WO2022021901A1 - 目标检测方法及装置、电子设备和存储介质 - Google Patents

目标检测方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2022021901A1
WO2022021901A1 PCT/CN2021/081674 CN2021081674W WO2022021901A1 WO 2022021901 A1 WO2022021901 A1 WO 2022021901A1 CN 2021081674 W CN2021081674 W CN 2021081674W WO 2022021901 A1 WO2022021901 A1 WO 2022021901A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
target
detection
detection network
parameters
Prior art date
Application number
PCT/CN2021/081674
Other languages
English (en)
French (fr)
Inventor
刘李洋
王波超
旷章辉
陈益民
张伟
Original Assignee
深圳市商汤科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市商汤科技有限公司 filed Critical 深圳市商汤科技有限公司
Publication of WO2022021901A1 publication Critical patent/WO2022021901A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Definitions

  • the present disclosure relates to the field of computer technology, and in particular, to a target detection method and device, an electronic device, and a storage medium.
  • target detection relies on large-scale training data, requiring a lot of manpower and material resources to collect and label the training data, and the more categories of objects, the higher the labeling cost.
  • data collection is also difficult, resulting in a small number of sample images.
  • the requirements are often dynamic, and detection categories may need to be dynamically increased, and the number of sample images of the increased categories may be small.
  • the present disclosure provides a target detection method and device, an electronic device and a storage medium.
  • a target detection method comprising:
  • the parameters of the detection network of the target category are obtained by inputting the training images of the target category into the parameter generation network.
  • the method further includes:
  • each target training set includes training images of K categories, each category includes M training images, and K is an integer greater than 0;
  • the parameter generation network is trained.
  • the parameters of the detection network can be easily obtained, and then a detection network with a smaller number of samples can be easily constructed.
  • the M training images include N support images and O query images, where N and O are integers greater than 0;
  • the parameter generation network is trained based on each target training set, include:
  • each support image of the target training set into the parameter generation network to be trained obtain the parameters of the general detection network of the target training set, and construct the general detection network of the target training set according to the parameters of the general detection network;
  • the parameter generation network to be trained is trained according to the detection loss of the general detection network.
  • the described each support image of the target training set is input to the parameter generation network to be trained, and the parameters of the general detection network of the target training set are obtained, including:
  • the parameters of the general detection network of the target training set are determined.
  • the accuracy of the parameters of the general detection network can be improved.
  • the method further includes:
  • the feature extraction network to be trained is trained according to the detection loss of the general detection network.
  • the feature discrimination ability of the feature extraction network can be improved.
  • the training of the feature extraction network to be trained according to the detection loss of the general detection network includes:
  • the feature extraction network to be trained is trained according to the detection loss of the general detection network and the detection loss of the reference detection network.
  • the feature extraction network can be trained by the detection loss of the general detection network and the detection loss of the reference detection network, so that the guidance of many samples to few samples can be realized, and the feature discrimination ability of the features extracted by the feature extraction network can be further improved.
  • the parameters of the reference detection network of the target training set are obtained, including:
  • the parameters of the trained detection network are determined as the reference detection network of the target training set.
  • the general detection network can be guided, so that the general detection network obtained by the few-sample training is closer to the reference detection network obtained by the multi-sample training, and the loss caused by the few samples can be reduced.
  • the training of the parameter generation network to be trained according to the detection loss of the general detection network includes:
  • the parameters of the general detection network of the target training set and the parameters of the reference detection network of the target training set determine the gap loss of the general detection network
  • the parameters of the to-be-trained parameter generation network are trained according to the detection loss and gap loss of the general detection network.
  • the method further includes:
  • the parameter generation network to be trained is trained according to the orthogonalization loss of the general detection network.
  • the construction of the detection network of the target category includes:
  • Each training image of the target category is respectively input into the parameter generation network to obtain the parameters of the detection network corresponding to each training sample of the target category;
  • the detection network of the target category is constructed.
  • a target detection device comprising:
  • building blocks configured to build detection networks for target classes
  • a detection module configured to use the detection network of the target category to detect the image to be detected, and obtain a target detection result of the image to be detected;
  • the parameters of the detection network of the target category are obtained from the input parameter generation network based on the training image of the target category.
  • the apparatus further includes:
  • an acquisition module configured to acquire one or more target training sets from the image set, wherein each target training set includes training images of K categories, each category includes M training images, and K is an integer greater than 0;
  • the first training module is configured to train the parameter generation network based on each target training set.
  • the M training images include N support images and O query images, where N and O are integers greater than 0; the first training module is further configured as:
  • each support image of the target training set into the parameter generation network to be trained obtain the parameters of the general detection network of the target training set, and construct the general detection network of the target training set according to the parameters of the general detection network;
  • the parameter generation network to be trained is trained according to the detection loss of the general detection network.
  • the first training module is further configured to:
  • the parameters of the general detection network of the target training set are determined.
  • the apparatus further includes:
  • the second training module is configured to train the feature extraction network to be trained according to the detection loss of the general detection network.
  • the second training module is further configured to:
  • the feature extraction network to be trained is trained according to the detection loss of the general detection network and the detection loss of the reference detection network.
  • the parameters of the reference detection network of the target training set are obtained, including:
  • the parameters of the trained detection network are determined as the reference detection network of the target training set.
  • the first training module is further configured to:
  • the parameters of the general detection network of the target training set and the parameters of the reference detection network of the target training set determine the gap loss of the general detection network
  • the parameters of the to-be-trained parameter generation network are trained according to the detection loss and gap loss of the general detection network.
  • the apparatus further includes:
  • a determination module configured to determine the orthogonalization loss of the universal detection network
  • the third training module is configured to train the parameter generation network to be trained according to the orthogonalization loss of the general detection network.
  • the building module is further configured to:
  • the detection network of the target category is constructed.
  • an electronic device comprising: a processor; a memory configured to store instructions executable by the processor; wherein the processor is configured to invoke the instructions stored in the memory to execute the above method.
  • a computer-readable storage medium having computer program instructions stored thereon, the computer program instructions implementing the above method when executed by a processor.
  • the parameters of the detection network of the target category can be obtained through the parameter generation network, and then the detection network of the target category can be constructed according to the parameters, so as to realize the target detection of the target category .
  • the labeling cost of training images is reduced, and the risk of overfitting caused by directly training the detection network with a small number of training images is reduced.
  • the embodiments of the present disclosure are conducive to dynamically adding new categories.
  • FIG. 1 shows a flowchart of a target detection method according to an embodiment of the present disclosure
  • FIG. 2 shows a schematic diagram of a network architecture according to an embodiment of the present disclosure
  • FIG. 3 shows a schematic diagram of a network architecture according to an embodiment of the present disclosure
  • FIG. 4 shows a block diagram of a target detection apparatus according to an embodiment of the present disclosure
  • FIG. 5 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure
  • FIG. 6 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • Object detection is a classic problem in computer vision. It mainly determines whether an image contains a certain type of object. If it does, the position of each object needs to be given. Object detection is the cornerstone of image content understanding and the basis for many more complex visual understanding tasks, such as tracking recognition, instance segmentation, scene classification, and event detection. With the development of technology, object detection has been widely used in real life, such as face recognition, automatic driving, security control and entertainment interaction. Typically, for a detection network to learn a new class, a large number of images of that class are required. However, in practical applications, a large number of images of new categories may not be obtained.
  • the target detection method provided by the embodiment of the present disclosure can construct a relatively accurate detection network for detecting the target of the small bird based on a small number of images containing the small bird. In this way, object detection can be performed on images of a large number of birds to determine whether such small birds have appeared.
  • FIG. 1 shows a flowchart of a target detection method according to an embodiment of the present disclosure.
  • the target detection method may include:
  • Step S11 construct a detection network of the target category.
  • Step S12 using the detection network of the target category to detect the image to be detected, to obtain a target detection result of the image to be detected.
  • the parameters of the detection network of the target category are obtained by inputting the training images of the target category into the parameter generation network.
  • the parameters of the detection network of the target category can be obtained through the parameter generation network, and then the detection network of the target category can be constructed according to the parameters, so as to realize the target detection of the target category .
  • the labeling cost of training images is reduced, and the risk of overfitting caused by directly training the detection network with a small number of training images is reduced.
  • the embodiments of the present disclosure are conducive to dynamically adding new categories.
  • the target category can also be a category with a large number of training images
  • the target detection method provided by the embodiment of the present disclosure can also be applied to a category with a large number of training images.
  • the target detection method may be performed by an electronic device such as a terminal device or a server
  • the terminal device may be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless Telephone, personal digital assistant (Personal Digital Assistant, PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc.
  • the method can be implemented by the processor calling the computer-readable instructions stored in the memory.
