸图像处理方法、装置及设备
CN112926654A (zh) * 2021-02-25 2021-06-08 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
CN112949767A (zh) * 2021-04-07 2021-06-11 北京百度网讯科技有限公司 样本图像增量、图像检测模型训练及图像检测方法
CN113011349A (zh) * 2021-03-24 2021-06-22 中国工商银行股份有限公司 一种票据的要素识别方法、装置及存储介质
CN113033415A (zh) * 2021-03-26 2021-06-25 北京百度网讯科技有限公司 数据队列动态更新方法、装置、电子设备及存储介质
CN113033442A (zh) * 2021-03-31 2021-06-25 清华大学 基于StyleGAN的高自由度人脸驱动方法和装置
CN113095434A (zh) * 2021-04-27 2021-07-09 深圳市商汤科技有限公司 目标检测方法及装置、电子设备、存储介质
CN113112498A (zh) * 2021-05-06 2021-07-13 东北农业大学 一种基于细粒度对抗生成网络的葡萄叶片病斑识别方法
CN113111960A (zh) * 2021-04-25 2021-07-13 北京文安智能技术股份有限公司 图像处理方法和装置以及目标检测模型的训练方法和系统
CN113132359A (zh) * 2021-03-30 2021-07-16 深圳市吉方工控有限公司 一种网络安全数据信息检测方法
CN113160060A (zh) * 2021-05-07 2021-07-23 京东方科技集团股份有限公司 图像处理方法及装置、设备、存储介质
CN113159209A (zh) * 2021-04-29 2021-07-23 深圳市商汤科技有限公司 目标检测方法、装置、设备和计算机可读存储介质
CN113221104A (zh) * 2021-05-12 2021-08-06 北京百度网讯科技有限公司 用户异常行为的检测方法及用户行为重构模型的训练方法
CN113255911A (zh) * 2021-06-07 2021-08-13 杭州海康威视数字技术股份有限公司 一种模型训练方法、装置、电子设备及存储介质
CN113269736A (zh) * 2021-05-17 2021-08-17 唐旸 紧固件尺寸的自动化检验方法、系统及介质
CN113326852A (zh) * 2021-06-11 2021-08-31 北京百度网讯科技有限公司 模型训练方法、装置、设备、存储介质及程序产品
CN113344214A (zh) * 2021-05-31 2021-09-03 北京百度网讯科技有限公司 数据处理模型的训练方法、装置、电子设备及存储介质
CN113361603A (zh) * 2021-06-04 2021-09-07 北京百度网讯科技有限公司 训练方法、类别识别方法、装置、电子设备以及存储介质
CN113469878A (zh) * 2021-09-02 2021-10-01 北京世纪好未来教育科技有限公司 一种文本擦除方法及其模型的训练方法、装置、存储介质
CN113486839A (zh) * 2021-07-20 2021-10-08 支付宝(杭州)信息技术有限公司 加密模型训练、图像加密和加密人脸图像识别方法及装置
CN113516046A (zh) * 2021-05-18 2021-10-19 平安国际智慧城市科技股份有限公司 区域内生物多样性的监测方法、装置、设备以及存储介质
CN113569691A (zh) * 2021-07-19 2021-10-29 新疆爱华盈通信息技术有限公司 人头检测模型生成方法、装置、人头检测模型及人头检测方法
CN113592696A (zh) * 2021-08-12 2021-11-02 支付宝(杭州)信息技术有限公司 加密模型训练、图像加密和加密人脸图像识别方法及装置
CN113591782A (zh) * 2021-08-12 2021-11-02 北京惠朗时代科技有限公司 一种基于训练式的人脸识别智能保险柜应用方法及系统
CN113657249A (zh) * 2021-08-13 2021-11-16 北京百度网讯科技有限公司 训练方法、预测方法、装置、电子设备以及存储介质
CN113704531A (zh) * 2021-03-10 2021-11-26 腾讯科技(深圳)有限公司 图像处理方法、装置、电子设备及计算机可读存储介质
CN113705596A (zh) * 2021-03-04 2021-11-26 腾讯科技(北京)有限公司 图像识别方法、装置、计算机设备和存储介质
CN113806613A (zh) * 2021-09-29 2021-12-17 中国平安人寿保险股份有限公司 训练图像集生成方法、装置、计算机设备及存储介质
CN113887530A (zh) * 2021-11-11 2022-01-04 重庆钢铁股份有限公司 液压站监控管理系统
CN113903044A (zh) * 2021-09-30 2022-01-07 京东方科技集团股份有限公司 识别物料标记的方法、训练方法、设备及存储介质
CN113902671A (zh) * 2021-08-31 2022-01-07 北京影谱科技股份有限公司 一种基于随机纹理的图像隐写方法和系统
