WO2022213846A1 - 识别模型的训练 - Google Patents

识别模型的训练 Download PDF

Info

Publication number
WO2022213846A1
WO2022213846A1 PCT/CN2022/083705 CN2022083705W WO2022213846A1 WO 2022213846 A1 WO2022213846 A1 WO 2022213846A1 CN 2022083705 W CN2022083705 W CN 2022083705W WO 2022213846 A1 WO2022213846 A1 WO 2022213846A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
fuzzy
recognition
training
recognition result
Prior art date
Application number
PCT/CN2022/083705
Other languages
English (en)
French (fr)
Inventor
夏华夏
乔健
Original Assignee
北京三快在线科技有限公司
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 北京三快在线科技有限公司 filed Critical 北京三快在线科技有限公司
Publication of WO2022213846A1 publication Critical patent/WO2022213846A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle

Definitions

  • This specification relates to the field of unmanned driving technology, and in particular, to a training method and device for a recognition model.
  • the driverless device During the driving process of the driverless device, it usually recognizes the target objects in the surrounding environment, and determines the movement strategy of the driverless device based on the detection results of the target objects so as not to violate the traffic laws.
  • the target can usually be an object such as an obstacle, a sign, a signal light, etc., which will affect the driving of the unmanned device.
  • a method for identifying a signal light is to determine the location of the signal light and the type of the signal light based on the detection model and the identification model. Specifically, the current image is first input into the detection model to obtain the position of each signal light in the current image, the current image is intercepted according to the position of each signal light in the current image, the image of each signal light is obtained, and the signal light image is input into the recognition model to determine Type of each signal light. Subsequently, the motion strategy of the unmanned device can be determined based on the position and type of each signal light.
  • the present specification provides a training method and device for a recognition model to partially solve the above problems existing in the prior art.
  • this specification provides a training method for a recognition model, including: acquiring an unlabeled target image; blurring the unlabeled target image to obtain the unlabeled target image. a first blurred image and a second blurred image, the blurring intensity of the first blurred image is higher than that of the second blurred image; the second blurred image is input into the recognition model to be trained, and the second blurred image is determined
  • the recognition result of the image according to the recognition result of the second fuzzy image, determine the label of the first fuzzy image; take the first fuzzy image as the first training sample, according to the label of the first fuzzy image, to treat The trained recognition model is trained.
  • performing a blurring process on the unmarked target image to obtain a second blurred image of the unmarked target image includes: performing multiple blurring on the unmarked target image processing to obtain multiple second fuzzy images of the unlabeled target image; inputting the second fuzzy images into the recognition model to be trained, and determining the recognition results of the second fuzzy images, including : input the plurality of second fuzzy images into the recognition model to be trained, and determine the respective recognition results of the plurality of second fuzzy images; according to the recognition results of the second fuzzy images, determine the first Annotation of a fuzzy image includes: according to the recognition results of the plurality of second fuzzy images and their weights, weighted average of the recognition results of the plurality of second fuzzy images to determine a weighted recognition result; The weighted recognition result is used to determine the label of the first blurred image.
  • the method further includes: using an iterative training method to train the recognition model to be trained, and the method for training the recognition model to be trained in one iteration includes: acquiring a marked target image as a first Two training samples; input the first training sample and the second training sample into the recognition model to be trained, and determine the recognition result of the first training sample and the recognition result of the second training sample; The first loss is determined according to the recognition result of the first training sample and its label, and the second loss is determined according to the recognition result of the second training sample and its label. loss, adjust the model parameter values of the recognition model to be trained.
  • determining the label of the first fuzzy image according to the recognition result of the second fuzzy image includes: acquiring a labeled target image as a second training sample; The second fuzzy image of the object image is input into the pre-recognition module determined based on the second training sample, and the pre-recognition result of the second fuzzy image is determined; according to the recognition result of the second fuzzy image and the According to the pre-recognition result, the label of the first blurred image is determined.
  • the method further includes: inputting the second training sample into a pre-recognition module to be trained, determining a pre-recognition result of the second training sample, and using the pre-recognition result of the second training sample Minimizing the difference between the recognition result and the labeling of the second training sample is the training objective, and adjusting the model parameter value of the pre-recognition module.
  • determining the label of the first fuzzy image according to the recognition result of the second fuzzy image includes: determining the second fuzzy image according to the recognition result of the second fuzzy image and its confidence level The classification result of the image; according to the classification result of the second fuzzy image, the label of the first fuzzy image is determined.
  • using an iterative training method to train the recognition model to be trained includes: for one iteration, determining the first fuzzy set according to the recognition results and labels of the first training samples of the iteration The sample weight of the image; when the recognition model is trained in the next iteration, in response to the sample weight of the first fuzzy image being greater than or equal to a preset threshold, the first fuzzy image is determined as the first training of the next iteration sample.
  • this specification provides a training device for a recognition model, including: an acquisition module for acquiring an unlabeled target image; a processing module for performing a fuzzification process on the unlabeled target image to obtain the the first blurred image and the second blurred image of the unmarked target image, the blurring intensity of the first blurred image is higher than that of the second blurred image; the identification module is used for inputting the second blurred image to the In the recognition model to be trained, the recognition result of the second fuzzy image is determined; the labeling module is used to determine the label of the first fuzzy image according to the recognition result of the second fuzzy image; the training module is used to The first fuzzy image is a first training sample, and the recognition model to be trained is trained according to the labeling of the first fuzzy image.
  • this specification provides a computer-readable storage medium, where the storage medium stores a computer program, and when the computer program is executed by a processor, implements the above-mentioned training method of the recognition model.
  • this specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, where the processor implements the above-mentioned training method for the recognition model when the program is executed .
  • the above-mentioned at least one technical solution adopted in this specification can achieve the following beneficial effects: in the training method of the recognition model provided in this specification, by performing blurring processing on the acquired image of the target object, the first blurred image and the second blurred image are determined. , through the recognition model to be trained, determine the recognition result of the second fuzzy image, and determine the label of the first fuzzy image according to the recognition result of the second fuzzy image, take the first fuzzy image as the first training sample, according to the first fuzzy image The annotation of the image, the recognition model to be trained is trained.
  • this method does not require manual labeling of samples, which reduces the time and cost of generating training samples and improves the training efficiency of the recognition model.
  • Fig. 3 is the schematic diagram of the recognition model provided by this specification.
  • FIG. 4 is a schematic diagram of a pre-identification module provided in this specification.
  • FIG. 5 is a structural diagram of determining the loss function provided in this specification.
  • Fig. 6 is the training device of the recognition model provided by this specification.
  • FIG. 7 is a schematic diagram of an electronic device corresponding to FIG. 1 provided in this specification.
  • FIG. 1 is a schematic flowchart of the training method of the recognition model provided in this specification, which specifically includes the following steps:
  • the environment image around the unmanned device can be detected by the detection model and the recognition model, and the position of the target object and the type of the target object can be determined.
  • Type which determines the motion strategy of the unmanned device.
  • the recognition model is pre-trained based on the training samples by the server used to train the model. While this specification provides a method for training a recognition model, similarly, the process of training the recognition model can be performed by a server for training the model.
  • the training model can be divided into a sample generation stage and a training model stage.
  • a training sample for training the model can be determined according to model needs and training needs.
  • the training samples for training the recognition model can be determined first, and since the recognition model usually recognizes the target based on the target image, each target image can be determined first to determine the training sample.
  • the target object image may be an image of a signal light, or may be a traffic sign during driving of the unmanned device, etc.
  • a signal light will be used as an example for description in the following.
  • the recognition model can recognize the target object based on the target object image obtained by inputting the surrounding image collected by the unmanned device into the detection model. Therefore, the target object image of the recognition model can be determined based on the detection model. Specifically, images collected by several unmanned devices may be obtained first (the images may include objects around the unmanned devices that collected the images), and the acquired images collected by each unmanned device may be input into the The detection model determines the position of each target (such as a signal light) in the images collected by each unmanned device. Therefore, each image can be cropped based on the determined position of each target, and each image containing the target can be determined. target image.
  • images collected by several unmanned devices may be obtained first (the images may include objects around the unmanned devices that collected the images), and the acquired images collected by each unmanned device may be input into the The detection model determines the position of each target (such as a signal light) in the images collected by each unmanned device. Therefore, each image can be cropped based on the determined position of each target, and each image containing the
  • an unmanned device may refer to a device that can realize automatic driving, such as an unmanned vehicle, a robot, and an automatic distribution device, or a semi-autonomous driving device or an auxiliary driving device, etc., which is not limited in this document.
  • the unmanned device using the recognition model trained by the training method provided in this specification can be used to perform distribution tasks, such as using the unmanned device to perform delivery tasks such as express delivery, logistics, and takeaway.
  • delivery tasks such as express delivery, logistics, and takeaway.
  • unmanned equipment needs to identify static objects with specific traffic indication functions such as traffic lights and traffic signs set on the road through the collected images.
  • S102 For each target image, perform a blurring process on the target image to determine a first blurred image and a second blurred image of the target image, where the blurring intensity of the first blurred image is higher than that of the second blurred image blurred images;
  • the target image may be blurred for each target image to determine the first blurred image of the target image and the second blurred image. So that the recognition model can be trained subsequently based on the first blurred image and the second blurred image of each target image.
  • the lower the blurring intensity of the blurred image of the target image the smaller the gap between the blurred image and the target, and the recognition result of the blurred image is the same as the target image.
  • the difference between the recognition results is also smaller.
  • the higher the blurring intensity of the blurred image of the target image the greater the gap between the blurred image and the target, and the greater the gap between the recognition result of the blurred image and the target image.