  • the method may be performed by a server.
  • the target category may represent a category to be subjected to target detection.
  • the target class may be a class with a small number of training images, eg, the target class may be a class with one or several training images.
  • the target category can be a dynamically added new category.
  • a detection network may represent a network for object detection.
  • the structure of the detection network may be a network capable of performing anchor-free object detection, such as an FCOS (Full Convolutional One Stage Object Detection) network.
  • FCOS Full Convolutional One Stage Object Detection
  • the embodiments of the present disclosure do not limit the detection network.
  • the detection network for object classes may represent a network for object detection on object classes. That is to say, the detection network of the target category can detect whether there is an object of the target category in the image to be detected.
  • the parameters of the detection network of the target category may be obtained first, and then the detection network of the target category may be constructed based on the parameters of the detection network of the target category. Among them, the parameters of the detection network of the target category are obtained from the input parameter generation network based on the training image of the target category.
  • the parameter generation network can be used to generate the parameters of the detection network.
  • the parameter generation network takes the training image as input and the parameter of the detection network as the output.
  • the training image of the target category is input into the parameter generation network, and the parameters of the detection network of the target category can be obtained.
  • the embodiments of the present disclosure do not limit the structure of the parameter generation network.
  • the process of constructing the detection network is as follows: first, acquiring training images of the target category, and inputting the training images of the target category into the parameter generation network respectively, to obtain the target category
  • the parameters of the detection network corresponding to each training sample of The parameters of the detection network are constructed, and the detection network of the target category is constructed.
  • each training image of the target category can be input into the parameter generation network respectively, and the parameters of the detection network corresponding to each training image of the target category can be obtained. Since these training images all belong to the target category, the parameters of the detection network of the target category can be determined according to the parameters of the detection network corresponding to these training images. In one example, the parameters of the detection network corresponding to the training images of the target category may be averaged, and the averaged parameters of the detection network may be determined as the parameters of the detection network of the target category.
  • the weight information of each training image of the target category may be determined according to information such as the position or size of the target area (the area where the object of the target category is located) in the training image; then, based on the weight information, the target category
  • the parameters of the detection network corresponding to each training image are weighted and averaged, and the parameters of the detection network after the weighted average are determined as the parameters of the detection network of the target category.
  • a corresponding detection network can be constructed based on the structure of the detection network. That is, after acquiring the parameters of the detection network of the target category, the detection network of the target category can be constructed based on the structure of the detection network.
  • the parameters may be set as the detection network of the detection network of the target category, and the detection network may be directly determined as the detection network of the target category. In this way, after inputting the training images of the target category into the parameter generation network, the detection network of the target category can be obtained conveniently and quickly.
  • the detection network whose parameters are set as the parameters of the detection network of the target category can be determined as the initialization detection network of the target category; then, the initialization detection network is fine-tuned to obtain the detection network of the target category.
  • the initialized detection network can be fine-tuned by loss minimization.
  • the loss can include detection loss and quadrature loss to initialize the detection network.
  • the detection loss of the initialized detection network can be determined according to the predicted label distribution result and the corresponding ground-truth label output after the initialized detection network is input to the training image of the target category.
  • the image to be detected may be input into the detection network of the target category to obtain the target detection result of the image to be detected.
  • the target detection result may include the probability that the image to be detected is of the target category and the position information of the object of the target category in the image to be detected.
  • the network is first generated based on the parameters to obtain the parameters of the detection network of the target category, and then the detection network of the target category is constructed according to the parameters of the detection network of the target category, thereby realizing the target detection on the target category.
  • the parameter generation network is an important tool of the target detection method of the embodiment of the present disclosure. The training process of the parameter generation network is described below.
  • the training process of the parameter generation network may include: acquiring one or more target training sets from an image set; and training the parameter generation network based on each target training set.
  • the image set includes training images of C categories (called C base categories), and each category includes A training images as an example for illustration.
  • the process of obtaining a target training set from the image set may include: randomly selecting K categories from C categories, and randomly selecting M training images from A training images of each category.
  • the target training set includes training images of K categories, and each category includes M training images.
  • C, A, K, and M are integers greater than 0, and C>K, A>M.
  • the number of K and M can be set as required. Since the target detection method of the embodiment of the present disclosure needs to solve the detection problem of a category with a small number of training images, the embodiment of the present disclosure adopts a small number of categories when training the parameter generation network, and the training images of each category are The number is also less.
  • K may be 5; M may be 11, 15, or 20, etc.
  • the corresponding C can be 1000 or 2000, etc., and A can be 5000 or 10000. It can be understood that the process of generating the network for training parameters provided by the embodiments of the present disclosure is also applicable to a large number of categories, therefore, the number of training images for each category may be larger, and M may also be 500 or 1000.
  • the M training images included in the category may include N support images and O query images, where N and O are integers greater than 0, and M ⁇ N+O.
  • N training images can be randomly selected from the M training images of this category as support images, and the remaining training images of this category can be used as Query images.
  • M>N+O for each category of the target training set, N training images can be randomly selected from the M training images of the category as support images, and randomly selected from the remaining training images of the category. Choose O training images as query images.
  • the following takes a target training set as an example to illustrate the training process of the parameter generation network.
  • the process of using multiple target training sets to train parameters to generate a network is actually a process of repeatedly using one target training set to train parameters to generate a network, which will not be repeated here.
  • training a parameter generation network based on a target training set may include: first, inputting each support image of the target training set into a parameter generation network to be trained to obtain a general detection method of the target training set Then, input each query image of the target training set into the feature extraction network to be trained to obtain each query image of the target training set image feature map; thirdly, input the feature map of each query image into the general detection network respectively, to obtain the predicted label distribution result of each query image; finally, according to the predicted label distribution result of each query image and ground truth label, determine the detection loss of the general detection network, and train the parameter generation network to be trained according to the detection loss of the general detection network.
  • inputting each support image of the target training set into the parameter generation network to be trained, and obtaining the parameters of the general detection network of the target training set may include: inputting each support image of the target training set into the parameter generation network to be trained respectively network to obtain the parameters of the detection network corresponding to each support image; according to the parameters of the detection network corresponding to each support image and the real category of each support image, determine the parameters of the detection network of each category of the target training set; and according to the The parameters of the detection network of each category of the target training set are determined, and the parameters of the general detection network of the target training set are determined.
  • the parameters of the detection network corresponding to the support images of the same category may be averaged or weighted averaged according to the real categories of the support images (the weights may be determined according to information such as the position or size of the target area in the support images), Get the parameters of the detection network for the corresponding category. Then, the parameters of each category of detection network are spliced into the parameters of the general detection network of the target training set.
  • FIG. 2 shows a schematic diagram of a network architecture according to an embodiment of the present disclosure.
  • the network architecture 200 includes a parameter generation network f ⁇ 201 and a feature extraction network 202g ⁇ .
  • the parameter of the parameter generation network f ⁇ is ⁇
  • the parameter of the feature extraction network g ⁇ is ⁇ .
  • the support set D s includes K categories of support images, and each category includes N support images.
  • x s represents the target region in the support image
  • y s represents the ground-truth label of x s
  • (x s , y s ) i represents the target region and ground-truth label of the ith support image in the support set D s , 1 ⁇ i ⁇ K*N, in, represents the category of x s
  • the query set Dq includes K categories of query images, and each category includes O query images.
  • x q represents the target area in the query image
  • y q represents the ground-truth label of x q
  • (x q , y q ) j represents the target area and ground-truth label of the j-th query image in the query set D q , 1 ⁇ j ⁇ K*O. in, represents the category of x q
  • the process of training the parameter generation network may include:
  • each support image in the support set D s Crop each support image in the support set D s to obtain the target area x s of each support image (in an example, the size of the target area may be 224 pixels*224 pixels), and input the target area x s of each support image
  • the parameters of the detection network corresponding to each support image can be obtained, and the parameters of the detection network corresponding to the support images x s of the same category can be averaged (or weighted average) to obtain the parameters of this category.
  • Equation (1) shows the parameters of the detection network of class k:
  • D represents the dimension of the parameters of the detection network, represents the parameters of the detection network of class k; That is to say, the category of the detection network is consistent with the category of the supporting image.
  • the parameters of the general detection network can be obtained. Then according to the parameters of the general detection network A general detection network for the target training set can be constructed.
  • Input feature extraction network g ⁇ the feature map g ⁇ (x q ) of each query image is obtained.