CN113971751A (zh) * 2021-10-28 2022-01-25 北京百度网讯科技有限公司 训练特征提取模型、检测相似图像的方法和装置
CN114219029A (zh) * 2021-12-16 2022-03-22 中国建设银行股份有限公司 一种模型训练系统、方法、装置、设备及介质
CN114266975A (zh) * 2021-12-23 2022-04-01 华南农业大学 无人机遥感图像的荔枝果实检测与计数方法
CN114283460A (zh) * 2021-09-26 2022-04-05 腾讯科技(深圳)有限公司 一种特征提取方法、装置、计算机设备及存储介质
CN114445682A (zh) * 2022-01-28 2022-05-06 北京百度网讯科技有限公司 训练模型的方法、装置、电子设备、存储介质及产品
CN114549392A (zh) * 2021-03-31 2022-05-27 正泰集团研发中心(上海)有限公司 太阳能电池片的色差检测方法、装置、设备和计算机介质
CN114596484A (zh) * 2022-02-22 2022-06-07 上海欧菲智能车联科技有限公司 目标检测能力训练方法、装置、存储介质及电子设备
CN114913354A (zh) * 2022-05-16 2022-08-16 武汉墨仗信息科技股份有限公司 一种智慧化综合管理方法及平台
CN115019334A (zh) * 2022-05-31 2022-09-06 中邮信息科技(北京)有限公司 一种任务处理方法、装置、设备及介质
CN115017994A (zh) * 2022-06-13 2022-09-06 蚂蚁区块链科技(上海)有限公司 分类模型训练方法及装置
CN115019374A (zh) * 2022-07-18 2022-09-06 北京师范大学 基于人工智能的智慧课堂学生专注度低耗检测方法及系统
CN115049546A (zh) * 2022-06-07 2022-09-13 苏州超云生命智能产业研究院有限公司 样本数据处理方法、装置、电子设备及存储介质
CN115081500A (zh) * 2021-03-12 2022-09-20 深圳海翼智新科技有限公司 用于对象识别模型的训练方法、装置及计算机存储介质
CN115100420A (zh) * 2022-07-22 2022-09-23 南京理工大学 一种视觉小目标外观特征的提取方法
CN115115972A (zh) * 2022-05-25 2022-09-27 腾讯科技(深圳)有限公司 视频处理方法、装置、计算机设备、介质及程序产品
CN115225413A (zh) * 2022-09-20 2022-10-21 北京微步在线科技有限公司 一种失陷指标的提取方法、装置、电子设备及存储介质
CN115272817A (zh) * 2022-07-12 2022-11-01 浙江大华技术股份有限公司 一种目标检测的方法、模型训练方法及设备
CN115273174A (zh) * 2022-06-21 2022-11-01 浙江大华技术股份有限公司 特征提取网络的训练方法、脸部识别方法及相关设备
CN115270151A (zh) * 2022-07-21 2022-11-01 上海商汤智能科技有限公司 扰动信息的生成方法、对象识别方法、装置、设备及介质
CN115409761A (zh) * 2021-05-28 2022-11-29 上海交通大学 目标计数方法及装置、电子设备、存储介质
CN115424355A (zh) * 2022-10-14 2022-12-02 通号通信信息集团有限公司 抽烟检测方法及装置、电子设备、计算机可读存储介质
WO2022262209A1 (zh) * 2021-06-17 2022-12-22 深圳市商汤科技有限公司 一种神经网络训练方法、装置、计算机设备及存储介质
CN116091797A (zh) * 2022-07-25 2023-05-09 网易(杭州)网络有限公司 图像相似度确定方法及其模型的训练方法及装置
CN116128954A (zh) * 2022-12-30 2023-05-16 上海强仝智能科技有限公司 一种基于生成网络的商品布局识别方法、装置及存储介质
CN116188256A (zh) * 2021-11-25 2023-05-30 北京字跳网络技术有限公司 超分图像处理方法、装置、设备及介质
CN116310286A (zh) * 2023-02-08 2023-06-23 杭州未名信科科技有限公司 一种纸箱唛头智能校对的方法、装置、设备及存储介质
CN116310656A (zh) * 2023-05-11 2023-06-23 福瑞泰克智能系统有限公司 训练样本确定方法、装置和计算机设备
CN116541707A (zh) * 2023-05-16 2023-08-04 中电信数智科技有限公司 一种图文匹配模型训练方法、装置、设备及存储介质
CN116977794A (zh) * 2023-08-28 2023-10-31 杭州一知智能科技有限公司 一种基于强化学习的数字人视频鉴别模型训练方法和系统
CN117132174A (zh) * 2023-10-26 2023-11-28 扬宇光电(深圳)有限公司 一种应用于工业流水线质量检测的模型训练方法与系统
CN117132922A (zh) * 2023-08-25 2023-11-28 中国联合网络通信集团有限公司 图像识别方法、装置、设备及存储介质