  • At least two blurring processing models with processing intensities may be preset, and after each target object image is determined, for each target object image, the blurring processing models with at least two processing intensities may be preset according to The model is used for blurring the image of the target object to obtain a first blurred image and a second blurred image with different blur intensities.
  • the left part is the image of the target object, and the image of the target object can be an image of a signal light.
  • the first blurring process and the second blurring process are respectively performed on the target object image, and the first blurring process of the target object image on the right part can be obtained.
  • a blurred image and a second blurred image Obviously, the blurring intensity of the first blurred image is higher than that of the second blurred image.
  • S104 Input the second fuzzy image into the recognition model to be trained, and determine the recognition result of the second fuzzy image.
  • the second blurred image may also be input into the recognition model to be trained to determine the second blurred image of the second blurred image. Identify the results.
  • the recognition result of the second blurred image of the target image is closer to the recognition result of the target image.
  • the recognition result of the second fuzzy image is determined, the label of the first fuzzy image is determined, and the recognition model is trained in the subsequent steps, so that the recognition model can determine the same recognition result for different fuzzy images of the same target image. Therefore, for each unlabeled target object image, the second fuzzy image of the target object image can be input into the recognition model to be trained, and the recognition result of the second fuzzy image can be determined.
  • the left part is the second blurred image.
  • the target image corresponding to the second fuzzy image The corresponding signal light status is green, the middle part is the recognition model to be trained, and the right part is the recognition result of the second fuzzy image.
  • the recognition result is the probability that the target object belongs to each preset classification, that is, the probability that the signal light state corresponding to the second fuzzy image is off, the signal light state is red, the signal light state is yellow, and the signal light state is green, respectively. Therefore, it can be determined that in the recognition result of the second blurred image, the probability that the status of the signal light is green is relatively high.
  • S106 Determine the label of the first blurred image according to the recognition result of the second blurred image.
  • the label of the first fuzzy image may also be determined according to the recognition result of the second fuzzy image.
  • the label of the first fuzzy image can be determined based on the recognition result of the second fuzzy image of the target image, and in the subsequent steps, The recognition model is trained based on the recognition result of each first fuzzy image and the annotation of the first fuzzy image.
  • the identification result of the second blurred image may be used as an annotation of the first blurred image.
  • the classification result of the second fuzzy image can be determined according to the recognition result of the second fuzzy image and its confidence; the label of the first fuzzy image can be determined according to the classification result of the second fuzzy image. For example, from the recognition result of the second blurred image, the state corresponding to the highest probability is determined as the label of the first blurred image. For example, the probability that the state corresponding to the signal light is off, red, yellow, and green is 10%, 80%, 2%, and 8%, respectively, then it can be determined that the first blurred image is marked with red.
  • the first blurred image may be the first blurred image.
  • the recognition model to be trained is trained according to the labeling of the first fuzzy image.
  • the first blurred image can be used as the first training sample and input into the recognition model to be trained to determine the recognition result of the first training sample.
  • the recognition result of the first training sample and its label the first loss is determined, and the model parameter value of the recognition model is adjusted according to the first loss, so as to complete the training of the recognition model.
  • the obtained target image is subjected to fuzzy processing to obtain a first fuzzy image and a second fuzzy image, and the recognition model to be trained is used to determine the recognition result of the second fuzzy image, And according to the recognition result of the second fuzzy image, determine the label of the first fuzzy image, take the first fuzzy image as the first training sample, and train the recognition model to be trained according to the label of the first fuzzy image.
  • This method does not require manual labeling of samples, reduces the generation time and cost of training samples, and improves the training efficiency of the recognition model.
  • step S102 when the second blurred image of the target image is determined, multiple blurring processes may also be performed on the target image to obtain multiple second blurred images of the target image. Therefore, in step S104, when the second blurred image is determined, the plurality of second blurred images may be input into the recognition model to be trained, and the respective recognition results of the plurality of second blurred images may be determined. Then, in step S106, the recognition results of the plurality of second blurred images and the preset weights of the second blurred images may be input into the weighting function Wherein, X i is the recognition result of each second fuzzy image, a i is the weight corresponding to the recognition result of each second fuzzy image, and k is the number of each second fuzzy image.
  • the weighted identification results of the second blurred images can be obtained as the labels of the first blurred images.
  • the weighted recognition result can be directly used as the annotation of the first blurred image.
  • the state corresponding to the highest probability in the weighted recognition result may be used as the label of the first fuzzy image.
  • the weighted identification result is a result obtained by weighted average calculation of the identification results of each second fuzzy image, that is, in this embodiment, the sum of the weights corresponding to the identification results of each second fuzzy image is 1.
  • the labeling of the blurred image may also be assisted to obtain based on the labeled training samples. Therefore, a pre-recognition module can be obtained by training based on the labeled training samples, and for each unlabeled target image, the second fuzzy image of the target image is input into the pre-recognition module, and the second fuzzy image is determined. The pre-recognition result is used to determine the label of the first blurred image.
  • a number of marked target images can be obtained first as second training samples, and then each second training sample is input into the pre-recognition module to be trained to determine the pre-recognition result of each second training sample, according to Minimizing the difference between the pre-recognition result of each second training sample and its label is the training target, and adjusting the model parameter values of the pre-recognition module to obtain the pre-recognition module.
  • each second fuzzy image is input into the pre-recognition module to obtain the recognition result of each second fuzzy image, as shown in FIG. 4 .
  • the left part is the second fuzzy image
  • the middle part is the pre-recognition module pre-trained based on each marked target image
  • the right part is the recognition result of the second fuzzy image , similar to FIG. 3
  • the state of the signal light corresponding to the second fuzzy image is green
  • the pre-recognition result is the probability that the target belongs to each preset classification, that is, the state of the signal light corresponding to the second fuzzy image is off and the state of the signal light is red.
  • the probability that the signal light state is yellow and the signal light state is green, respectively.
  • step S106 after the pre-recognition result of the second fuzzy image is determined, the recognition result of the second fuzzy image and its weight, and the pre-recognition result of the second fuzzy image and its weight (these two The sum of the weights is 1), and the weighted recognition result of the second blurred image is determined as the label of the first blurred image.
  • step S108 based on the annotation determined in step S106, the model can be trained with the recognition result of the first fuzzy image and its annotation.
  • this specification in order to obtain a more accurate recognition model, this specification can use labeled samples and unlabeled samples to train the recognition model, so that a number of labeled samples can also be obtained.
  • the target image is used as the second training sample.
  • the first training sample and the second training sample are input into the recognition model to be trained, and the recognition result of the first training sample and the recognition result of the second training sample are determined. Determine the first loss according to the recognition result and label of the first training sample of the pair of training samples, determine the second loss according to the recognition result and label of the second training sample, and adjust the adjustment according to the first loss and the second loss
  • the model parameters of the recognition model are used to complete the training of the recognition model, as shown in Figure 5.
  • Fig. 5 is a structural diagram of determining the loss function provided in this specification. Similar to Fig. 3 and Fig. 4, the unlabeled target image, the first blurred image of the unlabeled target image, and the signal light state corresponding to the second blurred image are all is green, and the signal light state corresponding to the second training sample is yellow.
  • Each recognition result is the probability that the target object belongs to each preset classification, that is, the corresponding probability that the signal light state corresponding to the second blurred image is off, the signal light state is red, the signal light state is yellow, and the signal light state is green, respectively.
  • each training sample pair can be determined according to each first training sample and each second training sample, and for each training sample pair, the first training sample and the second training sample in the training sample pair are used as input , input into the recognition model to be trained, determine the recognition result of the first training sample and the recognition result of the second training sample, and determine the first loss according to the recognition result of the first training sample and its label , and determine the second loss according to the recognition results of the second training samples and their labels.
  • the recognition model is trained with the minimum sum of the first loss and the second loss as the optimization goal.
  • the weighted recognition result can also be determined according to the recognition result and weight of each second fuzzy image, and the pre-recognition result and weight of at least one second fuzzy image, as
  • the first label of the first fuzzy image for example, the first label may be a probability value from 0 to 1
  • the weighted identification result determine the state corresponding to the highest probability as the second label of the first fuzzy image (For example, taking a signal light as an example, if the state corresponding to the highest probability determined according to the weighted recognition result is red, the red color is marked as 1, and the other colors are marked as 0).
  • the first loss may be determined according to the recognition result of the first training sample and the first label
  • the second loss may be determined according to the recognition result of the second training sample and its label
  • the third loss is determined according to the recognition result of the first training sample and the second label.
  • the recognition model is trained with the minimum sum of the first loss, the second loss, and the third loss as the optimization goal.
  • an iterative training method may be used to train the recognition model, and the method for one iteration of training the recognition model includes determining a first training sample, and training the recognition model based on each first training sample.
  • the sample weight of the first fuzzy image can be determined according to the identification result and the label of the first training sample of the iteration.
  • the recognition model in the next iteration determine whether the first fuzzy image can be used as the first training sample for the next iteration according to the sample weight of the first fuzzy image, and the sample weight and the first fuzzy image are used as the first training sample The probabilities of the samples are positively correlated.
  • the first blurred image is determined as the first training sample for the next iteration.
  • the current weight of the first training sample (which can be preset as 100%), and the recognition result of the second fuzzy image corresponding to the first fuzzy image, it can be determined whether the label of the first fuzzy image satisfies the first weight reduction
  • the first weight reduction condition may be that in the annotation of the first fuzzy image, the probability of each type is less than 30%. If the first weight reduction condition is satisfied, the current weight of the first blurred image is reduced; or it is determined whether the recognition result of the first fuzzy image satisfies the second weight reduction condition, and the second weight reduction condition may be that the first fuzzy image In the recognition result of the image, there are types with a probability of not less than 70%.