  • the input parameters of the feature map g ⁇ (x q ) of each query image are In the general detection network of , the predicted label distribution results of each query image can be obtained. According to the predicted label distribution results of each query image and the ground-truth label y q , the detection loss of the general detection network can be obtained. In one example, the detection loss of the general detection network can be obtained by Eq. (2).
  • L d represents the detection loss of the general detection network
  • loss((7) represents the loss function. It is represented by the ground-truth label y q of the query image and the feature map g ⁇ (x q ) of the query image.
  • the input parameters are The predicted label distribution obtained in the general detection network results in a loss function of the parameters.
  • the structure of the loss function is not limited in the embodiments of the present disclosure, for example, it may be a mean square error function, a cross entropy function, or the like.
  • the parameter ⁇ of the parameter generation network f ⁇ is adjusted to realize the training of the parameter generation network f ⁇ .
  • the parameter generation network f ⁇ trained with a small number of samples can be used to generate the parameters of the detection network of a new category, and has the potential to transfer the generation ability of the detection network to the new category.
  • the method further includes: training the feature extraction network to be trained according to the detection loss of the general detection network.
  • the feature extraction network g ⁇ can be trained at the same time. That is, the parameter ⁇ of the feature extraction network g ⁇ can also be updated with the goal of minimizing the detection loss of the general detection network.
  • training the feature extraction network to be trained according to the detection loss of the general detection network includes: acquiring parameters of the reference detection network of the target training set; With reference to the parameters of the detection network, a reference detection network of the target training set is constructed; the feature maps of the query images are respectively input into the reference detection network to obtain the reference label distribution results of the query images; Query the reference label distribution results and true value labels of the image to determine the detection loss of the reference detection network; train the feature extraction network to be trained according to the detection loss of the general detection network and the detection loss of the reference detection network .
  • the reference detection network may be used to represent the detection network obtained by training based on training images of all categories of the image set.
  • the distinguishing ability of the trained feature extraction network g ⁇ will be limited to the categories involved in each target training set, which will weaken its ability to extract features.
  • the number of training images involved in the training process is small.
  • the general detection network trained with a small number of training images has weaker target detection ability than the detection network obtained with a large number of training images. Therefore, in the embodiment of the present disclosure, a reference detection network obtained by training a large number of training images of various categories is introduced, and the training of the parameter generation network f ⁇ and the feature extraction network g ⁇ is optimized.
  • FIG. 3 shows a schematic diagram of a network architecture according to an embodiment of the present disclosure.
  • the network architecture shown in FIG. 3 adds a reference detection network 301 whose parameter is ⁇ on the basis of FIG. 2 .
  • the feature map g ⁇ (x q ) of each query image is input into the reference detection network whose parameter is ⁇ , and the distribution result of the reference label of each query image can be obtained.
  • the detection loss of the reference detection network can be obtained.
  • the detection loss of the reference detection network can be obtained by formula (3).
  • L r represents the detection loss of the reference detection network
  • loss((7) represents the loss function.
  • the reference label distribution obtained in the reference detection network with the input parameter ⁇ is the loss function of the parameters.
  • the structure of the loss function is not limited in the embodiments of the present disclosure, for example, it may be a mean square error function, a cross entropy function, or the like.
  • the reference detection network is trained based on training images of all categories, jointly training the feature extraction network to be trained according to the detection loss of the general detection network and the detection loss of the reference detection network can improve the The feature discrimination ability of the feature extraction network.
  • obtaining the parameters of the reference detection network of the target training set may include: obtaining a randomly initialized detection network; training the randomly initialized detection network based on all query images of the target training set; The parameters of the trained detection network are determined as the reference detection network of the target training set.
  • a detection network is randomly initialized as the detection network to be trained, and then based on all the query images of the target training set, the detection network to be trained is obtained, and the reference detection network of the target training set is obtained.
  • the reference detection network and the parameter generation network f ⁇ and the feature extraction network g ⁇ of the target training set can be trained simultaneously.
  • the parameters of the reference detection network of the target training set are also obtained by splicing the parameters of the detection networks of the K categories. Based on the parameters of the reference detection network of the target training set, the reference detection network of the target training set can be constructed. For example, suppose the parameters of the reference detection network for class k among the K classes of the target training set are Among them, D is the dimension of the parameters of the reference detection network. By splicing the parameters of the reference detection network of K categories, the parameters of the reference detection network of the target training set can be obtained
  • a data set including K categories may also be reconstructed for training the reference detection network.
  • reference may be made to the above-mentioned training process using query images, which will not be repeated here.
  • training the parameter generation network to be trained includes: according to the parameters of the general detection network of the target training set and the reference detection of the target training set The parameters of the network are used to determine the gap loss of the general detection network; according to the detection loss and gap loss of the general detection network, the parameters to be trained are trained to generate the parameters of the network.
  • the gap loss of the general detection network can be obtained by Equation (4) or Equation (5).
  • ⁇ is a conditional function. When the condition in parentheses is true, the value is 1, and when the condition in parentheses is false, the value is 0.
  • 1 indicates the first-order norm
  • 2 indicates the second-order norm.
  • f ⁇ (x s ) and ⁇ c represent the parameters of the generic detection network and the reference detection network corresponding to category c, respectively.
  • the reference detection network is trained based on all categories of training images, according to the detection loss and gap loss of the general detection network, jointly training the parameters of the parameter generation network to be trained can make the generation network based on the parameters The accuracy of the detection network obtained by the network is higher.
  • the method may further include: determining an orthogonalization loss of the general detection network; and training the parameter generation network to be trained according to the orthogonalization loss of the general detection network.
  • the orthogonalization loss of the generic detection network can be determined by Equation (6).
  • the distinguishing ability of the model can be improved.
  • the parameter generation network f ⁇ , the feature extraction network g ⁇ and the reference detection network shown in FIG. 3 can be trained simultaneously. Therefore, in the embodiment of the present disclosure, a total training loss can be determined by formula (7).
  • L represents the total training loss
  • L d represents the detection loss of the general detection network (see Equation (2))
  • L r represents the detection loss of the reference detection network (see Equation (3))
  • L g represents the detection loss of the general detection network Gap loss (see Equation (4) and Equation (5))
  • L o represents the Orthogonal Loss of the generic detection network (see Equation (6)).
  • ⁇ and ⁇ are hyperparameters. ⁇ and ⁇ can be set as required. In one example, ⁇ can take 0.01 and ⁇ can take 1.
  • the parameter generation network f ⁇ , the feature extraction network g ⁇ and the reference detection network can be trained simultaneously based on L, and the parameters ⁇ , ⁇ and ⁇ can be adjusted.
  • the present disclosure also provides target detection devices, electronic devices, computer-readable storage media, and programs, all of which can be used to implement any target detection method provided by the present disclosure.
  • target detection devices electronic devices, computer-readable storage media, and programs, all of which can be used to implement any target detection method provided by the present disclosure.
  • FIG. 4 shows a block diagram of a target detection apparatus according to an embodiment of the present disclosure.
  • the device 40 includes:
  • the building module 41 is configured to build a detection network of the target category
  • the detection module 42 is configured to use the detection network of the target category to detect the image to be detected, and obtain the target detection result of the image to be detected;
  • the parameters of the detection network of the target category are obtained from the input parameter generation network based on the training image of the target category.
  • the apparatus further includes:
  • an acquisition module configured to acquire one or more target training sets from the image set, wherein each target training set includes training images of K categories, each category includes M training images, and K is an integer greater than 0;
  • the first training module is configured to train the parameter generation network based on each target training set.
  • the M training images include N support images and O query images, where N and O are integers greater than 0; the first training module is further configured as:
  • each support image of the target training set into the parameter generation network to be trained obtain the parameters of the general detection network of the target training set, and construct the general detection network of the target training set according to the parameters of the general detection network;
  • the parameter generation network to be trained is trained according to the detection loss of the general detection network.
  • each support image of the target training set is input into the parameter generation network to be trained, and the parameters of the general detection network of the target training set are obtained, including:
  • the parameters of the general detection network of the target training set are determined.
  • the apparatus further includes:
  • the second training module is configured to train the feature extraction network to be trained according to the detection loss of the general detection network.
  • the second training module is further configured to:
  • the feature extraction network to be trained is trained according to the detection loss of the general detection network and the detection loss of the reference detection network.
  • the parameters of the reference detection network of the target training set are obtained, including:
  • the parameters of the trained detection network are determined as the reference detection network of the target training set.
  • training the parameter generation network to be trained includes:
  • the parameters of the general detection network of the target training set and the parameters of the reference detection network of the target training set determine the gap loss of the general detection network
  • the parameters of the to-be-trained parameter generation network are trained according to the detection loss and gap loss of the general detection network.
  • the apparatus further includes:
  • a determination module configured to determine the orthogonalization loss of the universal detection network
  • the third training module is configured to train the parameter generation network to be trained according to the orthogonalization loss of the general detection network.
  • the building module is further configured to:
  • the detection network of the target category is constructed.