CN117173385A (zh) * 2023-10-24 2023-12-05 四川思极科技有限公司 一种变电站的检测方法、装置、介质及设备
CN117788091A (zh) * 2023-10-18 2024-03-29 北京农夫铺子技术研究院 基于用户消费大数据的元宇宙电商平台构建方法及系统
CN118011405A (zh) * 2024-04-08 2024-05-10 厦门印海途海洋科技有限公司 海底地质成像方法、装置、设备及介质
CN118072289A (zh) * 2024-04-18 2024-05-24 智联信通科技股份有限公司 用于智能驾驶的图像采集优化方法
CN118279181A (zh) * 2024-05-31 2024-07-02 杭州海康威视数字技术股份有限公司 可调参图像复原模型的训练方法和可调参的图像复原方法
CN118397420A (zh) * 2024-07-01 2024-07-26 中国计量大学 一种图像目标识别方法
CN118537595A (zh) * 2024-05-08 2024-08-23 山东云海国创云计算装备产业创新中心有限公司 一种采样图像的验证方法、装置、计算机设备及存储介质
CN118692050A (zh) * 2023-03-22 2024-09-24 比亚迪股份有限公司 目标检测方法、车辆及存储介质
CN118840709A (zh) * 2024-09-24 2024-10-25 山东开创电气有限公司 一种基于图像识别的煤流量动态实时监测分析处理系统
CN118971892A (zh) * 2023-05-15 2024-11-15 Oppo广东移动通信有限公司 数据处理方法、装置、计算机设备及存储介质
CN120409589A (zh) * 2025-04-28 2025-08-01 上海智元新创技术有限公司 模型训练方法、移动设备、电子设备及介质

Families Citing this family (36)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111291817B (zh) * 2020-02-17 2024-01-23 北京迈格威科技有限公司 图像识别方法、装置、电子设备和计算机可读介质
CN111275011B (zh) * 2020-02-25 2023-12-19 阿波罗智能技术(北京)有限公司 移动红绿灯检测方法、装置、电子设备和存储介质
CN111368685B (zh) * 2020-02-27 2023-09-29 北京字节跳动网络技术有限公司 关键点的识别方法、装置、可读介质和电子设备
EP4116929A4 (en) * 2020-03-02 2023-07-05 Panasonic Intellectual Property Corporation of America Information processing method, information processing system, and information processing device
WO2021179198A1 (zh) * 2020-03-11 2021-09-16 深圳先进技术研究院 图像特征可视化方法、图像特征可视化装置及电子设备
CN111428875A (zh) * 2020-03-11 2020-07-17 北京三快在线科技有限公司 图像识别方法、装置及相应模型训练方法、装置
CN111401224B (zh) * 2020-03-13 2023-05-23 北京字节跳动网络技术有限公司 目标检测方法、装置及电子设备
CN111368934B (zh) * 2020-03-17 2023-09-19 腾讯科技(深圳)有限公司 图像识别模型训练方法、图像识别方法以及相关装置
CN113515980B (zh) * 2020-05-20 2022-07-05 阿里巴巴集团控股有限公司 模型训练方法、装置、设备和存储介质
CN111798414A (zh) * 2020-06-12 2020-10-20 北京阅视智能技术有限责任公司 显微图像的清晰度确定方法、装置、设备及存储介质
CN111709878B (zh) * 2020-06-17 2023-06-23 北京百度网讯科技有限公司 人脸超分辨率实现方法、装置、电子设备及存储介质
CN111768336B (zh) * 2020-07-09 2022-11-01 腾讯科技(深圳)有限公司 人脸图像处理方法、装置、计算机设备和存储介质
CN111967592B (zh) * 2020-07-09 2023-12-05 中国电子科技集团公司第三十六研究所 基于分离正负扰动生成对抗图像机器识别的方法
CN112085056B (zh) * 2020-08-05 2023-12-29 深圳市优必选科技股份有限公司 目标检测模型生成方法、装置、设备及存储介质
CN112149684B (zh) * 2020-08-19 2024-06-07 北京豆牛网络科技有限公司 图像处理方法和用于目标检测的图像预处理方法
CN111985565B (zh) * 2020-08-20 2023-01-10 上海风秩科技有限公司 图片分析方法和装置、存储介质及电子设备
CN112001355A (zh) * 2020-09-03 2020-11-27 杭州云栖智慧视通科技有限公司 室外视频下模糊人脸识别的训练数据预处理方法
CN112232420B (zh) * 2020-10-19 2025-04-08 深圳市华汉伟业科技有限公司 一种图像标注方法、目标检测方法及装置、存储介质