  • the training samples of types with a probability of not less than 70% in the recognition results can be considered as simple samples. If the second weight reduction condition is met, the first blur is reduced. The current weight of the image. The reduced current weight of the first blurred image is determined as the sample weight of the first blurred image, and if the sample weight of the first blurred image is greater than or equal to the probability threshold, the first blurred image is determined as the first training sample.
  • the detection model and the recognition model can be combined, and the recognition model can be applied to determine each target in the unmanned equipment environmental image type of scene, so that the unmanned equipment can determine the motion strategy of the unmanned equipment at the next moment based on the determined target type, and adopt the corresponding strategy to make the unmanned equipment normal at the next moment. drive without accident.
  • the specific method of determining the motion strategy of the unmanned device according to the type and position of the target is relatively mature, and this specification will not repeat it.
  • the number of the first training samples may be less than the number of the second training samples. Therefore, when determining the training sample pair, the selection of the first training samples may be repeated. Of course, in order to obtain more The training samples can also be selected repeatedly for the second training sample pair.
  • the specific method for determining the training samples can be set as required, which is not limited in this specification.
  • step S108 of this specification based on each first blurred image and the identification model obtained by its annotation training, the same identification result can be output for blurred images with different blurring intensities of the same target image. Therefore, the recognition model trained based on a large number of training samples can output more accurate recognition results for each input target image, but the accuracy of the recognition results is relative to the model, that is, the target is obtained through the recognition model.
  • the recognition result of the target object may be misaligned with the type of the target itself. For example, for several target images whose signal light status is red, the signal light status of the recognition result obtained by the recognition model is all green. Therefore, in actual use of the trained recognition model, the recognition result can be mapped to output the correct recognition result.
  • the blurring processing of the target image is not limited to blurring the target image, as long as the target image is subjected to the blurring processing to obtain the first blurred image and the target image whose similarity is the first similarity.
  • the second blurred image whose similarity with the target image is the second similarity, and the first similarity may be smaller than the second similarity.
  • the blurring processing performed on the image may also include corrosion processing, that is, determining the first corrosion image and the second corrosion image of the target image, the corrosion intensity of the first corrosion image is higher than that of the second corrosion image, based on the first corrosion image and the second corrosion image of the target image.
  • the second corrosion image determines the label of the first corrosion image, and the recognition model is trained based on the recognition result and label of the first corrosion image.
  • the blurring processing performed on the target image may also include affine transformation processing.
  • the above blurring processing, erosion processing, affine transformation processing, and the like are collectively referred to as blurring processing.
  • the blurred image of the target image may also be the result of the superposition of each blurring process, and how to determine the blurred image can be set as required, which is not limited in this specification.
  • weighting conditions and weighting reduction conditions can be set as required, which is not limited in this specification.
  • the training method of the recognition model provided above in one or more embodiments of this specification is based on the same idea, and the specification also provides a corresponding training device for the recognition model, as shown in FIG. 6 .
  • Fig. 6 is the training device of the recognition model provided by this specification, including:
  • the acquiring module 200 is used for acquiring an unmarked target image.
  • a processing module 202 configured to perform a blurring process on the unmarked target image to obtain a first blurred image and a second blurred image of the unmarked target image, wherein the first blurred image has a high blur intensity on the second blurred image.
  • the recognition module 204 is configured to input the second fuzzy image into the recognition model to be trained, and determine the recognition result of the second fuzzy image.
  • An annotation module 206 configured to determine an annotation of the first fuzzy image according to the recognition result of the second fuzzy image.
  • the training module 208 is configured to use the first fuzzy image as a first training sample, and train the recognition model to be trained according to the label of the first fuzzy image.
  • the processing module 202 is specifically configured to perform multiple blurring processes on the unmarked target image to obtain multiple second blurred images of the unmarked target image;
  • the recognition module 204 is specifically configured to input the plurality of second fuzzy images into the recognition model to be trained, and determine the respective recognition results of the plurality of second fuzzy images;
  • the labeling module 206 is specifically configured to perform a weighted average of the respective identification results of the multiple second blurry images according to the respective identification results of the multiple second blurry images and their weights, and determine the weighted identification. As a result, the labeling of the first blurred image is determined according to the weighted identification result.
  • the training module 208 is specifically configured to use an iterative training method to train the recognition model to be trained, and the method for training the recognition model to be trained in one iteration includes: acquiring an image of an object marked as Second training sample, input the first training sample and the second training sample into the recognition model to be trained, and determine the recognition result of the first training sample and the recognition result of the second training sample , according to the recognition result of the first training sample and its label, determine the first loss, and according to the recognition result of the second training sample and its label, determine the second loss, according to the first loss and the first loss
  • the second loss is to adjust the model parameter values of the recognition model to be trained.
  • the labeling module 206 is specifically configured to obtain a labeled target image as a second training sample, and input the second blurred image of the unlabeled target image to a In the pre-recognition module determined by the training sample, the pre-recognition result of the second fuzzy image is determined, and the label of the first fuzzy image is determined according to the recognition result of the second fuzzy image and the pre-recognition result.
  • the training module 208 is further configured to input the second training sample into the pre-recognition module to be trained, and determine the pre-recognition result of the second training sample, using the second training sample as the pre-recognition result. Minimizing the difference between the pre-recognition result and the labeling of the second training sample is the training target, and the model parameter value of the pre-recognition module is adjusted.
  • the labeling module 206 is specifically configured to determine the classification result of the second fuzzy image according to the recognition result of the second fuzzy image and its confidence, and according to the classification result of the second fuzzy image, An annotation of the first blurred image is determined.
  • the training module 208 is further configured to, for one iteration, determine the sample weight of the first fuzzy image according to the identification result of the first training sample and its label of the iteration, and train the next iteration.
  • the recognition model in response to the sample weight of the first blurred image being greater than or equal to a preset threshold, the first blurred image is determined as the first training sample of the next iteration.
  • the present specification also provides a computer-readable storage medium, where a computer program is stored in the storage medium, and the computer program can be used to execute the training method of the recognition model provided in FIG. 1 above.
  • the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and of course, may also include hardware required by other services.
  • the processor reads the corresponding computer program from the non-volatile memory into the memory and then executes it, so as to realize the training method of the recognition model described in FIG. 1 above.
  • this specification does not exclude other implementations, such as logic devices or the combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic device.
  • a Programmable Logic Device (such as a Field Programmable Gate Array (FPGA)) is an integrated circuit whose logic function is determined by user programming of the device.
  • HDL Hardware Description Language
  • ABEL Advanced Boolean Expression Language
  • AHDL Altera Hardware Description Language
  • HDCal JHDL
  • Lava Lava
  • Lola MyHDL
  • PALASM RHDL
  • VHDL Very-High-Speed Integrated Circuit Hardware Description Language
  • Verilog Verilog
  • the controller may be implemented in any suitable manner, for example, the controller may take the form of eg a microprocessor or processor and a computer readable medium storing computer readable program code (eg software or firmware) executable by the (micro)processor , logic gates, switches, application specific integrated circuits (ASICs), programmable logic controllers and embedded microcontrollers, examples of controllers include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20 and Silicon Labs C8051F320, the memory controller can also be implemented as part of the control logic of the memory.
  • the controller may take the form of eg a microprocessor or processor and a computer readable medium storing computer readable program code (eg software or firmware) executable by the (micro)processor , logic gates, switches, application specific integrated circuits (ASICs), programmable logic controllers and embedded microcontrollers
  • ASICs application specific integrated circuits
  • controllers include but are not limited to
  • the controller in addition to implementing the controller in the form of pure computer-readable program code, the controller can be implemented as logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded devices by logically programming the method steps.
  • the same function can be realized in the form of a microcontroller, etc. Therefore, such a controller can be regarded as a hardware component, and the devices included therein for realizing various functions can also be regarded as a structure within the hardware component. Or even, the means for implementing various functions can be regarded as both a software module implementing a method and a structure within a hardware component.
  • a typical implementation device is a computer.
  • the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or A combination of any of these devices.
  • embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media having computer-usable program code embodied therein, including but not limited to disk storage, CD-ROM, optical storage, and the like.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions
  • the apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.
  • a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
  • processors CPUs
  • input/output interfaces network interfaces
  • memory volatile and non-volatile memory
  • Memory may include forms of non-persistent memory, random access memory (RAM) and/or non-volatile memory in computer readable media, such as read only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.
  • RAM random access memory
  • ROM read only memory
  • flash RAM flash memory
  • Computer-readable media includes both persistent and non-permanent, removable and non-removable media, and storage of information may be implemented by any method or technology.
  • Information may be computer readable instructions, data structures, modules of programs, or other data.
  • Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Flash Memory or other memory technology, Compact Disc Read Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.
  • computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.
  • the embodiments of the present specification may be provided as a method, a system or a computer program product. Accordingly, this description may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
  • computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.
  • program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types.
  • the specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network.
  • program modules may be located in both local and remote computer storage media including storage devices.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