  • the functions or modules included in the apparatuses provided in the embodiments of the present disclosure may be used to execute the methods described in the above method embodiments.
  • Embodiments of the present disclosure further provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure further provides an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to invoke the instructions stored in the memory to execute the above method.
  • Embodiments of the present disclosure also provide a computer program product, including computer-readable codes.
  • a processor in the device executes a method for implementing the target detection method provided in any of the above embodiments. instruction.
  • Embodiments of the present disclosure further provide another computer program product for storing computer-readable instructions, which, when executed, cause the computer to perform the operations of the target detection method provided by any of the foregoing embodiments.
  • the electronic device may be provided as a terminal, server or other form of device.
  • FIG. 5 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • electronic device 800 may be a mobile phone, computer, digital broadcast terminal, messaging device, game console, tablet device, medical device, fitness device, personal digital assistant, etc. terminal.
  • electronic device 800 may include one or more of the following components: processing component 802, memory 804, power supply component 806, multimedia component 808, audio component 810, input/output (I/O) interface 812, sensor component 814 , and the communication component 816 .
  • the processing component 802 generally controls the overall operation of the electronic device 800, such as operations associated with display, phone calls, data communications, camera operations, and recording operations.
  • the processing component 802 can include one or more processors 820 to execute instructions to perform all or some of the steps of the methods described above.
  • processing component 802 may include one or more modules that facilitate interaction between processing component 802 and other components.
  • processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.
  • Memory 804 is configured to store various types of data to support operation at electronic device 800 . Examples of such data include instructions for any application or method operating on electronic device 800, contact data, phonebook data, messages, pictures, videos, and the like. Memory 804 may be implemented by any type of volatile or nonvolatile storage device or combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic or Optical Disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read only memory
  • EPROM erasable Programmable Read Only Memory
  • PROM Programmable Read Only Memory
  • ROM Read Only Memory
  • Magnetic Memory Flash Memory
  • Magnetic or Optical Disk Magnetic Disk
  • Power supply assembly 806 provides power to various components of electronic device 800 .
  • Power supply components 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 800 .
  • Multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user.
  • the touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundaries of a touch or swipe action, but also detect the duration and pressure associated with the touch or swipe action.
  • the multimedia component 808 includes a front-facing camera and/or a rear-facing camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera may receive external multimedia data. Each of the front and rear cameras can be a fixed optical lens system or have focal length and optical zoom capability.
  • Audio component 810 is configured to output and/or input audio signals.
  • audio component 810 includes a microphone (MIC) that is configured to receive external audio signals when electronic device 800 is in operating modes, such as calling mode, recording mode, and voice recognition mode.
  • the received audio signal may be further stored in memory 804 or transmitted via communication component 816 .
  • audio component 810 also includes a speaker for outputting audio signals.
  • the I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module, which may be a keyboard, a click wheel, a button, or the like. These buttons may include, but are not limited to: home button, volume buttons, start button, and lock button.
  • Sensor assembly 814 includes one or more sensors for providing status assessment of various aspects of electronic device 800 .
  • the sensor assembly 814 can detect the on/off state of the electronic device 800, the relative positioning of the components, such as the display and the keypad of the electronic device 800, the sensor assembly 814 can also detect the electronic device 800 or one of the electronic device 800 Changes in the position of components, presence or absence of user contact with the electronic device 800 , orientation or acceleration/deceleration of the electronic device 800 and changes in the temperature of the electronic device 800 .
  • Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects in the absence of any physical contact.
  • Sensor assembly 814 may also include a light sensor, such as a complementary metal oxide semiconductor (CMOS) or charge coupled device (CCD) image sensor, for use in imaging applications.
  • CMOS complementary metal oxide semiconductor
  • CCD charge coupled device
  • the sensor assembly 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • Communication component 816 is configured to facilitate wired or wireless communication between electronic device 800 and other devices.
  • the electronic device 800 may access a wireless network based on a communication standard, such as wireless network (WiFi), second generation mobile communication technology (2G) or third generation mobile communication technology (3G), or a combination thereof.
  • the communication component 816 receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module may be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra-wideband
  • Bluetooth Bluetooth
  • electronic device 800 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A programmed gate array (FPGA), controller, microcontroller, microprocessor or other electronic component implementation is used to perform the above method.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field programmable A programmed gate array
  • controller microcontroller, microprocessor or other electronic component implementation is used to perform the above method.
  • a non-volatile computer-readable storage medium such as a memory 804 comprising computer program instructions executable by the processor 820 of the electronic device 800 to perform the above method is also provided.
  • FIG. 6 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • the electronic device 1900 may be provided as a server.
  • electronic device 1900 includes processing component 1922, which further includes one or more processors, and a memory resource represented by memory 1932 for storing instructions executable by processing component 1922, such as applications.
  • An application program stored in memory 1932 may include one or more modules, each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above-described methods.
  • the electronic device 1900 may also include a power supply assembly 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as a Microsoft server operating system (Windows Server TM ), a graphical user interface based operating system (Mac OS X TM ) introduced by Apple, a multi-user multi-process computer operating system (Unix TM ), Free and Open Source Unix-like Operating System (Linux TM ), Open Source Unix-like Operating System (FreeBSD TM ) or the like.
  • Microsoft server operating system Windows Server TM
  • Mac OS X TM graphical user interface based operating system
  • Uniix TM multi-user multi-process computer operating system
  • Free and Open Source Unix-like Operating System Linux TM
  • FreeBSD TM Open Source Unix-like Operating System
  • a non-volatile computer-readable storage medium such as memory 1932 comprising computer program instructions executable by processing component 1922 of electronic device 1900 to perform the above-described method.
  • the present disclosure may be a system, method and/or computer program product.
  • the computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present disclosure.
  • a computer-readable storage medium may be a tangible device that can hold and store instructions for use by the instruction execution device.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM) or flash memory), static random access memory (SRAM), portable compact disk read only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically coded devices, such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • RAM random access memory
  • ROM read only memory
  • EPROM erasable programmable read only memory
  • flash memory static random access memory
  • SRAM static random access memory
  • CD-ROM compact disk read only memory
  • DVD digital versatile disk
  • memory sticks floppy disks
  • mechanically coded devices such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • Computer-readable storage media are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (eg, light pulses through fiber optic cables), or through electrical wires transmitted electrical signals.
  • the computer readable program instructions described herein may be downloaded to various computing/processing devices from a computer readable storage medium, or to an external computer or external storage device over a network such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • a network adapter card or network interface in each computing/processing device receives computer-readable program instructions from a network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing/processing device .
  • Computer program instructions for carrying out operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or instructions in one or more programming languages.
  • Source or object code written in any combination, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as the "C" language or similar programming languages.
  • the computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server implement.
  • the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider through the Internet connect).
  • LAN local area network
  • WAN wide area network
  • custom electronic circuits such as programmable logic circuits, field programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) can be personalized by utilizing state information of computer readable program instructions.
  • Computer readable program instructions are executed to implement various aspects of the present disclosure.
  • These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer or other programmable data processing apparatus to produce a machine that causes the instructions when executed by the processor of the computer or other programmable data processing apparatus , resulting in means for implementing the functions/acts specified in one or more blocks of the flowchart and/or block diagrams.
  • These computer readable program instructions can also be stored in a computer readable storage medium, these instructions cause a computer, programmable data processing apparatus and/or other equipment to operate in a specific manner, so that the computer readable medium on which the instructions are stored includes An article of manufacture comprising instructions for implementing various aspects of the functions/acts specified in one or more blocks of the flowchart and/or block diagrams.
  • Computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other equipment to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other equipment to produce a computer-implemented process , thereby causing instructions executing on a computer, other programmable data processing apparatus, or other device to implement the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams.
  • each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more functions for implementing the specified logical function(s) executable instructions.
  • the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented in dedicated hardware-based systems that perform the specified functions or actions , or can be implemented in a combination of dedicated hardware and computer instructions.
  • the computer program product can be specifically implemented by hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (Software Development Kit, SDK), etc. Wait.
  • a software development kit Software Development Kit, SDK
  • the present disclosure provides a target detection method and device, an electronic device and a storage medium, wherein a detection network of the target category is constructed; the detection network of the target category is used to detect the image to be detected, and the target detection of the to-be-detected image is obtained. The result; wherein, the parameters of the detection network of the target category are obtained from the input parameter generation network based on the training image of the target category.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Debugging And Monitoring (AREA)
  • Signal Processing For Digital Recording And Reproducing (AREA)