CN112488947B (zh) * 2020-12-04 2025-09-05 北京字跳网络技术有限公司 模型训练和图像处理方法、装置、设备和计算机可读介质
CN112561879B (zh) * 2020-12-15 2024-01-09 北京百度网讯科技有限公司 模糊度评价模型训练方法、图像模糊度评价方法及装置
CN112633276B (zh) * 2020-12-25 2024-06-21 北京百度网讯科技有限公司 训练方法、识别方法、装置、设备、介质
CN112529114B (zh) * 2021-01-13 2021-06-29 北京云真信科技有限公司 基于gan的目标信息识别方法、电子设备和介质
CN112926399A (zh) * 2021-01-28 2021-06-08 上海商汤智能科技有限公司 目标对象检测方法、装置、电子设备和存储介质
CN113723437B (zh) * 2021-04-02 2022-06-07 荣耀终端有限公司 一种ai模型的自动化训练方法及ai模型训练系统
CN113177469B (zh) * 2021-04-27 2024-04-12 北京百度网讯科技有限公司 人体属性检测模型的训练方法、装置、电子设备及介质
CN113240376A (zh) * 2021-05-31 2021-08-10 中邮信息科技(北京)有限公司 物品信息确定方法、装置、电子设备及介质
CN113361543B (zh) * 2021-06-09 2024-05-21 北京工业大学 Ct图像特征提取方法、装置、电子设备和存储介质
CN113409207B (zh) * 2021-06-15 2023-12-08 广州光锥元信息科技有限公司 一种人脸图像清晰度提升方法及装置
TWI821715B (zh) * 2021-07-20 2023-11-11 和碩聯合科技股份有限公司 生成器網路模型的訓練方法及其執行的電子裝置
CN113486858B (zh) * 2021-08-03 2024-01-23 济南博观智能科技有限公司 一种人脸识别模型训练方法、装置、电子设备及存储介质
CN113963216A (zh) * 2021-11-10 2022-01-21 广东电网能源发展有限公司 钢丝绳缺陷识别方法、装置、设备和介质
CN114118412A (zh) * 2021-11-30 2022-03-01 平安科技(深圳)有限公司 证件识别模型训练及证件识别的方法、系统、设备及介质
CN114299371A (zh) * 2021-11-30 2022-04-08 深圳壹账通智能科技有限公司 证件识别模型训练及证件识别的方法、系统、设备及介质
CN114612728B (zh) * 2022-03-25 2025-12-02 Oppo广东移动通信有限公司 模型训练方法、装置、计算机设备及存储介质
CN114821739A (zh) * 2022-05-16 2022-07-29 北京百度网讯科技有限公司 一种模型训练的方法、装置、电子设备及存储介质
CN115050086B (zh) * 2022-08-15 2022-11-04 北京百度网讯科技有限公司 样本图像生成方法、模型训练方法、图像处理方法和装置

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108334848A (zh) * 2018-02-06 2018-07-27 哈尔滨工业大学 一种基于生成对抗网络的微小人脸识别方法
CN108805809A (zh) * 2018-05-28 2018-11-13 天津科技大学 一种基于生成对抗网络的红外人脸图像超分辨率重建方法
US10325201B1 (en) * 2019-01-31 2019-06-18 StradVision, Inc. Method and device for generating deceivable composite image by using GAN including generating neural network and discriminating neural network to allow surveillance system to recognize surroundings and detect rare event more accurately

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108108753B (zh) * 2017-12-15 2022-08-19 京北方信息技术股份有限公司 一种基于支持向量机的复选框选择状态的识别方法及装置
CN108510194B (zh) * 2018-03-30 2022-11-29 平安科技(深圳)有限公司 风控模型训练方法、风险识别方法、装置、设备及介质
CN108764082A (zh) * 2018-05-17 2018-11-06 淘然视界(杭州)科技有限公司 一种飞机目标检测方法、电子设备、存储介质及系统
CN108985155A (zh) * 2018-06-06 2018-12-11 平安科技(深圳)有限公司 嘴巴模型训练方法、嘴巴识别方法、装置、设备及介质