本说明书公开了一种识别模型的训练方法及装置，通过对获取到的无标注的目标物图像进行模糊化处理，确定第一模糊图像和第二模糊图像，通过待训练的识别模型，确定第二模糊图像的识别结果，并根据第二模糊图像的识别结果，确定第一模糊图像的标注，以第一模糊图像为第一训练样本，根据第一模糊图像的标注，对待训练的识别模型进行训练。不需要人工对样本进行标注，减少了训练样本的生成时间与成本，提高了识别模型的训练效率。

Description

识别模型的训练 技术领域
本说明书涉及无人驾驶技术领域，尤其涉及一种识别模型的训练方法及装置。
背景技术
无人驾驶设备行驶过程中，通常会对周围环境中的目标物进行识别，以基于目标物的检测结果，确定无人驾驶设备的运动策略以不违反交通法规。其中，目标物通常可为障碍物、指示牌、信号灯等会影响无人驾驶设备行驶的对象。
以信号灯为例进行说明，一种对信号灯进行识别的方法是基于检测模型和识别模型确定信号灯所在位置和信号灯类型。具体的，首先将当前图像输入检测模型，得到当前图像中各信号灯的位置，根据当前图像中各信号灯的位置对当前图像进行截取，得到各信号灯图像，将信号灯图像输入到识别模型中，确定出各信号灯的类型。后续则可基于各信号灯的位置和类型确定无人驾驶设备的运动策略。
但是，在实现对目标物的检测和识别的过程中，需要人工对训练样本进行标注。而人工标注成本高、时间长，这使得训练模型的周期过长，降低了模型训练的效率。
发明内容
本说明书提供一种识别模型的训练方法及装置，以部分的解决现有技术存在的上述问题。
本说明书采用下述技术方案。
第一方面，本说明书提供了一种识别模型的训练方法，包括：获取无标注的目标物图像；对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像；将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果；根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注；以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
在一些实施例中，对所述无标注的目标物图像进行模糊化处理，获得所述无标注的 目标物图像的第二模糊图像，包括：对所述无标注的目标物图像进行多种模糊化处理，获得所述无标注的目标物图像的多个第二模糊图像；将所述第二模糊图像输入到所述待训练的识别模型中，确定所述第二模糊图像的识别结果，包括：将所述多个第二模糊图像输入到所述待训练的识别模型中，确定所述多个第二模糊图像各自的识别结果；根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：根据所述多个第二模糊图像各自的识别结果及其权重，对所述多个第二模糊图像各自的识别结果进行加权平均，确定加权识别结果；根据所述加权识别结果，确定所述第一模糊图像的标注。
在一些实施例中，所述方法还包括：采用迭代训练的方法训练所述待训练的识别模型，一次迭代训练所述待训练的识别模型的方法包括：获取有标注的目标物图像，作为第二训练样本；将所述第一训练样本和所述第二训练样本输入到所述待训练的识别模型中，确定所述第一训练样本的识别结果和所述第二训练样本的识别结果；根据所述第一训练样本的识别结果及其标注，确定第一损失，以及根据所述第二训练样本的识别结果及其标注，确定第二损失，根据所述第一损失与所述第二损失，调整所述待训练的识别模型的模型参数值。
在一些实施例中，根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：获取有标注的目标物图像，作为第二训练样本；将所述无标注的目标物图像的所述第二模糊图像输入到基于所述第二训练样本确定的预识别模块中，确定所述第二模糊图像的预识别结果；根据所述第二模糊图像的识别结果和所述预识别结果，确定所述第一模糊图像的标注。
在一些实施例中，所述方法还包括：将所述第二训练样本输入到待训练的预识别模块中，确定所述第二训练样本的预识别结果，以所述第二训练样本的预识别结果和所述第二训练样本的标注之间的差异最小化为训练目标，调整所述预识别模块的模型参数值。
在一些实施例中，根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：根据所述第二模糊图像的识别结果及其置信度，确定所述第二模糊图像的分类结果；根据所述第二模糊图像的分类结果，确定所述第一模糊图像的标注。
在一些实施例中，采用迭代训练的方法训练所述待训练的识别模型，包括：针对一次迭代，根据该次迭代的所述第一训练样本的识别结果及其标注，确定所述第一模糊图像的样本权重；当下一次迭代训练所述识别模型时，响应于所述第一模糊图像的样本权重大于等于预设阈值，将所述第一模糊图像确定为所述下一次迭代的第一训练样本。
第二方面，本说明书提供了一种识别模型的训练装置，包括：获取模块，用于获取无标注的目标物图像；处理模块，对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像；识别模块，用于将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果；标注模块，用于根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注；训练模块，用于以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
第三方面，本说明书提供了一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述识别模型的训练方法。
第四方面，本说明书提供了一种电子设备，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程序，所述处理器执行所述程序时实现上述识别模型的训练方法。
本说明书采用的上述至少一个技术方案能够达到以下有益效果：在本说明书提供的识别模型的训练方法中，通过对获取到的目标物图像进行模糊化处理，确定第一模糊图像和第二模糊图像，通过待训练的识别模型，确定第二模糊图像的识别结果，并根据第二模糊图像的识别结果，确定第一模糊图像的标注，以第一模糊图像为第一训练样本，根据第一模糊图像的标注，对待训练的识别模型进行训练。
从上述方法中可以看出，本方法不需要人工对样本进行标注，减少了训练样本的生成时间与成本，提高了识别模型的训练效率。
附图说明
此处所说明的附图用来提供对本说明书的进一步理解，构成本说明书的一部分，本说明书的示意性实施例及其说明用于解释本说明书，并不构成对本说明书的不当限定。在附图中：
图1为本说明书提供的识别模型的训练方法的流程示意图；
图2为本说明书提供的对目标物图像进行模糊处理的示意图；
图3为本说明书提供的识别模型的示意图；
图4为本说明书提供的预识别模块的示意图；
图5为本说明书提供的确定损失函数的结构图；
图6为本说明书提供的识别模型的训练装置；
图7为本说明书提供的对应于图1的电子设备示意图。
具体实施方式
为使本说明书的目的、技术方案和优点更加清楚，下面将结合本说明书具体实施例及相应的附图对本说明书技术方案进行清楚、完整地描述。显然，所描述的实施例仅是本说明书一部分实施例，而不是全部的实施例。基于本说明书中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本说明书保护的范围。
以下结合附图，详细说明本说明书各实施例提供的技术方案。
图1为本说明书提供的识别模型的训练方法的流程示意图，具体包括以下步骤：
S100：获取若干无标注的目标物图像；
一般的，在无人驾驶技术领域，可通过检测模型和识别模型对无人驾驶设备周围的环境图像进行检测，确定目标物的位置及目标物的类型，以根据确定出的目标物的位置及类型，确定无人驾驶设备的运动策略。
通常，识别模型是由用于训练模型的服务器，基于训练样本预先训练得到的。而本说明书提供了一种识别模型的训练方法，同样的，可由用于训练模型的服务器执行该训练识别模型的过程。
训练模型可以分为样本生成阶段以及训练模型阶段，在样本生成阶段可根据模型需要以及训练需要，确定用于训练模型的训练样本。在本说明书中，首先可确定用于训练识别模型的训练样本，并且由于通常识别模型是基于目标物图像对目标物进行识别的，因此，可首先确定各目标物图像，以确定训练样本。其中，该目标物图像可为信号灯图像，也可为无人驾驶设备行驶过程中的交通指示标志等，为了方便描述，后续以信号灯为例进行说明。
由于通常该识别模型可基于将无人驾驶设备采集到的周围图像输入检测模型得到的目标物图像，对目标物进行识别，因此，该识别模型的目标物图像可基于检测模型确定。具体的，可首先获取若干无人驾驶设备采集到的图像(该图像中可以包括采集该图像的 无人驾驶设备周围的物体)，并将获取到的各无人驾驶设备采集到的图像输入该检测模型，确定各无人驾驶设备采集到的图像中的各目标物(如，信号灯)的位置，于是，可基于确定出的各目标物的位置对各图像进行裁剪，确定出各包含目标物的目标物图像。
在本说明书中，无人驾驶设备可以是指无人车、机器人、自动配送设备等能够实现自动驾驶的设备，也可以是半自动驾驶设备或辅助驾驶设备等，在本文不做限定。基于此，应用本说明书提供的训练方法训练完成的识别模型的无人驾驶设备可以用于执行配送任务，如，使用无人驾驶设备进行快递、物流、外卖等配送任务。为了保证无人驾驶设备在配送任务中的安全行驶，无人驾驶设备需要通过采集到的图像，对道路中设置的交通灯、交通指示牌等具体交通指示功能的静态目标物进行识别。
S102：针对每个目标物图像，对该目标物图像进行模糊化处理，确定该目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像；
在本说明书提供的一个或多个实施例中，在确定出各目标物图像后，可针对每个目标物图像，对该目标物图像进行模糊化处理，确定该目标物图像的第一模糊图像和第二模糊图像。以便后续可基于各目标物图像的第一模糊图像和第二模糊图像，对该识别模型进行训练。
具体的，针对每个无标注的目标物图像，该目标物图像的模糊图像的模糊强度越低，则该模糊图像与该目标物的差距越小，该模糊图像的识别结果与该目标物图像的识别结果差距也越小。而该目标物图像的模糊图像的模糊强度越高，则该模糊图像与该目标物的差距越大，该模糊图像的识别结果与该目标物图像的识别结果差距也越大。因此，在实施例中，可预设至少两种处理强度的模糊处理模型，在确定出各目标物图像后，可针对每个目标物图像，根据预设的具有至少两种处理强度的模糊处理模型，对该目标物图像进行模糊处理，得到模糊强度不同的第一模糊图像和第二模糊图像。如图2所示，左边部分为目标物图像，该目标物图像可以为信号灯图像，对该目标物图像分别进行第一模糊处理和第二模糊处理，可得到右边部分的该目标物图像的第一模糊图像和第二模糊图像。显然，第一模糊图像的模糊强度高于第二模糊图像。
S104：将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果。
在本说明书提供的一个或多个实施例中，在确定出各目标物图像的第二模糊图像后，还可将第二模糊图像输入到待训练的识别模型中，确定该第二模糊图像的识别结果。