Abstract

一种目标检测方法及装置、电子设备和存储介质，所述方法包括：构建目标类别的检测网络（S11）；采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果（S12）；其中，所述目标类别的检测网络的参数是将目标类别的训练图像输入参数生成网络中而得到的。所述方法有利于动态增加新类别。

Description

目标检测方法及装置、电子设备和存储介质
相关申请的交叉引用
本公开基于申请号为202010751150.1、申请日为2020年7月30日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本公开作为参考。
技术领域
本公开涉及计算机技术领域，尤其涉及一种目标检测方法及装置、电子设备和存储介质。
背景技术
在相关技术中，目标检测依赖于大规模的训练数据，需要大量的人力物力对训练数据进行搜集和标注，且对象的类别越多，标注成本也就越高。在一些特定的场景下，数据的搜集也很困难，从而造成样本图像的数量较少。且实际应用场景中，需求往往是动态变化的，可能需要动态地增加检测类别，而增加的类别的样本图像的数量可能较少。
发明内容
本公开提出了一种目标检测方法及装置、电子设备和存储介质。
根据本公开的一方面，提供了一种目标检测方法，包括：
构建目标类别的检测网络；
采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；
其中，所述目标类别的检测网络的参数是将目标类别的训练图像输入参数生成网络中而得到的。
在一种可能的实现方式中，所述方法还包括：
从图像集中获取一个或多个目标训练集，其中，每个目标训练集包括K个类别的训练图像，每个类别包括M个训练图像，K为大于0的整数；
基于各目标训练集，训练所述参数生成网络。
通过较少的样本训练参数生成网络，可以方便的获取到检测网络的参数，进而方便的构建出样本数量较少的类别的检测网络。
在一种可能的实现方式中，所述M个训练图像包括N个支持图像和O个查询图像，N和O为大于0的整数；所述基于各目标训练集，训练所述参数生成网络，包括：
针对每个目标训练集：
将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，并根据该通用检测网络的参数，构建该目标训练集的通用检测网络；
将该目标训练集的各查询图像输入待训练的特征提取网络，得到该目标训练集的各查询图像的特征图；
将所述各查询图像的特征图分别输入所述通用检测网络，得到所述各查询图像的预测标签分布结果；
根据所述各查询图像的预测标签分布结果和真值标签，确定所述通用检测网络的检测损失；
根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络。
通过通用检测网络的检测损失，可以快速实现收敛，从而快速完成参数生成网络的训练。
在一种可能的实现方式中，所述将该目标训练集的各支持图像输入待训练的参数生 成网络，得到该目标训练集的通用检测网络的参数，包括：
将该目标训练集的各支持图像分别输入待训练的参数生成网络，得到每个支持图像对应的检测网络的参数；
根据各支持图像对应的检测网络的参数和各支持图像的真实类别，确定该目标训练集的每个类别的检测网络的参数；
根据该目标训练集的各类别的检测网络的参数，确定该目标训练集的通用检测网络的参数。
通过基于多个支持图像获取目标训练集的通用检测网络的参数，可以提高通用检测网络的参数的准确性。
在一种可能的实现方式中，所述方法还包括：
根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络。
通过通用检测网络的检测损失训练特征提取网络，可以提高特征提取网络的特征区分能力。
在一种可能的实现方式中，所述根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络，包括：
获取目标训练集的参考检测网络；
将所述各查询图像的特征图分别输入所述参考检测网络，得到所述各查询图像的参考标签分布结果；
根据所述各查询图像的参考标签分布结果和真值标签，确定所述参考检测网络的检测损失；
根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，训练所述待训练的特征提取网络。
通过通用检测网络的检测损失和所述参考检测网络的检测损失训练特征提取网络，可以实现多样本对少样本的指导，进一步提升特征提取网络提取特征的特征区分能力。
在一种可能的实现方式中，获取该目标训练集的参考检测网络的参数，包括：
获取随机初始化的检测网络；
基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练；
将训练完成的检测网络的参数，确定为该目标训练集的参考检测网络。
通过获取参考检测网络，可以对通用检测网络进行指导，使得少样本训练得到的通用检测网络更加接近多样本训练得到的参考检测网络，缩小少样本带来的损失。
在一种可能的实现方式中，所述根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络，包括：
根据该目标训练集的通用检测网络的参数和该目标训练集的参考检测网络的参数，确定所述通用检测网络的差距损失；
根据所述通用检测网络的检测损失和差距损失，训练所述待训练的参数生成网络的参数。
这样，根据所述通用检测网络的检测损失和差距损失，共同训练所述待训练的参数生成网络的参数，可以使基于参数生成网络得到的检测网络的准确度更高。
在一种可能的实现方式中，所述方法还包括：
确定所述通用检测网络的正交化损失；
根据所述通用检测网络的正交化损失，训练所述待训练的参数生成网络。
通过使不同类别的检测网络之间彼此正交，可以提升模型的区分能力。
在一种可能的实现方式中，所述构建目标类别的检测网络，包括：
获取所述目标类别的训练图像；
将所述目标类别的各训练图像分别输入所述参数生成网络中，得到所述目标类别的 每个训练样本对应的检测网络的参数；
根据所述目标类别的每个训练样本对应的检测网络的参数，确定所述目标类别的检测网络的参数；
根据所述目标类别的检测网络的参数，构建所述目标类别的检测网络。
根据本公开的一方面，提供了一种目标检测装置，包括：
构建模块，配置为构建目标类别的检测网络；
检测模块，配置为采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；
其中，所述目标类别的检测网络的参数是基于目标类别的训练图像输入参数生成网络中而得到的。
在一种可能的实现方式中，所述装置还包括：
获取模块，配置为从图像集中获取一个或多个目标训练集，其中，每个目标训练集包括K个类别的训练图像，每个类别包括M个训练图像，K为大于0的整数；
第一训练模块，配置为基于各目标训练集，训练所述参数生成网络。
在一种可能的实现方式中，所述M个训练图像包括N个支持图像和O个查询图像，N和O为大于0的整数；所述第一训练模块还配置为：
针对每个目标训练集：
将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，并根据该通用检测网络的参数，构建该目标训练集的通用检测网络；
将该目标训练集的各查询图像输入待训练的特征提取网络，得到该目标训练集的各查询图像的特征图；
将所述各查询图像的特征图分别输入所述通用检测网络，得到所述各查询图像的预测标签分布结果；
根据所述各查询图像的预测标签分布结果和真值标签，确定所述通用检测网络的检测损失；
根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络。
在一种可能的实现方式中，所述第一训练模块还配置为：
将该目标训练集的各支持图像分别输入待训练的参数生成网络，得到每个支持图像对应的检测网络的参数；
根据各支持图像对应的检测网络的参数和各支持图像的真实类别，确定该目标训练集的每个类别的检测网络的参数；
根据该目标训练集的各类别的检测网络的参数，确定该目标训练集的通用检测网络的参数。
在一种可能的实现方式中，所述装置还包括：
第二训练模块，配置为根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络。
在一种可能的实现方式中，所述第二训练模块还配置为：
获取目标训练集的参考检测网络；
将所述各查询图像的特征图分别输入所述参考检测网络，得到所述各查询图像的参考标签分布结果；
根据所述各查询图像的参考标签分布结果和真值标签，确定所述参考检测网络的检测损失；
根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，训练所述待训练的特征提取网络。
在一种可能的实现方式中，获取该目标训练集的参考检测网络的参数，包括：
获取随机初始化的检测网络；
基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练；
将训练完成的检测网络的参数，确定为该目标训练集的参考检测网络。
在一种可能的实现方式中，所述第一训练模块还配置为：
根据该目标训练集的通用检测网络的参数和该目标训练集的参考检测网络的参数，确定所述通用检测网络的差距损失；
根据所述通用检测网络的检测损失和差距损失，训练所述待训练的参数生成网络的参数。
在一种可能的实现方式中，所述装置还包括：
确定模块，配置为确定所述通用检测网络的正交化损失；
第三训练模块，配置为根据所述通用检测网络的正交化损失，训练所述待训练的参数生成网络。
在一种可能的实现方式中，所述构建模块还配置为：
获取所述目标类别的训练图像；
将所述目标类别的各训练图像分别输入所述参数生成网络中，得到所述目标类别的每个训练样本对应的检测网络的参数；
根据所述目标类别的每个训练样本对应的检测网络的参数，确定所述目标类别的检测网络的参数；
根据所述目标类别的检测网络的参数，构建所述目标类别的检测网络。
根据本公开的一方面，提供了一种电子设备，包括：处理器；配置为存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。
在本公开实施例中，对于训练图像数量较少的目标类别，可以先通过参数生成网络得到目标类别的检测网络的参数，然后根据该参数构建目标类别的检测网络，从而实现目标类别的目标检测。这样，既降低了训练图像的标注成本，又降低了采用少量训练图像直接训练检测网络而带来的过拟合的风险。进一步的，本公开实施例有利于动态增加新的类别。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的目标检测方法的流程图；
图2示出根据本公开实施例的网络架构示意图；
图3示出根据本公开实施例的网络架构示意图；
图4示出根据本公开实施例的目标检测装置的框图；
图5示出根据本公开实施例的一种电子设备800的框图；
图6示出根据本公开实施例的一种电子设备1900的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的 附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
目标检测是计算机视觉里面的经典问题，主要判断图像中是否包含某一类对象，如果包含还需要给出每个对象的位置。目标检测是图像内容理解的基石，是很多更加复杂视觉理解任务的基础，如跟踪识别、实例分割、场景分类和事件检测等。随着技术的发展，目标检测在实际生活中有着广泛的引用，如人脸识别、自动驾驶、安防布控和娱乐互动等。通常来说，让检测网络学习新的类别，需要大量该类别的图像。然而在实际应用中，可能并不能获取到大量新类别的图像。例如，摄影师拍摄到一种珍惜的小鸟(或者罕见的场景、深海动物等)后，科研人员在研究的过程中可能需要从大量鸟类的图像中，检测是否出现过该类小鸟。此时，如果以人工的方式去确认大量鸟类的图像中是否出现过该类小鸟，非常费时费力。因此，需要通过一个检测网络进行该类小鸟的目标检测。由于这类小鸟的数量较少，已确认包含该类小鸟的图像也较少，因此无法直接通过已确认包含该类小鸟的图片，训练出能够准确对该类小鸟进行目标检测的检测网络。而本公开实施例提供的目标检测方法，可以基于少量的包含该类小鸟的图像，构建出较为准确的对该类小鸟进行目标检测的检测网络。这样，就可以对大量鸟类的图像进行目标检测，确定是否出现过该类小鸟。
图1示出根据本公开实施例的目标检测方法的流程图。如图1所示，所述目标检测方法可以包括：
步骤S11，构建目标类别的检测网络。
步骤S12，采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果。
其中，所述目标类别的检测网络的参数是将目标类别的训练图像输入参数生成网络中而得到的。
在本公开实施例中，对于训练图像数量较少的目标类别，可以先通过参数生成网络得到目标类别的检测网络的参数，然后根据该参数构建目标类别的检测网络，从而实现目标类别的目标检测。这样，既降低了训练图像的标注成本，又降低了采用少量训练图像直接训练检测网络而带来的过拟合的风险。进一步的，本公开实施例有利于动态增加新的类别。
可以理解的是，目标类别也可以为具有训练图像数量较多的类别，本公开实施例提供的目标检测方法同样可以适用于具有训练图像数量较多的类别。