CN109711258A (zh) * 2018-11-27 2019-05-03 哈尔滨工业大学(深圳) 基于卷积网络的轻量级人脸关键点检测方法、系统及存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108334848A (zh) * 2018-02-06 2018-07-27 哈尔滨工业大学 一种基于生成对抗网络的微小人脸识别方法
CN108805809A (zh) * 2018-05-28 2018-11-13 天津科技大学 一种基于生成对抗网络的红外人脸图像超分辨率重建方法
US10325201B1 (en) * 2019-01-31 2019-06-18 StradVision, Inc. Method and device for generating deceivable composite image by using GAN including generating neural network and discriminating neural network to allow surveillance system to recognize surroundings and detect rare event more accurately

Cited By (96)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926654A (zh) * 2021-02-25 2021-06-08 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
CN112926654B (zh) * 2021-02-25 2023-08-01 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
CN113705596A (zh) * 2021-03-04 2021-11-26 腾讯科技(北京)有限公司 图像识别方法、装置、计算机设备和存储介质
CN113704531A (zh) * 2021-03-10 2021-11-26 腾讯科技(深圳)有限公司 图像处理方法、装置、电子设备及计算机可读存储介质
CN115081500A (zh) * 2021-03-12 2022-09-20 深圳海翼智新科技有限公司 用于对象识别模型的训练方法、装置及计算机存储介质
CN112883925A (zh) * 2021-03-23 2021-06-01 杭州海康威视数字技术股份有限公司 一种人脸图像处理方法、装置及设备
CN112883925B (zh) * 2021-03-23 2023-08-29 杭州海康威视数字技术股份有限公司 一种人脸图像处理方法、装置及设备
CN113011349A (zh) * 2021-03-24 2021-06-22 中国工商银行股份有限公司 一种票据的要素识别方法、装置及存储介质
CN113033415B (zh) * 2021-03-26 2023-11-28 北京百度网讯科技有限公司 数据队列动态更新方法、装置、电子设备及存储介质
CN113033415A (zh) * 2021-03-26 2021-06-25 北京百度网讯科技有限公司 数据队列动态更新方法、装置、电子设备及存储介质
CN113132359A (zh) * 2021-03-30 2021-07-16 深圳市吉方工控有限公司 一种网络安全数据信息检测方法
CN113033442A (zh) * 2021-03-31 2021-06-25 清华大学 基于StyleGAN的高自由度人脸驱动方法和装置
CN114549392B (zh) * 2021-03-31 2024-05-31 正泰集团研发中心(上海)有限公司 太阳能电池片的色差检测方法、装置、设备和计算机介质
CN114549392A (zh) * 2021-03-31 2022-05-27 正泰集团研发中心(上海)有限公司 太阳能电池片的色差检测方法、装置、设备和计算机介质
CN112949767B (zh) * 2021-04-07 2023-08-11 北京百度网讯科技有限公司 样本图像增量、图像检测模型训练及图像检测方法
CN112949767A (zh) * 2021-04-07 2021-06-11 北京百度网讯科技有限公司 样本图像增量、图像检测模型训练及图像检测方法
CN113111960A (zh) * 2021-04-25 2021-07-13 北京文安智能技术股份有限公司 图像处理方法和装置以及目标检测模型的训练方法和系统
CN113111960B (zh) * 2021-04-25 2024-04-26 北京文安智能技术股份有限公司 图像处理方法和装置以及目标检测模型的训练方法和系统
CN113095434B (zh) * 2021-04-27 2024-06-11 深圳市商汤科技有限公司 目标检测方法及装置、电子设备、存储介质
CN113095434A (zh) * 2021-04-27 2021-07-09 深圳市商汤科技有限公司 目标检测方法及装置、电子设备、存储介质
CN113159209A (zh) * 2021-04-29 2021-07-23 深圳市商汤科技有限公司 目标检测方法、装置、设备和计算机可读存储介质
CN113159209B (zh) * 2021-04-29 2024-05-24 深圳市商汤科技有限公司 目标检测方法、装置、设备和计算机可读存储介质
CN113112498A (zh) * 2021-05-06 2021-07-13 东北农业大学 一种基于细粒度对抗生成网络的葡萄叶片病斑识别方法