具体的，在本说明书中，针对每个无标注的目标物图像，该目标物图像的第二模糊图像的识别结果更接近于该目标物图像的识别结果，于是，可采用根据该目标物图像的第二模糊图像的识别结果，确定该第一模糊图像的标注，在后续步骤中对该识别模型进行训练，使该识别模型可对于同一目标物图像的不同模糊图像，确定同一识别结果。因此，可针对每个无标注的目标物图像，将该目标物图像的第二模糊图像输入到待训练的识别模型中，确定该第二模糊图像的识别结果。如图3所示，左边部分为第二模糊图像，假设从上向下的信号灯分别为红灯、黄灯、绿灯，黑色代表亮灯，白色代表熄灭，则第二模糊图像对应的目标物图像对应的信号灯状态为绿色，中间部分为待训练的识别模型，而右边部分为第二模糊图像的识别结果。其中，识别结果为目标物分别属于各预设分类的概率，即，第二模糊图像分别对应的信号灯状态为熄灭、信号灯状态为红色、信号灯状态为黄色、信号灯状态为绿色分别对应的概率。于是，可确定该第二模糊图像的识别结果中，信号灯状态为绿色的概率较高。
S106：根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注。
在本说明书提供的一个或多个实施例中，在确定出第二模糊图像的识别结果后，还可根据第二模糊图像的识别结果，确定第一模糊图像的标注。
具体的，如前所述的，在本说明书中，针对每个目标物图像，可基于该目标物图像的第二模糊图像的识别结果，确定第一模糊图像的标注，并在后续步骤中，基于各第一模糊图像的识别结果与第一模糊图像的标注对识别模型进行训练。在一个例子中，在确定出第二模糊图像的识别结果后，可将该第二模糊图像的识别结果，作为该第一模糊图像的标注。
因为确定出的第二模糊图像的识别结果为目标物分别属于各预设分类的概率，而通常目标物属于预设的某一分类，如，信号灯状态为绿色，或信号灯状态为熄灭等。于是，可根据该第二模糊图像的识别结果及其置信度，确定该第二模糊图像的分类结果；根据该第二模糊图像的分类结果，确定该第一模糊图像的标注。例如，从该第二模糊图像的识别结果中，确定概率最高对应的状态，作为该第一模糊图像的标注。如，信号灯对应的状态为熄灭、红色、黄色、绿色的概率分别为10％、80％、2％、8％，则可确定该第一模糊图像的标注为红色。
S108：以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
在本说明书提供的一个或多个实施例中，在确定出各目标物图像的第一模糊图像及第一模糊图像的标注后，针对每个目标物图像，可以该第一模糊图像为第一训练样本，根据该第一模糊图像的标注，对待训练的识别模型进行训练。
具体的，可将第一模糊图像作为第一训练样本，输入到待训练的识别模型中，确定第一训练样本的识别结果。根据第一训练样本的识别结果及其标注，确定第一损失，并根据第一损失调整该识别模型的模型参数值，以完成对该识别模型的训练。
基于图1的识别模型的训练方法，通过对获取到的目标物图像进行模糊化处理，获得第一模糊图像和第二模糊图像，通过待训练的识别模型，确定第二模糊图像的识别结果，并根据第二模糊图像的识别结果，确定第一模糊图像的标注，以第一模糊图像为第一训练样本，根据第一模糊图像的标注，对待训练的识别模型进行训练。此方法不需要人工对样本进行标注，减少了训练样本的生成时间与成本，提高了识别模型的训练效率。
在步骤S102中，在确定该目标物图像的第二模糊图像时，还可对该目标物图像进行多种模糊化处理，获得该目标物图像的多个第二模糊图像。于是，在步骤S104中，确定第二模糊图像时，可将多个第二模糊图像输入待训练的识别模型中，确定多个第二模糊图像各自的识别结果。然后，在步骤S106中，可将多个第二模糊图像各自的识别结果及预设的各第二模糊图像的权重，输入加权函数
Figure PCTCN2022083705-appb-000001
其中，X i为各第二模糊图像的识别结果，a i为各第二模糊图像的识别结果对应的权重，k为各第二模糊图像的个数。可得到这些第二模糊图像的加权识别结果，作为该第一模糊图像的标注。在一个例子中，可以直接将该加权识别结果作为该第一模糊图像的标注。在另一个例子中，可将该加权识别结果中最高概率对应的状态作为该第一模糊图像的标注。其中所述加权识别结果为各第二模糊图像的识别结果经过加权平均计算后得到的结果，即在本实施例中，各第二模糊图像的识别结果对应的权重之和为1。最后，在步骤S108中，可基于步骤S106中确定出的标注，以该第一模糊图像的识别结果与其标注，对该模型进行训练。
在本说明书，为了使训练得到的识别模型对目标物图像的识别更准确，在步骤S106中，还可基于有标注的训练样本，辅助获得模糊图像的标注。于是，可基于有标注的训练样本，训练得到预识别模块，并针对每个无标注的目标物图像，将该目标物图像的第二模糊图像输入到预识别模块中，确定第二模糊图像的预识别结果，进而确定该第一模糊图像的标注。
具体的，可首先获取若干有标注的目标物图像，作为第二训练样本，然后，将各第 二训练样本输入到待训练的预识别模块中，确定各第二训练样本的预识别结果，根据各第二训练样本的预识别结果和其标注之间的差异最小化为训练目标，调整预识别模块的模型参数值，以得到预识别模块。并在确定出该预识别模块后，将各第二模糊图像输入到该预识别模块中，得到各第二模糊图像的识别结果，如图4所示。
图4为本说明书提供的预识别模块的示意图，左边部分为第二模糊图像，中间部分为基于各有标注的目标物图像预先训练得到的预识别模块，右边部分为第二模糊图像的识别结果，与图3类似，第二模糊图像对应的信号灯状态为绿色，预识别结果为目标物分别属于各预设分类的概率，即，第二模糊图像分别对应的信号灯状态为熄灭、信号灯状态为红色、信号灯状态为黄色、信号灯状态为绿色分别对应的概率。于是，步骤S106中，在确定出第二模糊图像的预识别结果后，可根据该第二模糊图像的识别结果及其权重，以及该第二模糊图像的预识别结果及其权重(这两个权重之和为1)，确定该第二模糊图像的加权识别结果，作为第一模糊图像的标注。最后，在步骤S108中，可基于步骤S106中确定出的标注，以该第一模糊图像的识别结果与其标注，对该模型进行训练。
在本说明书提供的一个或多个实施例中，为了得到更加准确的识别模型，本说明书可使用有标注的样本与无标注的样本一起对识别模型进行训练，于是，还可获取若干有标注的目标物图像，作为第二训练样本。将第一训练样本和第二训练样本作为输入，输入到待训练的识别模型中，确定第一训练样本的识别结果和第二训练样本的识别结果。根据该训练样本对的第一训练样本的识别结果及其标注，确定第一损失，以及根据第二训练样本的识别结果及其标注，确定第二损失，根据第一损失与第二损失，调整识别模型的模型参数，以完成对该识别模型的训练，如图5所示。
图5为本说明书提供的确定损失函数的结构图，与图3、图4类似，无标注的目标物图像、无标注的目标物图像的第一模糊图像、第二模糊图像对应的信号灯状态均为绿色，第二训练样本对应的信号灯状态为黄色。各识别结果为目标物分别属于各预设分类的概率，即，第二模糊图像分别对应的信号灯状态为熄灭、信号灯状态为红色、信号灯状态为黄色、信号灯状态为绿色分别对应的概率。于是，可以根据各第一训练样本和各第二训练样本确定各训练样本对，针对每个训练样本对，将该训练样本对中的所述第一训练样本和所述第二训练样本作为输入，输入到所述待训练的识别模型中，确定所述第一训练样本的识别结果和所述第二训练样本的识别结果，可根据第一训练样本的识别结果及其标注，确定第一损失，根据第二训练样本的识别结果及其标注，确定第二损失。 并以第一损失与第二损失之和最小为优化目标，对该识别模型进行训练。
当然，在步骤S106中确定第一模糊图像的标注时，还可根据各第二模糊图像的识别结果及其权重，以及至少一个第二模糊图像的预识别结果及权重，确定加权识别结果，作为该第一模糊图像的第一标注(例如，第一标注可以是0到1的一个概率值)，以及根据该加权识别结果，确定最高概率对应的状态，作为该第一模糊图像的第二标注(例如，以信号灯为例，如果根据加权识别结果确定的最高概率对应的状态为红色，则红色的标注为1，其他颜色的标注为0)。于是，在步骤S108中对识别模型进行训练时，可根据第一训练样本的识别结果及第一标注，确定第一损失，根据第二训练样本的识别结果及其标注，确定第二损失，以及根据第一训练样本的识别结果及第二标注，确定第三损失。并以第一损失、第二损失、与第三损失之和最小为优化目标，对该识别模型进行训练。
在训练模型的过程中，可采用迭代训练的方法训练识别模型，一次迭代训练识别模型的方法包括，确定第一训练样本，并基于各第一训练样本对识别模型进行训练。而针对一次迭代，可根据该次迭代的第一训练样本的识别结果及其标注，确定第一模糊图像的样本权重。当下一次迭代训练识别模型时，根据第一模糊图像的样本权重，确定该第一模糊图像是否可以用作下一次迭代的第一训练样本，该样本权重与第一模糊图像用于作为第一训练样本的概率正相关。响应于第一模糊图像的样本权重大于等于预设阈值，将该第一模糊图像确定为下一次迭代的第一训练样本。
具体的，可根据第一训练样本的当前权重(可预设为100％)，以及该第一模糊图像对应的第二模糊图像的识别结果确定该第一模糊图像的标注是否满足第一降权条件，所述第一降权条件可以是，该第一模糊图像的标注中，各类型的概率均小于30％，如，可认为标注中各类型的概率均小于30％的训练样本为背景，若满足第一降权条件，则降低第一模糊图像的当前权重；或确定该第一模糊图像的识别结果是否满足第二降权条件，所述第二降权条件可以是，该第一模糊图像的识别结果中，存在概率不小于70％的类型，如，可认为识别结果中存在概率不小于70％的类型的训练样本为简单样本，若满足第二降权条件，则降低第一模糊图像的当前权重。将第一模糊图像的降低后的当前权重确定为该第一模糊图像的样本权重，如果第一模糊图像的样本权重大于或等于概率阈值，则将该第一模糊图像确定为第一训练样本。
在本说明书提供的一个或多个实施例中，在训练好该识别模型后，可将检测模型与识别模型进行组合，并将该识别模型应用于确定无人驾驶设备环境图像中的各目标物类 型的场景中，以使无人驾驶设备能够基于确定出的目标物类型，确定下一时刻该无人驾驶设备的运动策略，并通过采取相应的策略，以使下一时刻无人驾驶设备正常行驶而不发生交通事故。具体的根据目标物类型与位置等确定无人驾驶设备运动策略的方法已较为成熟，本说明书对此不再赘述。
需要说明的是，在本说明书中，第一训练样本的数量可能少于第二训练样本，因此，在确定训练样本对时，可重复对第一训练样本进行选择，当然，为了得到更多的训练样本，还可重复对第二训练样本对进行选择，具体的确定训练样本的方法可根据需要进行设置，本说明书对此不做限制。
另外，在本说明书的步骤S108中，基于各第一模糊图像及其标注训练得到的识别模型，可对同一目标物图像的不同模糊强度的模糊图像，输出同一识别结果。于是，基于大量训练样本训练得到的识别模型，可对各输入的目标物图像输出较为准确的识别结果，但是该识别结果的准确性是相对于模型来说的，即，目标物通过识别模型得到的识别结果与该目标物本身所属类型可能是错位的。如，若干信号灯状态为红色的目标物图像，经该识别模型得到的识别结果的信号灯状态都为绿色。因此，训练完成的识别模型在实际使用中，可对其识别结果进行映射，以输出正确的识别结果。