在一种可能的实现方式中，所述目标检测方法可以由终端设备或服务器等电子设备执行，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、 蜂窝电话、无绳电话、个人数字处理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等，所述方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。或者，可通过服务器执行所述方法。
在步骤S11中，目标类别可以表示待进行目标检测的类别。在一个示例中，目标类别可以为具有训练图像数量较少的类别，例如，目标类别可以为具有一个或几个训练图像的类别。在实际应用场景中，目标类别可以为动态增加的新类别。
检测网络可以表示用于进行目标检测的网络。在一个示例中，检测网络的结构可以为能够进行无候选框(Anchor-Free)的目标检测的网络，例如FCOS(Full Convolutional One Stage Object Detection，全卷积一级目标检测)网络。本公开实施例对检测网络不做限制。
目标类别的检测网络可以表示用于对目标类别进行目标检测的网络。也就是说，通过目标类别的检测网络可以检测出待检测图像中是否存在目标类别的对象。在本公开实施例中，可以首先获取目标类别的检测网络的参数，然后基于该目标类别的检测网络的参数，构建目标类别的检测网络。其中，目标类别的检测网络的参数是基于目标类别的训练图像输入参数生成网络中而得到的。
参数生成网络可以用于生成检测网络的参数。参数生成网络以训练图像作为输入，以检测网络的参数作为输出，将目标类别的训练图像输入参数生成网络中，可以得到目标类别的检测网络的参数。本公开实施例对参数生成网络的结构不做限制。
在一种可能的实现方式中，构建检测网络的过程为：首先，获取所述目标类别的训练图像，将所述目标类别的各训练图像分别输入所述参数生成网络中，得到所述目标类别的每个训练样本对应的检测网络的参数；然后，根据所述目标类别的每个训练样本对应的检测网络的参数，确定所述目标类别的检测网络的参数；最后，根据所述目标类别的检测网络的参数，构建所述目标类别的检测网络。
在本公开实施例中，可以将目标类别的各训练图像分别输入参数生成网络中，得到目标类别的每个训练图像对应的检测网络的参数。由于这些训练图像均属于目标类别，因此，可以根据这些训练图像对应的检测网络的参数，确定目标类别的检测网络的参数。在一个示例中，可以将目标类别的训练图像对应的检测网络的参数进行平均，将平均后的检测网络的参数确定为目标类别的检测网络的参数。在又一示例中，可以首先，根据训练图像中目标区域(目标类别的对象所在区域)的位置或者大小等信息，确定目标类别的各训练图像的权重信息；然后，基于权重信息，对目标类别的各训练图像对应的检测网络的参数进行加权平均，将加权平均后的检测网络的参数确定为目标类别的检测网络的参数。
在获取到检测网络的参数之后，可以基于检测网络的结构，构建出相应的检测网络。也就是说，在获取了目标类别的检测网络的参数之后，可以基于检测网络的结构，构建出目标类别的检测网络。
在一种可能的实现方式中，可以将参数设置为目标类别的检测网络的参数的检测网络，直接确定为目标类别的检测网络。这样，在将目标类别的训练图像输入参数生成网络后，即可方便、快捷的得到目标类别的检测网络。
在一种可能的实现方式中，可以先将参数设置为目标类别的检测网络的参数的检测网络，确定为目标类别的初始化检测网络；然后，对该初始化检测网络进行微调，得到为目标类别的检测网络。在一个示例中，可以通过损失最小化对初始化检测网络进行微调。这里的损失可以包括初始化检测网络的检测损失和正交损失。其中，初始化检测网络的检测损失可以根据目标类别的训练图像输入初始化检测网络后输出的预测标签分布结果和对应的真值标签确定。
这样，可以在较短的时间内得到优化的检测网络，从而提升了目标类别的检测网络的准确性。
在步骤S12中，可以将待检测图像输入目标类别的检测网络得到待检测图像的目标检测结果。一个示例中，目标检测结果可以包括待检测图像为目标类别的概率以及待检测图像中目标类别的对象的位置信息。
在本公开实施例中，首先基于参数生成网络，得到目标类别的检测网络的参数，然后根据目标类别的检测网络的参数，构建出目标类别的检测网络，从而实现了目标类别上的目标检测。参数生成网络是本公开实施例的目标检测方法的重要工具。下面对参数生成网络的训练过程进行说明。
在一种可能的实现方式中，参数生成网络的训练过程可以包括：从图像集中获取一个或多个目标训练集；并基于各目标训练集，训练所述参数生成网络。
以图像集包括C个类别(称为C个基类)的训练图像，每个类别包括A个训练图像为例进行说明。从图像集中获取一个目标训练集的过程可以包括：从C个类别中随机选取K个类别，并从每个类别的A个训练图像中，随机选取M个训练图像。此时，目标训练集包括K个类别的训练图像，每个类别包括M个训练图像。重复该过程，则可以从图像集中获取到多个目标训练集。
其中，C、A、K、M为大于0的整数，且C＞K，A＞M。
K和M的数量可以根据需要进行设置。由于本公开实施例的目标检测方法要解决的是训练图像数量较少的类别的检测问题，因此，本公开实施例在训练参数生成网络时，采用的类别数量较少，每个类别的训练图像的数量也较少。在一个示例中，K可以取5；M可以取11、15或者20等。而相应的C可以为1000或者2000等，A可以取5000或者10000等。可以理解的是，本公开实施例提供的训练参数生成网络的过程对数量较多的类别同样适用，因此，每个类别的训练图像的数量可以较多，M还可以取500或者1000等。
需要说明的是，针对目标训练集的每个类别，该类别包括的M个训练图像可以包括N个支持图像和O个查询图像，N和O为大于0的整数，且M≥N+O。在M＝N+O的情况下，针对目标训练集的每个类别，可以随机从该类别的M个训练图像中，选取N个训练图像作为支持图像，并将该类别剩下的训练图像作为查询图像。在M＞N+O的情况下，针对目标训练集的每个类别，可以随机从类别的M个训练图像中，选取N个训练图像作为支持图像，并从该类别剩下的训练图像中随机选取O个训练图像作为查询图像。
下面以一个目标训练集为例，对参数生成网络的训练过程进行说明。采用多个目标训练集训练参数生成网络的过程，实际上是多次重复采用一个目标训练集训练参数生成网络的过程，这里不再赘述。
在一种可能的实现方式中，基于一个目标训练集，训练参数生成网络，可以包括：首先，将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，并根据该通用检测网络的参数，构建该目标训练集的通用检测网络；其次，将该目标训练集的各查询图像输入待训练的特征提取网络，得到该目标训练集的各查询图像的特征图；再次，将所述各查询图像的特征图分别输入所述通用检测网络，得到所述各查询图像的预测标签分布结果；最后，根据所述各查询图像的预测标签分布结果和真值标签，确定所述通用检测网络的检测损失，并根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络。
其中，将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，可以包括：将该目标训练集的各支持图像分别输入待训练的参数生成网络，得到每个支持图像对应的检测网络的参数；根据各支持图像对应的检测网络的参数和各支持图像的真实类别，确定该目标训练集的每个类别的检测网络的参数；并根据该目标训练集的各类别的检测网络的参数，确定该目标训练集的通用检测网络的参数。
在一个示例中，可以按照各支持图像的真实类别，将同一类别的支持图像对应的检 测网络的参数进行平均或者加权平均(权值可以根据支持图像中目标区域的位置或者大小等信息确定)，得到对应类别的检测网络的参数。然后，将各类别的检测网络的参数拼接为目标训练集的通用检测网络的参数。
图2示出根据本公开实施例的网络架构示意图。如图2所示，该网络架构200包括参数生成网络f ψ201和特征提取网络202g φ。其中，参数生成网络f ψ的参数为ψ，特征提取网络g φ的参数为φ。
如图2所示，从图像集中获取了目标训练集D＝{(x s,y s) i,(x q,y q) j}，该目标训练集包括支持集D s＝{(x s,y s) i}和查询集D q＝{(x q,y q) j}。
其中，支持集D s包括K个类别的支持图像，每个类别包括N个支持图像。x s表示支持图像中的目标区域，y s表示x s的真值标签，(x s,y s) i表示支持集D s中第i个支持图像的目标区域和真值标签，1≤i≤K*N，
Figure PCTCN2021081674-appb-000001
其中，
Figure PCTCN2021081674-appb-000002
表示x s的类别，
Figure PCTCN2021081674-appb-000003
表示x s的位置信息。
查询集D q包括K个类别的查询图像，每个类别包括O个查询图像。x q表示查询图像中的目标区域，y q表示x q的真值标签，(x q,y q) j表示查询集D q中第j个查询图像的目标区域和真值标签，1≤j≤K*O。
Figure PCTCN2021081674-appb-000004
其中，
Figure PCTCN2021081674-appb-000005
表示x q的类别，
Figure PCTCN2021081674-appb-000006
表示x q的位置信息。
结合图2，采用目标训练集D，训练参数生成网络的过程可以包括：
(1)构建目标训练集的通用检测网络，具体地：
将支持集D s中的各支持图像进行裁剪得到各支持图像的目标区域x s(在一个示例中，目标区域的尺寸可以为224像素*224像素)，将各支持图像的目标区域x s输入待训练的参数生成网络f ψ中，可以得到每个支持图像对应的检测网络的参数，将同一类别的支持图像x s对应的检测网络的参数进行平均(或者加权平均)，可以得到该类别的检测网络的参数。公式(1)示出了类别k的检测网络的参数：
Figure PCTCN2021081674-appb-000007
其中，D表示检测网络的参数的维度，
Figure PCTCN2021081674-appb-000008
表示类别k的检测网络的参数；
Figure PCTCN2021081674-appb-000009
也就是说检测网络的类别与支持图像的类别一致。
将K个类别的检测网络的参数进行拼接，可以得到通用检测网络的参数
Figure PCTCN2021081674-appb-000010
然后根据该通用检测网络的参数
Figure PCTCN2021081674-appb-000011
可以构建出目标训练集的通用检测网络。
(2)获取查询图像的特征图，具体地：
将查询集D q中的各查询图像进行裁剪后得到各查询图像的目标区域x q(在一个示例中，目标区域的短边为600像素，长边不超过1000像素)输入特征提取网络g φ中，得到各查询图像的特征图g φ(x q)。
(3)确定通用检测网络的检测损失，具体地：
将各查询图像的特征图g φ(x q)输入参数为
Figure PCTCN2021081674-appb-000012
的通用检测网络中，可以得到各查询图像的预测标签分布结果。根据各查询图像的预测标签分布结果和真值标签y q，可以得到通用检测网络的检测损失。在一个示例中，可以通过公式(2)得到通用检测网络的检测损失。
Figure PCTCN2021081674-appb-000013
其中，L d表示通用检测网络的检测损失，loss(…)表示损失函数。
Figure PCTCN2021081674-appb-000014
表示以查询图像的真值标签y q和查询图像的特征图g φ(x q)输入参数为
Figure PCTCN2021081674-appb-000015
的通用检测网络中得到的预测标签分布结果为参数的损失函数。本公开实施例中对损失函数的结构不做限制，例如可以为均方误差函数、交叉熵函数等。
(4)根据通用检测网络的检测损失L d，训练待训练的参数生成网络f ψ。
以通用检测网络的检测损失最小化为目标，调整参数生成网络f ψ的参数ψ，以实现对参数生成网络f ψ的训练。
这样，通过少量的样本训练出来的参数生成网络f ψ，可以用来生成新类别的检测网络的参数，具有将其检测网络生成能力转移到新类上的潜力。
在一种可能的实现方式中，所述方法还包括：根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络。
由图2所示的网络架构可知，在训练参数生成网络f ψ的过程中，可以同时对特征提取网络g φ进行训练。也就是说，还可以以通用检测网络的检测损失最小化为目标，更新特征提取网络g φ的参数φ。
在一种可能的实现方式中，根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络，包括：获取该目标训练集的参考检测网络的参数；根据所述目标训练集的参考检测网络的参数，构建所述目标训练集的参考检测网络；将所述各查询图像的特征图分别输入所述参考检测网络，得到所述各查询图像的参考标签分布结果；根据所述 各查询图像的参考标签分布结果和真值标签，确定所述参考检测网络的检测损失；根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，训练所述待训练的特征提取网络。
其中，参考检测网络可以用于表示基于所述图像集的所有类别的训练图像，训练得到的检测网络。