CN113112498B (zh) * 2021-05-06 2024-01-19 东北农业大学 一种基于细粒度对抗生成网络的葡萄叶片病斑识别方法
CN113160060A (zh) * 2021-05-07 2021-07-23 京东方科技集团股份有限公司 图像处理方法及装置、设备、存储介质
CN113221104B (zh) * 2021-05-12 2023-07-28 北京百度网讯科技有限公司 用户异常行为的检测方法及用户行为重构模型的训练方法
CN113221104A (zh) * 2021-05-12 2021-08-06 北京百度网讯科技有限公司 用户异常行为的检测方法及用户行为重构模型的训练方法
CN113269736B (zh) * 2021-05-17 2023-01-13 唐旸 紧固件尺寸的自动化检验方法、系统及介质
CN113269736A (zh) * 2021-05-17 2021-08-17 唐旸 紧固件尺寸的自动化检验方法、系统及介质
CN113516046A (zh) * 2021-05-18 2021-10-19 平安国际智慧城市科技股份有限公司 区域内生物多样性的监测方法、装置、设备以及存储介质
CN115409761A (zh) * 2021-05-28 2022-11-29 上海交通大学 目标计数方法及装置、电子设备、存储介质
CN113344214A (zh) * 2021-05-31 2021-09-03 北京百度网讯科技有限公司 数据处理模型的训练方法、装置、电子设备及存储介质
CN113361603A (zh) * 2021-06-04 2021-09-07 北京百度网讯科技有限公司 训练方法、类别识别方法、装置、电子设备以及存储介质
CN113361603B (zh) * 2021-06-04 2024-05-10 北京百度网讯科技有限公司 训练方法、类别识别方法、装置、电子设备以及存储介质
CN113255911A (zh) * 2021-06-07 2021-08-13 杭州海康威视数字技术股份有限公司 一种模型训练方法、装置、电子设备及存储介质
CN113255911B (zh) * 2021-06-07 2023-10-13 杭州海康威视数字技术股份有限公司 一种模型训练方法、装置、电子设备及存储介质
CN113326852A (zh) * 2021-06-11 2021-08-31 北京百度网讯科技有限公司 模型训练方法、装置、设备、存储介质及程序产品
WO2022262209A1 (zh) * 2021-06-17 2022-12-22 深圳市商汤科技有限公司 一种神经网络训练方法、装置、计算机设备及存储介质
CN113569691A (zh) * 2021-07-19 2021-10-29 新疆爱华盈通信息技术有限公司 人头检测模型生成方法、装置、人头检测模型及人头检测方法
CN113486839A (zh) * 2021-07-20 2021-10-08 支付宝(杭州)信息技术有限公司 加密模型训练、图像加密和加密人脸图像识别方法及装置
CN113592696A (zh) * 2021-08-12 2021-11-02 支付宝(杭州)信息技术有限公司 加密模型训练、图像加密和加密人脸图像识别方法及装置
CN113591782A (zh) * 2021-08-12 2021-11-02 北京惠朗时代科技有限公司 一种基于训练式的人脸识别智能保险柜应用方法及系统
CN113657249A (zh) * 2021-08-13 2021-11-16 北京百度网讯科技有限公司 训练方法、预测方法、装置、电子设备以及存储介质
CN113657249B (zh) * 2021-08-13 2024-05-17 北京神州数码云科信息技术有限公司 训练方法、预测方法、装置、电子设备以及存储介质
CN113902671A (zh) * 2021-08-31 2022-01-07 北京影谱科技股份有限公司 一种基于随机纹理的图像隐写方法和系统
CN113469878A (zh) * 2021-09-02 2021-10-01 北京世纪好未来教育科技有限公司 一种文本擦除方法及其模型的训练方法、装置、存储介质
CN113469878B (zh) * 2021-09-02 2021-11-12 北京世纪好未来教育科技有限公司 一种文本擦除方法及其模型的训练方法、装置、存储介质
CN114283460A (zh) * 2021-09-26 2022-04-05 腾讯科技(深圳)有限公司 一种特征提取方法、装置、计算机设备及存储介质
CN113806613B (zh) * 2021-09-29 2023-07-25 中国平安人寿保险股份有限公司 训练图像集生成方法、装置、计算机设备及存储介质
CN113806613A (zh) * 2021-09-29 2021-12-17 中国平安人寿保险股份有限公司 训练图像集生成方法、装置、计算机设备及存储介质
CN113903044A (zh) * 2021-09-30 2022-01-07 京东方科技集团股份有限公司 识别物料标记的方法、训练方法、设备及存储介质
CN113971751A (zh) * 2021-10-28 2022-01-25 北京百度网讯科技有限公司 训练特征提取模型、检测相似图像的方法和装置
CN113887530A (zh) * 2021-11-11 2022-01-04 重庆钢铁股份有限公司 液压站监控管理系统