需要说明的是，对目标物图像的模糊处理不限于使目标物图像变得模糊，只要目标物图像经模糊处理后，得到与目标物图像的相似度为第一相似度的第一模糊图像和与目标物图像的相似度为第二相似度的第二模糊图像，所述第一相似度小于所述第二相似度即可。在实际使用中，对图像进行的模糊处理还可以包括腐蚀处理，即，确定目标物图像的第一腐蚀图像与第二腐蚀图像，第一腐蚀图像的腐蚀强度高于第二腐蚀图像，基于第二腐蚀图像确定第一腐蚀图像的标注，并以第一腐蚀图像的识别结果与标注对该识别模型进行训练等。同样，对目标物图像进行的模糊处理还可以包括仿射变换处理。在本说明书中，将以上模糊处理、腐蚀处理、仿射变换处理等统称为模糊处理。当然，目标物图像的模糊图像还可为各模糊处理叠加的结果，具体如何确定模糊图像可根据需要进行设置，本说明书对此不做限制。
需要说明的是，具体的加权条件和降权条件等可根据需要进行设置，本说明书对此不做限制。
以上为本说明书的一个或多个实施例提供的识别模型的训练方法基于同样的思路，本说明书还提供了相应的识别模型的训练装置，如图6所示。
图6为本说明书提供的识别模型的训练装置，包括：
获取模块200，用于获取无标注的目标物图像。
处理模块202，用于对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像。
识别模块204，用于将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果。
标注模块206，用于根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注。
训练模块208，用于以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
可选地，所述处理模块202，具体用于对所述无标注的目标物图像进行多种模糊化处理，获得所述无标注的目标物图像的多个第二模糊图像；
可选的，所述识别模块204，具体用于将所述多个第二模糊图像输入到所述待训练的识别模型中，确定所述多个第二模糊图像各自的识别结果；
可选地，所述标注模块206，具体用于根据所述多个第二模糊图像各自的识别结果及其权重，对所述多个第二模糊图像各自的识别结果进行加权平均，确定加权识别结果，根据所述加权识别结果，确定所述第一模糊图像的标注。
可选地，所述训练模块208，具体用于采用迭代训练的方法训练所述待训练的识别模型，一次迭代训练所述待训练的识别模型的方法包括：获取有标注的目标物图像，作为第二训练样本，将所述第一训练样本和所述第二训练样本输入到所述待训练的识别模型中，确定所述第一训练样本的识别结果和所述第二训练样本的识别结果，根据所述第一训练样本的识别结果及其标注，确定第一损失，以及根据所述第二训练样本的识别结果及其标注，确定第二损失，根据所述第一损失与所述第二损失，调整所述待训练的识别模型的模型参数值。
可选地，所述标注模块206，具体用于获取有标注的目标物图像，作为第二训练样本，将所述无标注的目标物图像的所述第二模糊图像输入到基于所述第二训练样本确定的预识别模块中，确定所述第二模糊图像的预识别结果，根据所述第二模糊图像的识别 结果和所述预识别结果，确定该第一模糊图像的标注。
可选地，所述训练模块208，还用于将所述第二训练样本输入到待训练的预识别模块中，确定所述第二训练样本的预识别结果，以所述第二训练样本的预识别结果和所述第二训练样本的标注之间的差异最小化为训练目标，调整所述预识别模块的模型参数值。
可选地，所述标注模块206，具体用于根据所述第二模糊图像的识别结果及其置信度，确定所述第二模糊图像的分类结果，根据所述第二模糊图像的分类结果，确定所述第一模糊图像的标注。
可选地，所述训练模块208，还用于针对一次迭代，根据该次迭代的所述第一训练样本的识别结果及其标注，确定所述第一模糊图像的样本权重，当下一次迭代训练所述识别模型时，响应于所述第一模糊图像的样本权重大于等于预设阈值，将所述第一模糊图像确定为所述下一次迭代的第一训练样本。
本说明书还提供了一种计算机可读存储介质，该存储介质存储有计算机程序，计算机程序可用于执行上述图1提供的识别模型的训练方法。
本说明书还提供了图7所示的电子设备的示意结构图。如图7所述，在硬件层面，该电子设备包括处理器、内部总线、网络接口、内存以及非易失性存储器，当然还可能包括其他业务所需要的硬件。处理器从非易失性存储器中读取对应的计算机程序到内存中然后运行，以实现上述图1所述的识别模型的训练方法。当然，除了软件实现方式之外，本说明书并不排除其他实现方式，比如逻辑器件抑或软硬件结合的方式等等，也就是说以下处理流程的执行主体并不限定于各个逻辑单元，也可以是硬件或逻辑器件。
在20世纪90年代，对于一个技术的改进可以很明显地区分是硬件上的改进(例如，对二极管、晶体管、开关等电路结构的改进)还是软件上的改进(对于方法流程的改进)。然而，随着技术的发展，当今的很多方法流程的改进已经可以视为硬件电路结构的直接改进。设计人员几乎都通过将改进的方法流程编程到硬件电路中来得到相应的硬件电路结构。因此，不能说一个方法流程的改进就不能用硬件实体模块来实现。例如，可编程逻辑器件(Programmable Logic Device,PLD)(例如现场可编程门阵列(Field Programmable Gate Array，FPGA))就是这样一种集成电路，其逻辑功能由用户对器件编程来确定。由设计人员自行编程来把一个数字系统“集成”在一片PLD上，而不需要请芯片制造厂商来设计和制作专用的集成电路芯片。而且，如今，取代手工地制作集成电路芯片，这种编程也多半改用“逻辑编译器(logic compiler)”软件来实现，它与程序开 发撰写时所用的软件编译器相类似，而要编译之前的原始代码也得用特定的编程语言来撰写，此称之为硬件描述语言(Hardware Description Language，HDL)，而HDL也并非仅有一种，而是有许多种，如ABEL(Advanced Boolean Expression Language)、AHDL(Altera Hardware Description Language)、Confluence、CUPL(Cornell University Programming Language)、HDCal、JHDL(Java Hardware Description Language)、Lava、Lola、MyHDL、PALASM、RHDL(Ruby Hardware Description Language)等，目前最普遍使用的是VHDL(Very-High-Speed Integrated Circuit Hardware Description Language)与Verilog。本领域技术人员也应该清楚，只需要将方法流程用上述几种硬件描述语言稍作逻辑编程并编程到集成电路中，就可以很容易得到实现该逻辑方法流程的硬件电路。
控制器可以按任何适当的方式实现，例如，控制器可以采取例如微处理器或处理器以及存储可由该(微)处理器执行的计算机可读程序代码(例如软件或固件)的计算机可读介质、逻辑门、开关、专用集成电路(Application Specific Integrated Circuit，ASIC)、可编程逻辑控制器和嵌入微控制器的形式，控制器的例子包括但不限于以下微控制器：ARC 625D、Atmel AT91SAM、Microchip PIC18F26K20以及Silicone Labs C8051F320，存储器控制器还可以被实现为存储器的控制逻辑的一部分。本领域技术人员也知道，除了以纯计算机可读程序代码方式实现控制器以外，完全可以通过将方法步骤进行逻辑编程来使得控制器以逻辑门、开关、专用集成电路、可编程逻辑控制器和嵌入微控制器等的形式来实现相同功能。因此这种控制器可以被认为是一种硬件部件，而对其内包括的用于实现各种功能的装置也可以视为硬件部件内的结构。或者甚至，可以将用于实现各种功能的装置视为既可以是实现方法的软件模块又可以是硬件部件内的结构。
上述实施例阐明的系统、装置、模块或单元，具体可以由计算机芯片或实体实现，或者由具有某种功能的产品来实现。一种典型的实现设备为计算机。具体的，计算机例如可以为个人计算机、膝上型计算机、蜂窝电话、相机电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任何设备的组合。
为了描述的方便，描述以上装置时以功能分为各种单元分别描述。当然，在实施本说明书时可以把各单元的功能在同一个或多个软件和/或硬件中实现。
本领域内的技术人员应明白，本发明的实施例可提供为方法、系统、或计算机程序产品。因此，本发明可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本发明可采用在一个或多个其中包含有计算机可用程序代码 的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
在一个典型的配置中，计算设备包括一个或多个处理器(CPU)、输入/输出接口、网络接口和内存。
内存可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/或非易失性内存等形式，如只读存储器(ROM)或闪存(flash RAM)。内存是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备 访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可读媒体(transitory media)，如调制的数据信号和载波。
还需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、商品或者设备中还存在另外的相同要素。
本领域技术人员应明白，本说明书的实施例可提供为方法、系统或计算机程序产品。因此，本说明书可采用完全硬件实施例、完全软件实施例或结合软件和硬件方面的实施例的形式。而且，本说明书可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本说明书可以在由计算机执行的计算机可执行指令的一般上下文中描述，例如程序模块。一般地，程序模块包括执行特定任务或实现特定抽象数据类型的例程、程序、对象、组件、数据结构等等。也可以在分布式计算环境中实践本说明书，在这些分布式计算环境中，由通过通信网络而被连接的远程处理设备来执行任务。在分布式计算环境中，程序模块可以位于包括存储设备在内的本地和远程计算机存储介质中。
本说明书中的各个实施例均采用递进的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于系统实施例而言，由于其基本相似于方法实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
以上所述仅为本说明书的实施例而已，并不用于限制本说明书。对于本领域技术人员来说，本说明书可以有各种更改和变化。凡在本说明书的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本说明书的权利要求范围之内。