采用目标训练集训练参考检测网络和特征提取网络时，一次训练过程仅涉及K个类别，多次训练仍然被限制在有限数量的类别中。这样，会导致训练出来的特征提取网络g φ的区分能力被限制在各目标训练集涉及到的类别中，使其提取特征的能力减弱。同时，采用目标训练集训练参考检测网络和特征提取网络时，训练过程涉及的训练图像的数量较少。而采用少量训练图像训练得到的通用检测网络相较于采用大量训练图像得到的检测网络的目标检测能力较弱。因此，在本公开实施例中，引入通过多种类别的大量训练图像训练得到的参考检测网络，对参数生成网络f ψ和特征提取网络g φ的训练进行优化。
图3示出根据本公开实施例的网络架构示意图。图3所示的网络架构在图2的基础上增加了参数为θ的参考检测网络301。将各查询图像的特征图g φ(x q)输入参数为θ的参考检测网络中，可以得到各查询图像的参考标签分布结果。根据各查询图像的参考标签分布结果和真值标签y q，可以得到参考检测网络的检测损失。在一个示例中，可以通过公式(3)示得到参考检测网络的检测损失。
Figure PCTCN2021081674-appb-000016
其中，L r表示参考检测网络的检测损失，loss(…)表示损失函数。
Figure PCTCN2021081674-appb-000017
表示以查询图像的真值标签
Figure PCTCN2021081674-appb-000018
和查询图像的特征图g φ(x q)输入参数为θ的参考检测网络中得到的参考标签分布结果为参数的损失函数。本公开实施例中对损失函数的结构不做限制，例如可以为均方误差函数、交叉熵函数等。
需要说明的是公式(2)中的y q和公式(3)中的
Figure PCTCN2021081674-appb-000019
均可以查询图像的真值标签，区别是y q是K个类别中的一个，
Figure PCTCN2021081674-appb-000020
是所有类别中的一个。
这样，由于参考检测网络是基于所有类别的训练图像训练出来的，因此根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，共同训练所述待训练的特征提取网络，可以提升特征提取网络的特征区分能力。
在一种可能的实现方式中，获取该目标训练集的参考检测网络的参数可以包括：获取随机初始化的检测网络；基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练；将训练完成的检测网络的参数，确定为该目标训练集的参考检测网络。
首先随机初始化一个检测网络作为待训练的检测网络，然后基于该目标训练集的所有查询图像，对待训练的检测网络，得到目标训练集的参考检测网络。该目标训练集的参考检测网络与参数生成网络f ψ和特征提取网络g φ可以同时进行训练。基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练的过程可以参照相关技术中训 练检测网络的训练方法，例如YOLO、SSD等，对此本公开不做限制。
目标训练集的参考检测网络的参数同样由K个类别的检测网络的参数拼接得到。基于目标训练集的参考检测网络的参数可以构建出目标训练集的参考检测网络。举例来说，假设目标训练集的K个类别中的类别k的参考检测网络的参数为
Figure PCTCN2021081674-appb-000021
其中，D为参考检测网络的参数的维度。将K个类别的参考检测网络的参数进行拼接，可以得到目标训练集的参考检测网络的参数
Figure PCTCN2021081674-appb-000022
需要说明的是，在本公开实施例中还可以重新构建一个包括K个类别的数据集进行参考检测网络的训练。训练过程可以参照上述采用查询图像进行训练的过程，这里不再赘述。
在一种可能的实现方式中，根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络，包括：根据该目标训练集的通用检测网络的参数和该目标训练集的参考检测网络的参数，确定所述通用检测网络的差距损失；根据所述通用检测网络的检测损失和差距损失，训练所述待训练的参数生成网络的参数。
在一个示例中，可以通过公式(4)或者公式(5)得到通用检测网络的差距损失。
Figure PCTCN2021081674-appb-000023
Figure PCTCN2021081674-appb-000024
其中，
Figure PCTCN2021081674-appb-000025
和
Figure PCTCN2021081674-appb-000026
为通用检测网络的差距损失L g的两个表现形式。∏是一个条件函数，括号中条件为真时，取值为1，括号中条件为假时，取值为0。||…|| 1表示一阶范数，||…|| 2表示二阶范数。f ψ(x s)和θ c分别表示类别c对应的通用检测网络和参考检测网络的参数。
这样，由于参考检测网络是基于所有类别的训练图像训练出来的，因此，根据所述通用检测网络的检测损失和差距损失，共同训练所述待训练的参数生成网络的参数，可以使基于参数生成网络得到的检测网络的准确度更高。
在一种可能的实现方式中，所述方法还可以包括：确定所述通用检测网络的正交化损失；根据所述通用检测网络的正交化损失，训练所述待训练的参数生成网络。
在一个示例中，可以通过公式(6)，确定通用检测网络的正交化损失。
Figure PCTCN2021081674-appb-000027
其中，
Figure PCTCN2021081674-appb-000028
是
Figure PCTCN2021081674-appb-000029
的行标准化版本，||…|| 1表示1阶范数，I是单位矩阵。
在本公开实施例中，通过使不同类别的检测网络之间彼此正交，可以提升模型的区分能力。
考虑到本公开实施例中，图3所示的参数生成网络f ψ、特征提取网络g φ和参考检测网络可以同时进行训练。因此，本公开实施例中可以通过公式(7)，确定一个总的训练损失。
L＝L d+L r+αL g+βL o      (7)；
其中，L表示总的训练损失，L d表示通用检测网络的检测损失(参见公式(2))，L r表示参考检测网络的检测损失(参见公式(3))，L g表示通用检测网络的差距损失(参见公式(4)和公式(5))，L o表示通用检测网络的正交损失(参见公式(6))。α和β为超参数。α和β可以根据需要进行设置。在一个示例中，α可以取0.01，β可以取1。
在本公开实施例中，可以基于L同时对参数生成网络f ψ、特征提取网络g φ和参考检测网络进行训练，调整参数ψ、φ和θ。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
此外，本公开还提供了目标检测装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种目标检测方法，相应技术方案和描述和参见方法部分的相应记载，不再赘述。
图4示出根据本公开实施例的目标检测装置的框图。如图4所示，所述装置40包括：
构建模块41，配置为构建目标类别的检测网络；
检测模块42，配置为采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；
其中，所述目标类别的检测网络的参数是基于目标类别的训练图像输入参数生成网络中而得到的。
在一种可能的实现方式中，所述装置还包括：
获取模块，配置为从图像集中获取一个或多个目标训练集，其中，每个目标训练集包括K个类别的训练图像，每个类别包括M个训练图像，K为大于0的整数；
第一训练模块，配置为基于各目标训练集，训练所述参数生成网络。
在一种可能的实现方式中，所述M个训练图像包括N个支持图像和O个查询图像，N和O为大于0的整数；所述第一训练模块还配置为：
针对每个目标训练集：
将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，并根据该通用检测网络的参数，构建该目标训练集的通用检测网络；
将该目标训练集的各查询图像输入待训练的特征提取网络，得到该目标训练集的各查询图像的特征图；
将所述各查询图像的特征图分别输入所述通用检测网络，得到所述各查询图像的预测标签分布结果；
根据所述各查询图像的预测标签分布结果和真值标签，确定所述通用检测网络的检测损失；
根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络。
在一种可能的实现方式中，将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，包括：
将该目标训练集的各支持图像分别输入待训练的参数生成网络，得到每个支持图像对应的检测网络的参数；
根据各支持图像对应的检测网络的参数和各支持图像的真实类别，确定该目标训练集的每个类别的检测网络的参数；
根据该目标训练集的各类别的检测网络的参数，确定该目标训练集的通用检测网络的参数。
在一种可能的实现方式中，所述装置还包括：
第二训练模块，配置为根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络。
在一种可能的实现方式中，所述第二训练模块还配置为：
获取目标训练集的参考检测网络；
将所述各查询图像的特征图分别输入所述参考检测网络，得到所述各查询图像的参考标签分布结果；
根据所述各查询图像的参考标签分布结果和真值标签，确定所述参考检测网络的检测损失；
根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，训练所述待训练的特征提取网络。
在一种可能的实现方式中，获取该目标训练集的参考检测网络的参数，包括：
获取随机初始化的检测网络；
基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练；
将训练完成的检测网络的参数，确定为该目标训练集的参考检测网络。
在一种可能的实现方式中，根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络，包括：
根据该目标训练集的通用检测网络的参数和该目标训练集的参考检测网络的参数，确定所述通用检测网络的差距损失；
根据所述通用检测网络的检测损失和差距损失，训练所述待训练的参数生成网络的参数。
在一种可能的实现方式中，所述装置还包括：
确定模块，配置为确定所述通用检测网络的正交化损失；
第三训练模块，配置为根据所述通用检测网络的正交化损失，训练所述待训练的参数生成网络。
在一种可能的实现方式中，所述构建模块还配置为：
获取所述目标类别的训练图像；
将所述目标类别的各训练图像分别输入所述参数生成网络中，得到所述目标类别的每个训练样本对应的检测网络的参数；
根据所述目标类别的每个训练样本对应的检测网络的参数，确定所述目标类别的检测网络的参数；
根据所述目标类别的检测网络的参数，构建所述目标类别的检测网络。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，当计算机可读 代码在设备上运行时，设备中的处理器执行用于实现如上任一实施例提供的目标检测方法的指令。
本公开实施例还提供了另一种计算机程序产品，用于存储计算机可读指令，指令被执行时使得计算机执行上述任一实施例提供的目标检测方法的操作。
电子设备可以被提供为终端、服务器或其它形态的设备。
图5示出根据本公开实施例的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图5，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设 备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如互补金属氧化物半导体(CMOS)或电荷耦合装置(CCD)图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如无线网络(WiFi)，第二代移动通信技术(2G)或第三代移动通信技术(3G)，或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
图6示出根据本公开实施例的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图6，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如微软服务器操作系统(Windows Server TM)，苹果公司推出的基于图形用户界面操作系统(Mac OS X TM)，多用户多进程的计算机操作系统(Unix TM),自由和开放原代码的类Unix操作系统(Linux TM)，开放原代码的类Unix操作系统(FreeBSD TM)或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是――但不限于――电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也 不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术的改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。
工业实用性
本公开提供了一种目标检测方法及装置、电子设备和存储介质，其中，构建目标类别的检测网络；采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；其中，所述目标类别的检测网络的参数是基于目标类别的训练图像输入参数生成网络中而得到的。