CN116188256A (zh) * 2021-11-25 2023-05-30 北京字跳网络技术有限公司 超分图像处理方法、装置、设备及介质
CN114219029A (zh) * 2021-12-16 2022-03-22 中国建设银行股份有限公司 一种模型训练系统、方法、装置、设备及介质
CN114266975B (zh) * 2021-12-23 2024-04-16 华南农业大学 无人机遥感图像的荔枝果实检测与计数方法
CN114266975A (zh) * 2021-12-23 2022-04-01 华南农业大学 无人机遥感图像的荔枝果实检测与计数方法
CN114445682A (zh) * 2022-01-28 2022-05-06 北京百度网讯科技有限公司 训练模型的方法、装置、电子设备、存储介质及产品
CN114596484A (zh) * 2022-02-22 2022-06-07 上海欧菲智能车联科技有限公司 目标检测能力训练方法、装置、存储介质及电子设备
CN114913354A (zh) * 2022-05-16 2022-08-16 武汉墨仗信息科技股份有限公司 一种智慧化综合管理方法及平台
CN115115972A (zh) * 2022-05-25 2022-09-27 腾讯科技(深圳)有限公司 视频处理方法、装置、计算机设备、介质及程序产品
CN115019334A (zh) * 2022-05-31 2022-09-06 中邮信息科技(北京)有限公司 一种任务处理方法、装置、设备及介质
CN115049546A (zh) * 2022-06-07 2022-09-13 苏州超云生命智能产业研究院有限公司 样本数据处理方法、装置、电子设备及存储介质
CN115017994A (zh) * 2022-06-13 2022-09-06 蚂蚁区块链科技(上海)有限公司 分类模型训练方法及装置
CN115273174A (zh) * 2022-06-21 2022-11-01 浙江大华技术股份有限公司 特征提取网络的训练方法、脸部识别方法及相关设备
CN115272817A (zh) * 2022-07-12 2022-11-01 浙江大华技术股份有限公司 一种目标检测的方法、模型训练方法及设备
CN115019374B (zh) * 2022-07-18 2022-10-11 北京师范大学 基于人工智能的智慧课堂学生专注度低耗检测方法及系统
CN115019374A (zh) * 2022-07-18 2022-09-06 北京师范大学 基于人工智能的智慧课堂学生专注度低耗检测方法及系统
CN115270151A (zh) * 2022-07-21 2022-11-01 上海商汤智能科技有限公司 扰动信息的生成方法、对象识别方法、装置、设备及介质
CN115100420A (zh) * 2022-07-22 2022-09-23 南京理工大学 一种视觉小目标外观特征的提取方法
CN116091797A (zh) * 2022-07-25 2023-05-09 网易(杭州)网络有限公司 图像相似度确定方法及其模型的训练方法及装置
CN115225413A (zh) * 2022-09-20 2022-10-21 北京微步在线科技有限公司 一种失陷指标的提取方法、装置、电子设备及存储介质
CN115424355A (zh) * 2022-10-14 2022-12-02 通号通信信息集团有限公司 抽烟检测方法及装置、电子设备、计算机可读存储介质
CN116128954B (zh) * 2022-12-30 2023-12-05 上海强仝智能科技有限公司 一种基于生成网络的商品布局识别方法、装置及存储介质
CN116128954A (zh) * 2022-12-30 2023-05-16 上海强仝智能科技有限公司 一种基于生成网络的商品布局识别方法、装置及存储介质
CN116310286A (zh) * 2023-02-08 2023-06-23 杭州未名信科科技有限公司 一种纸箱唛头智能校对的方法、装置、设备及存储介质
CN118692050A (zh) * 2023-03-22 2024-09-24 比亚迪股份有限公司 目标检测方法、车辆及存储介质
CN116310656B (zh) * 2023-05-11 2023-08-15 福瑞泰克智能系统有限公司 训练样本确定方法、装置和计算机设备
CN116310656A (zh) * 2023-05-11 2023-06-23 福瑞泰克智能系统有限公司 训练样本确定方法、装置和计算机设备
CN118971892A (zh) * 2023-05-15 2024-11-15 Oppo广东移动通信有限公司 数据处理方法、装置、计算机设备及存储介质
CN116541707A (zh) * 2023-05-16 2023-08-04 中电信数智科技有限公司 一种图文匹配模型训练方法、装置、设备及存储介质
CN117132922A (zh) * 2023-08-25 2023-11-28 中国联合网络通信集团有限公司 图像识别方法、装置、设备及存储介质
CN116977794B (zh) * 2023-08-28 2024-04-02 杭州一知智能科技有限公司 一种基于强化学习的数字人视频鉴别模型训练方法和系统
CN116977794A (zh) * 2023-08-28 2023-10-31 杭州一知智能科技有限公司 一种基于强化学习的数字人视频鉴别模型训练方法和系统
CN117788091A (zh) * 2023-10-18 2024-03-29 北京农夫铺子技术研究院 基于用户消费大数据的元宇宙电商平台构建方法及系统