Claims (10)

  1. 一种识别模型的训练方法，其特征在于，所述方法包括：
    获取无标注的目标物图像；
    对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像；
    将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果；
    根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注；
    以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
  2. 如权利要求1所述的方法，其特征在于，对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第二模糊图像，包括：
    对所述无标注的目标物图像进行多种模糊化处理，获得所述无标注的目标物图像的多个第二模糊图像；
    将所述第二模糊图像输入到所述待训练的识别模型中，确定所述第二模糊图像的识别结果，包括：
    将所述多个第二模糊图像输入到所述待训练的识别模型中，确定所述多个第二模糊图像各自的识别结果；
    根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：
    根据所述多个第二模糊图像各自的识别结果及其权重，对所述多个第二模糊图像各自的识别结果进行加权平均，确定加权识别结果；
    根据所述加权识别结果，确定所述第一模糊图像的标注。
  3. 如权利要求1所述的方法，其特征在于，所述方法还包括：
    采用迭代训练的方法训练所述待训练的识别模型，一次迭代训练所述待训练的识别模型的方法包括：
    获取有标注的目标物图像，作为第二训练样本；
    将所述第一训练样本和所述第二训练样本输入到所述待训练的识别模型中，确定所述第一训练样本的识别结果和所述第二训练样本的识别结果；
    根据所述第一训练样本的识别结果及其标注，确定第一损失，以及根据所述第二训练样本的识别结果及其标注，确定第二损失，根据所述第一损失与所述第二损失，调整所述待训练的识别模型的模型参数值。
  4. 如权利要求1所述的方法，其特征在于，根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：
    获取有标注的目标物图像，作为第二训练样本；
    将所述无标注的目标物图像的所述第二模糊图像输入到基于所述第二训练样本确定的预识别模块中，确定所述第二模糊图像的预识别结果；
    根据所述第二模糊图像的识别结果和所述预识别结果，确定所述第一模糊图像的标注。
  5. 如权利要求4所述的方法，其特征在于，所述方法还包括：
    将所述第二训练样本输入到待训练的预识别模块中，确定所述第二训练样本的预识别结果，以所述第二训练样本的预识别结果和所述第二训练样本的标注之间的差异最小化为训练目标，调整所述预识别模块的模型参数值。
  6. 如权利要求1所述的方法，其特征在于，根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注，包括：
    根据所述第二模糊图像的识别结果及其置信度，确定所述第二模糊图像的分类结果；
    根据所述第二模糊图像的分类结果，确定所述第一模糊图像的标注。
  7. 如权利要求3所述的方法，其特征在于，采用迭代训练的方法训练所述待训练的识别模型，包括：
    针对一次迭代，根据该次迭代的所述第一训练样本的识别结果及其标注，确定所述第一模糊图像的样本权重；
    当下一次迭代训练所述识别模型时，响应于所述第一模糊图像的样本权重大于等于预设阈值，将所述第一模糊图像确定为所述下一次迭代的第一训练样本。
  8. 一种识别模型的训练装置，其特征在于，所述装置包括：
    获取模块，用于获取无标注的目标物图像；
    处理模块，对所述无标注的目标物图像进行模糊化处理，获得所述无标注的目标物图像的第一模糊图像和第二模糊图像，所述第一模糊图像的模糊强度高于所述第二模糊图像；
    识别模块，用于将所述第二模糊图像输入到待训练的识别模型中，确定所述第二模糊图像的识别结果；
    标注模块，用于根据所述第二模糊图像的识别结果，确定所述第一模糊图像的标注；
    训练模块，用于以所述第一模糊图像为第一训练样本，根据所述第一模糊图像的标注，对待训练的识别模型进行训练。
  9. 一种计算机可读存储介质，其特征在于，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述权利要求1～7任一项所述的方法。
  10. 一种电子设备，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程序，其特征在于，所述处理器执行所述程序时实现上述权利要求1～7任一项所述的方法。
PCT/CN2022/083705 2021-04-07 2022-03-29 识别模型的训练 WO2022213846A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110369658.X 2021-04-07
CN202110369658.XA CN112801229B (zh) 2021-04-07 2021-04-07 一种识别模型的训练方法及装置