Claims (13)

  1. 一种目标检测方法，包括：
    构建目标类别的检测网络；
    采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；
    其中，所述目标类别的检测网络的参数是将目标类别的训练图像输入参数生成网络中而得到的。
  2. 根据权利要求1所述的方法，所述方法还包括：
    从图像集中获取一个或多个目标训练集，其中，每个目标训练集包括K个类别的训练图像，每个类别包括M个训练图像，K为大于0的整数；
    基于各目标训练集，训练所述参数生成网络。
  3. 根据权利要求2所述的方法，所述M个训练图像包括N个支持图像和O个查询图像，N和O为大于0的整数；所述基于各目标训练集，训练所述参数生成网络，包括：
    针对每个目标训练集：
    将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，并基于该通用检测网络的参数，构建该目标训练集的通用检测网络；
    将该目标训练集的各查询图像输入待训练的特征提取网络，得到该目标训练集的各查询图像的特征图；
    将所述各查询图像的特征图分别输入所述通用检测网络，得到所述各查询图像的预测标签分布结果；
    根据所述各查询图像的预测标签分布结果和真值标签，确定所述通用检测网络的检测损失；
    根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络。
  4. 根据权利要求3所述的方法，所述将该目标训练集的各支持图像输入待训练的参数生成网络，得到该目标训练集的通用检测网络的参数，包括：
    将该目标训练集的各支持图像分别输入待训练的参数生成网络，得到每个支持图像对应的检测网络的参数；
    根据各支持图像对应的检测网络的参数和各支持图像的真实类别，确定该目标训练集的每个类别的检测网络的参数；
    根据该目标训练集的各类别的检测网络的参数，确定该目标训练集的通用检测网络的参数。
  5. 根据权利要求3或4所述的方法，所述方法还包括：
    根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络。
  6. 根据权利要求5所述的方法，所述根据所述通用检测网络的检测损失，训练所述待训练的特征提取网络，包括：
    获取目标训练集的参考检测网络；
    将所述各查询图像的特征图分别输入所述参考检测网络，得到所述各查询图像的参考标签分布结果；
    根据所述各查询图像的参考标签分布结果和真值标签，确定所述参考检测网络的检测损失；
    根据所述通用检测网络的检测损失和所述参考检测网络的检测损失，训练所述待训练的特征提取网络。
  7. 根据权利要求6所述的方法，获取该目标训练集的参考检测网络的参数，包括：
    获取随机初始化的检测网络；
    基于该目标训练集的所有查询图像对所述随机初始化的检测网络进行训练；
    将训练完成的检测网络的参数，确定为该目标训练集的参考检测网络。
  8. 根据权利要求6或7所述的方法，所述根据所述通用检测网络的检测损失，训练所述待训练的参数生成网络，包括：
    根据该目标训练集的通用检测网络的参数和该目标训练集的参考检测网络的参数，确定所述通用检测网络的差距损失；
    根据所述通用检测网络的检测损失和差距损失，训练所述待训练的参数生成网络的参数。
  9. 根据权利要求3至8中任一项所述的方法，所述方法还包括：
    确定所述通用检测网络的正交化损失；
    根据所述通用检测网络的正交化损失，训练所述待训练的参数生成网络。
  10. 根据权利要求2所述的方法，所述构建目标类别的检测网络，包括：
    获取所述目标类别的训练图像；
    将所述目标类别的各训练图像分别输入所述参数生成网络中，得到所述目标类别的每个训练样本对应的检测网络的参数；
    根据所述目标类别的每个训练样本对应的检测网络的参数，确定所述目标类别的检测网络的参数；
    根据所述目标类别的检测网络的参数，构建所述目标类别的检测网络。
  11. 一种目标检测装置，包括：
    构建模块，配置为构建目标类别的检测网络；
    检测模块，配置为采用所述目标类别的检测网络对待检测图像进行检测，得到所述待检测图像的目标检测结果；
    其中，所述目标类别的检测网络的参数是将目标类别的训练图像输入参数生成网络中而得到的。
  12. 一种电子设备，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为调用所述存储器存储的指令，以执行权利要求1至10中任意一项所述的方法。
  13. 一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现权利要求1至10中任意一项所述的方法。
PCT/CN2021/081674 2020-07-30 2021-03-19 目标检测方法及装置、电子设备和存储介质 WO2022021901A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010751150.1A CN111898676B (zh) 2020-07-30 2020-07-30 目标检测方法及装置、电子设备和存储介质
CN202010751150.1 2020-07-30

Publications (1)

Publication Number Publication Date
WO2022021901A1 true WO2022021901A1 (zh) 2022-02-03

Family

ID=73183433

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/081674 WO2022021901A1 (zh) 2020-07-30 2021-03-19 目标检测方法及装置、电子设备和存储介质

Country Status (3)

Country Link
CN (1) CN111898676B (zh)
TW (1) TWI785638B (zh)
WO (1) WO2022021901A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115100419A (zh) * 2022-07-20 2022-09-23 中国科学院自动化研究所 目标检测方法、装置、电子设备及存储介质

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111898676B (zh) * 2020-07-30 2022-09-20 深圳市商汤科技有限公司 目标检测方法及装置、电子设备和存储介质
CN113128429A (zh) * 2021-04-24 2021-07-16 新疆爱华盈通信息技术有限公司 基于立体视觉的活体检测方法和相关设备
TWI803205B (zh) * 2022-02-23 2023-05-21 神通資訊科技股份有限公司 支援多感測態樣的計數系統及其方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109145958A (zh) * 2018-07-27 2019-01-04 哈尔滨工业大学 一种基于多任务生成对抗网络的真实场景小物体检测方法
CN110210535A (zh) * 2019-05-21 2019-09-06 北京市商汤科技开发有限公司 神经网络训练方法及装置以及图像处理方法及装置
US20190347828A1 (en) * 2018-05-09 2019-11-14 Beijing Kuangshi Technology Co., Ltd. Target detection method, system, and non-volatile storage medium
CN110580487A (zh) * 2018-06-08 2019-12-17 Oppo广东移动通信有限公司 神经网络的训练方法、构建方法、图像处理方法和装置
CN111898676A (zh) * 2020-07-30 2020-11-06 深圳市商汤科技有限公司 目标检测方法及装置、电子设备和存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108122027B (zh) * 2016-11-29 2021-01-12 华为技术有限公司 一种神经网络模型的训练方法、装置及芯片
CN108764370B (zh) * 2018-06-08 2021-03-12 Oppo广东移动通信有限公司 图像处理方法、装置、计算机可读存储介质和计算机设备
CN110009090B (zh) * 2019-04-02 2022-12-02 北京市商汤科技开发有限公司 神经网络训练与图像处理方法及装置
CN111242217A (zh) * 2020-01-13 2020-06-05 支付宝实验室(新加坡)有限公司 图像识别模型的训练方法、装置、电子设备及存储介质
CN111353580B (zh) * 2020-02-03 2023-06-20 中国人民解放军国防科技大学 目标检测网络的训练方法、电子设备及存储介质
CN111325343B (zh) * 2020-02-20 2022-09-09 北京市商汤科技开发有限公司 神经网络的确定、目标检测、智能行驶控制方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190347828A1 (en) * 2018-05-09 2019-11-14 Beijing Kuangshi Technology Co., Ltd. Target detection method, system, and non-volatile storage medium
CN110580487A (zh) * 2018-06-08 2019-12-17 Oppo广东移动通信有限公司 神经网络的训练方法、构建方法、图像处理方法和装置
CN109145958A (zh) * 2018-07-27 2019-01-04 哈尔滨工业大学 一种基于多任务生成对抗网络的真实场景小物体检测方法
CN110210535A (zh) * 2019-05-21 2019-09-06 北京市商汤科技开发有限公司 神经网络训练方法及装置以及图像处理方法及装置
CN111898676A (zh) * 2020-07-30 2020-11-06 深圳市商汤科技有限公司 目标检测方法及装置、电子设备和存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115100419A (zh) * 2022-07-20 2022-09-23 中国科学院自动化研究所 目标检测方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
TWI785638B (zh) 2022-12-01
CN111898676A (zh) 2020-11-06
TW202205127A (zh) 2022-02-01
CN111898676B (zh) 2022-09-20

Similar Documents

Publication Publication Date Title
WO2022021901A1 (zh) 目标检测方法及装置、电子设备和存储介质
TWI759722B (zh) 神經網路訓練方法及裝置、圖像處理方法及裝置、電子設備和計算機可讀存儲介質
TWI766286B (zh) 圖像處理方法及圖像處理裝置、電子設備和電腦可讀儲存媒介
KR102421820B1 (ko) 키포인트 검출 방법, 장치, 전자 기기 및 기억 매체
WO2022011892A1 (zh) 网络训练方法及装置、目标检测方法及装置和电子设备
WO2019141042A1 (zh) 图像分类方法、装置及终端
TWI771645B (zh) 文本識別方法及裝置、電子設備、儲存介質
CN110009090B (zh) 神经网络训练与图像处理方法及装置
WO2022134382A1 (zh) 图像分割方法及装置、电子设备和存储介质、计算机程序
KR20200128165A (ko) 비디오 처리 방법 및 장치, 전자 기기 및 저장 매체
CN110633755A (zh) 网络训练方法、图像处理方法及装置、电子设备
WO2022227405A1 (zh) 道路病害检测方法及装置、电子设备和存储介质
WO2021036382A9 (zh) 图像处理方法及装置、电子设备和存储介质
CN111931844B (zh) 图像处理方法及装置、电子设备和存储介质
CN109615006B (zh) 文字识别方法及装置、电子设备和存储介质
CN113326768B (zh) 训练方法、图像特征提取方法、图像识别方法及装置
CN110781813B (zh) 图像识别方法及装置、电子设备和存储介质
CN110781934A (zh) 监督学习、标签预测方法及装置、电子设备和存储介质
WO2021098256A1 (zh) 一种预测方法及装置、电子设备和存储介质
CN113792207A (zh) 一种基于多层次特征表示对齐的跨模态检索方法
CN111582383B (zh) 属性识别方法及装置、电子设备和存储介质
CN111259967A (zh) 图像分类及神经网络训练方法、装置、设备及存储介质
CN111523599B (zh) 目标检测方法及装置、电子设备和存储介质
WO2022141969A1 (zh) 图像分割方法及装置、电子设备、存储介质和程序
TWI751593B (zh) 網路訓練方法及裝置、圖像處理方法及裝置、電子設備、電腦可讀儲存媒體及電腦程式

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: 21850352

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

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

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 09.05.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 21850352

Country of ref document: EP

Kind code of ref document: A1