CN117173385B (zh) * 2023-10-24 2024-01-26 四川思极科技有限公司 一种变电站的检测方法、装置、介质及设备
CN117173385A (zh) * 2023-10-24 2023-12-05 四川思极科技有限公司 一种变电站的检测方法、装置、介质及设备
CN117132174A (zh) * 2023-10-26 2023-11-28 扬宇光电(深圳)有限公司 一种应用于工业流水线质量检测的模型训练方法与系统
CN117132174B (zh) * 2023-10-26 2024-01-30 扬宇光电(深圳)有限公司 一种应用于工业流水线质量检测的模型训练方法与系统
CN118011405A (zh) * 2024-04-08 2024-05-10 厦门印海途海洋科技有限公司 海底地质成像方法、装置、设备及介质
CN118072289A (zh) * 2024-04-18 2024-05-24 智联信通科技股份有限公司 用于智能驾驶的图像采集优化方法
CN118537595A (zh) * 2024-05-08 2024-08-23 山东云海国创云计算装备产业创新中心有限公司 一种采样图像的验证方法、装置、计算机设备及存储介质
CN118279181A (zh) * 2024-05-31 2024-07-02 杭州海康威视数字技术股份有限公司 可调参图像复原模型的训练方法和可调参的图像复原方法
CN118397420A (zh) * 2024-07-01 2024-07-26 中国计量大学 一种图像目标识别方法
CN118840709A (zh) * 2024-09-24 2024-10-25 山东开创电气有限公司 一种基于图像识别的煤流量动态实时监测分析处理系统
CN120409589A (zh) * 2025-04-28 2025-08-01 上海智元新创技术有限公司 模型训练方法、移动设备、电子设备及介质

Also Published As

Publication number Publication date
CN110569721A (zh) 2019-12-13
CN110569721B (zh) 2023-08-29

Similar Documents

Publication Publication Date Title
WO2021017261A1 (zh) 识别模型训练方法、图像识别方法、装置、设备及介质
US10936911B2 (en) Logo detection
CN110399799B (zh) 图像识别和神经网络模型的训练方法、装置和系统
CN111832581B (zh) 肺部特征识别方法、装置、计算机设备及存储介质
CN110852349B (zh) 一种图像处理方法、检测方法、相关设备及存储介质
WO2019237846A1 (zh) 图像处理方法、人脸识别方法、装置和计算机设备
US12073537B1 (en) Image data enhancement method and apparatus, computer device, and storage medium
WO2021175006A1 (zh) 车辆图像检测方法、装置、计算机设备及存储介质
WO2021139324A1 (zh) 图像识别方法、装置、计算机可读存储介质及电子设备
CN112668462B (zh) 车损检测模型训练、车损检测方法、装置、设备及介质
WO2021012382A1 (zh) 配置聊天机器人的方法、装置、计算机设备和存储介质
WO2016054779A1 (en) Spatial pyramid pooling networks for image processing
CN111709313B (zh) 基于局部和通道组合特征的行人重识别方法
US11756319B2 (en) Shift invariant loss for deep learning based image segmentation
CN111191649A (zh) 一种识别弯曲多行文本图像的方法与设备
WO2020098257A1 (zh) 一种图像分类方法、装置及计算机可读存储介质
WO2021189770A1 (zh) 基于人工智能的图像增强处理方法、装置、设备及介质
CN112884782B (zh) 生物对象分割方法、装置、计算机设备和存储介质
CN114743121A (zh) 图像处理方法、图像处理模型的训练方法及装置
CN115082667A (zh) 图像处理方法、装置、设备及存储介质
CN112464873A (zh) 模型的训练方法、人脸活体识别方法、系统、设备及介质
CN112597909A (zh) 一种用于人脸图片质量评价的方法与设备
CN112836682B (zh) 视频中对象的识别方法、装置、计算机设备和存储介质
CN111860582B (zh) 图像分类模型构建方法、装置、计算机设备和存储介质
CN112749702B (zh) 一种图像识别方法、装置、终端及存储介质

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: 19939164

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19939164

Country of ref document: EP

Kind code of ref document: A1