Publications (1)

Publication Number Publication Date
WO2022213846A1 true WO2022213846A1 (zh) 2022-10-13

Family

ID=75816425

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/083705 WO2022213846A1 (zh) 2021-04-07 2022-03-29 识别模型的训练

Country Status (2)

Country Link
CN (1) CN112801229B (zh)
WO (1) WO2022213846A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115828247A (zh) * 2023-02-10 2023-03-21 支付宝(杭州)信息技术有限公司 一种小程序的异常检测方法、装置、设备及可读存储介质
CN116309924A (zh) * 2023-05-25 2023-06-23 之江实验室 一种模型训练的方法、图像展示的方法及装置
CN116402113A (zh) * 2023-06-08 2023-07-07 之江实验室 一种任务执行的方法、装置、存储介质及电子设备
CN116661574A (zh) * 2023-07-13 2023-08-29 之江实验室 一种计算设备散热的方法、装置、存储介质及电子设备
CN117036869A (zh) * 2023-10-08 2023-11-10 之江实验室 一种基于多样性和随机策略的模型训练方法及装置
CN117036870A (zh) * 2023-10-09 2023-11-10 之江实验室 一种基于积分梯度多样性的模型训练和图像识别方法
CN117237744A (zh) * 2023-11-10 2023-12-15 之江实验室 一种图像分类模型的训练方法、装置、介质及电子设备

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801229B (zh) * 2021-04-07 2021-07-30 北京三快在线科技有限公司 一种识别模型的训练方法及装置
CN113850301A (zh) * 2021-09-02 2021-12-28 支付宝(杭州)信息技术有限公司 训练数据的获取方法和装置、模型训练方法和装置
CN116755862B (zh) * 2023-08-11 2023-12-19 之江实验室 一种算子优化调度模型的训练方法、装置、介质及设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108875768A (zh) * 2018-01-23 2018-11-23 北京迈格威科技有限公司 数据标注方法、装置和系统及存储介质
CN109784391A (zh) * 2019-01-04 2019-05-21 杭州比智科技有限公司 基于多模型的样本标注方法及装置
CN111539309A (zh) * 2020-04-21 2020-08-14 广州云从鼎望科技有限公司 一种基于ocr的数据处理方法、系统、平台、设备及介质
CN111967429A (zh) * 2020-08-28 2020-11-20 清华大学 一种基于主动学习的行人重识别模型训练方法及装置
CN112801229A (zh) * 2021-04-07 2021-05-14 北京三快在线科技有限公司 一种识别模型的训练方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109460795A (zh) * 2018-12-17 2019-03-12 北京三快在线科技有限公司 分类器训练方法、装置、电子设备及计算机可读介质
CN112308144A (zh) * 2020-10-30 2021-02-02 江苏云从曦和人工智能有限公司 一种筛选样本的方法、系统、设备及介质
CN112308143A (zh) * 2020-10-30 2021-02-02 江苏云从曦和人工智能有限公司 一种基于多样性的样本筛选方法、系统、设备及介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108875768A (zh) * 2018-01-23 2018-11-23 北京迈格威科技有限公司 数据标注方法、装置和系统及存储介质
CN109784391A (zh) * 2019-01-04 2019-05-21 杭州比智科技有限公司 基于多模型的样本标注方法及装置
CN111539309A (zh) * 2020-04-21 2020-08-14 广州云从鼎望科技有限公司 一种基于ocr的数据处理方法、系统、平台、设备及介质
CN111967429A (zh) * 2020-08-28 2020-11-20 清华大学 一种基于主动学习的行人重识别模型训练方法及装置
CN112801229A (zh) * 2021-04-07 2021-05-14 北京三快在线科技有限公司 一种识别模型的训练方法及装置

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115828247A (zh) * 2023-02-10 2023-03-21 支付宝(杭州)信息技术有限公司 一种小程序的异常检测方法、装置、设备及可读存储介质
CN116309924A (zh) * 2023-05-25 2023-06-23 之江实验室 一种模型训练的方法、图像展示的方法及装置
CN116309924B (zh) * 2023-05-25 2023-08-18 之江实验室 一种模型训练的方法、图像展示的方法及装置
CN116402113A (zh) * 2023-06-08 2023-07-07 之江实验室 一种任务执行的方法、装置、存储介质及电子设备
CN116402113B (zh) * 2023-06-08 2023-10-03 之江实验室 一种任务执行的方法、装置、存储介质及电子设备
CN116661574B (zh) * 2023-07-13 2023-12-12 之江实验室 一种计算设备散热的方法、装置、存储介质及电子设备
CN116661574A (zh) * 2023-07-13 2023-08-29 之江实验室 一种计算设备散热的方法、装置、存储介质及电子设备
CN117036869A (zh) * 2023-10-08 2023-11-10 之江实验室 一种基于多样性和随机策略的模型训练方法及装置
CN117036869B (zh) * 2023-10-08 2024-01-09 之江实验室 一种基于多样性和随机策略的模型训练方法及装置
CN117036870A (zh) * 2023-10-09 2023-11-10 之江实验室 一种基于积分梯度多样性的模型训练和图像识别方法
CN117036870B (zh) * 2023-10-09 2024-01-09 之江实验室 一种基于积分梯度多样性的模型训练和图像识别方法
CN117237744A (zh) * 2023-11-10 2023-12-15 之江实验室 一种图像分类模型的训练方法、装置、介质及电子设备
CN117237744B (zh) * 2023-11-10 2024-01-30 之江实验室 一种图像分类模型的训练方法、装置、介质及电子设备

Also Published As

Publication number Publication date
CN112801229A (zh) 2021-05-14
CN112801229B (zh) 2021-07-30

Similar Documents

Publication Publication Date Title
WO2022213846A1 (zh) 识别模型的训练
Wang et al. Improved YOLOv5 network for real-time multi-scale traffic sign detection
WO2019228211A1 (zh) 基于车道线的智能驾驶控制方法和装置、电子设备
JP7073247B2 (ja) 車線境界線検出モデルを生成するための方法、車線境界線を検出するための方法、車線境界線検出モデルを生成するための装置、車線境界線を検出するための装置、機器、コンピュータ読み取り可能な記憶媒体およびコンピュータプログラム
CN107358242B (zh) 目标区域颜色识别方法、装置及监控终端
CN107851213B (zh) 神经网络中的转移学习
US20210271862A1 (en) Expression recognition method and related apparatus
WO2022213867A1 (zh) 目标物识别
CN109460704B (zh) 一种基于深度学习的疲劳检测方法、系统和计算机设备
CN107545263B (zh) 一种物体检测方法及装置
CN110782420A (zh) 一种基于深度学习的小目标特征表示增强方法
EP4145345A1 (en) Data annotation method and device, and fine granularity identification method and device
CN112686274B (zh) 目标对象的检测方法及设备
Chen et al. An attention based YOLOv5 network for small traffic sign recognition
Muthalagu et al. Vehicle lane markings segmentation and keypoint determination using deep convolutional neural networks
CN115937879A (zh) 基于多尺度特征融合网络的学术内容目标检测方法及系统
Ren et al. A new multi-scale pedestrian detection algorithm in traffic environment
EP4016394A1 (en) Object identification device and object identification method
CN114332808A (zh) 一种预测转向意图的方法及装置
Gao et al. Real-time chinese traffic warning signs recognition based on cascade and CNN
CN114495050A (zh) 一种面向自动驾驶前向视觉检测的多任务集成检测方法
CN114972492A (zh) 一种基于鸟瞰图的位姿确定方法、设备和计算机存储介质
CN114187355A (zh) 一种图像标定方法及装置
CN113505763A (zh) 关键点检测方法、装置、电子设备及存储介质
CN111783811A (zh) 伪标签生成方法和装置

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: 22783913

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: 22783913

Country of ref document: EP

Kind code of ref document: A1