WO2020048242A1 - 基于gan网络的车损图像生成方法和装置 - Google Patents

基于gan网络的车损图像生成方法和装置 Download PDF

Info

Publication number
WO2020048242A1
WO2020048242A1 PCT/CN2019/096526 CN2019096526W WO2020048242A1 WO 2020048242 A1 WO2020048242 A1 WO 2020048242A1 CN 2019096526 W CN2019096526 W CN 2019096526W WO 2020048242 A1 WO2020048242 A1 WO 2020048242A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
target frame
model
vehicle
vehicle damage
Prior art date
Application number
PCT/CN2019/096526
Other languages
English (en)
French (fr)
Inventor
徐娟
Original Assignee
阿里巴巴集团控股有限公司
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 阿里巴巴集团控股有限公司 filed Critical 阿里巴巴集团控股有限公司
Priority to EP19857070.7A priority Critical patent/EP3848888A4/en
Priority to SG11202100107QA priority patent/SG11202100107QA/en
Publication of WO2020048242A1 publication Critical patent/WO2020048242A1/zh
Priority to US17/141,790 priority patent/US11972599B2/en

Links

Images

Classifications

    • 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
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • 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
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/40—Filling a planar surface by adding surface attributes, e.g. colour or texture
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/60—Editing figures and text; Combining figures or text
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06V30/19173—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/26—Techniques for post-processing, e.g. correcting the recognition result
    • G06V30/262—Techniques for post-processing, e.g. correcting the recognition result using context analysis, e.g. lexical, syntactic or semantic context
    • G06V30/274—Syntactic or semantic context, e.g. balancing
    • 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
    • G06V2201/08—Detecting or categorising vehicles

Definitions

  • the embodiments of the present specification relate to the field of image processing technology, and more particularly, to a method and device for training a car damage image discrimination model, a method and device for training an image filling model, and a method and device for generating a car damage image. .
  • an insurance company needs to send a professional inspector to determine the damage at the accident site, provide a repair plan and compensation amount for the vehicle, take a photo of the scene, and keep the fixed damage photo for the record The inspector verifies the damage and price. Due to the need to manually inspect and determine the damage, insurance companies need to invest a lot of labor costs and professional training costs. From the experience of ordinary users, the process of claim settlement is long due to waiting for a manual surveyor to take pictures on the scene, the rater to determine the damage at the repair site, and the rater to verify the loss in the background.
  • the embodiments of the present specification are intended to provide a more effective solution for acquiring a car damage image, so as to solve the deficiencies in the prior art.
  • one aspect of the present specification provides a computer-implemented method for generating a car damage image, including:
  • the second image being obtained by labeling a target frame on the first image and pruning a partial image within the target frame;
  • the first image includes or does not include a local image of a vehicle injury.
  • the first image includes at least one local image of vehicle damage
  • labeling a target frame on the first image includes randomly labeling the target frame at the at least one local image of vehicle damage.
  • labeling the target frame on the first image includes labeling the target frame at a first position, wherein the first position is a position randomly determined from a plurality of second positions, and the second The position is a highly probable position where the vehicle is damaged by statistics.
  • removing the partial image in the target frame includes removing a partial image in the target frame by performing a dot product operation on the first image using a mask.
  • the image filling model is obtained by training a generation model in a GAN model, and the GAN model further includes a discriminant model, wherein the discriminant model is used to discriminate whether an output image of the generated model is real The image, and whether the target image of the output image is a local image of the vehicle damage.
  • the generation model in the training GAN model includes:
  • the discriminant model is trained in the following manner:
  • a plurality of positive samples and a plurality of negative samples are obtained, and the positive samples and the negative samples are both vehicle images including a target frame, wherein the positive samples are real images, and the local areas within the target frame of the positive samples are The image is a local image of the vehicle damage, wherein the plurality of negative samples include at least one first negative sample, and the first negative sample is a non-real image obtained by replacing a partial image in a target frame of a real image with another partial image Images; and
  • a classification model is trained using the plurality of positive samples and the plurality of negative samples as the discrimination model.
  • the first negative sample includes at least one of the following characteristics: inconsistent components inside and outside the target frame, inconsistent models inside and outside the target frame, inconsistent colors inside and outside the target frame, and inconsistent textures inside and outside the target frame.
  • the plurality of negative samples further include at least one second negative sample, and the second negative sample is a real image that does not contain a local image of the vehicle damage in the target frame.
  • the discrimination model includes a semantic recognition model, and the semantic recognition model is used to determine whether the sample target frame contains a local image of a vehicle damage.
  • the method further comprises, after generating the car damage image, using the car damage image for training a vehicle damage recognition model, wherein the vehicle damage recognition model is used for vehicle-based car damage Image identifies vehicle damage.
  • Another aspect of this specification provides a computer-implemented device for generating a car damage image, including:
  • a first obtaining unit configured to obtain a first image, where the first image is a real image of a vehicle
  • a second obtaining unit configured to obtain a second image based on the first image, the second image being obtained by labeling a target frame on the first image and pruning a partial image within the target frame;
  • the input unit is configured to input the second image into an image filling model to obtain a vehicle damage image with a target frame from an output of the image filling model, wherein the image filling model passes the second image
  • the target frame of the vehicle is filled with a partial image of the vehicle damage and the vehicle damage image with the target frame is output.
  • the first image includes or does not include a local image of a vehicle injury.
  • the first image includes at least one partial image of a vehicle damage
  • the second acquisition unit is further configured to randomly mark the target frame at the at least one partial image of a vehicle damage.
  • the second obtaining unit is further configured to mark a target frame at a first position, wherein the first position is a position randomly determined from a plurality of second positions, and the second position is A highly probable location of vehicle damage through statistical acquisition of a vehicle.
  • the second obtaining unit is further configured to perform a dot product operation on the first image by using a mask to cut out a partial image within the target frame.
  • the image filling model is trained by a first training device, the first training device is used to train a generative model in a GAN model, and the GAN model further includes a discriminant model, wherein the discriminant model is used
  • the first training device includes: determining whether an output image of the generated model is a real image, and whether a target image of the output image is a local image of a vehicle injury;
  • a third acquisition unit configured to acquire a plurality of third images, where the third images are real images of the vehicle
  • a fourth obtaining unit configured to obtain a plurality of fourth images based on the plurality of third images, where the fourth image is marked with a target frame on the corresponding third image, and the target frame is cut out; Partial image acquisition; and
  • the first training unit is configured to use the at least the plurality of fourth images to train the generated model as the image filling model based on the discriminant model.
  • the discrimination model is trained by a second training device, and the second training device includes:
  • a fifth obtaining unit is configured to obtain a plurality of positive samples and a plurality of negative samples, where the positive samples and the negative samples are vehicle images including a target frame, wherein the positive samples are real images, and the The local image in the target frame of the positive sample is a local image of the vehicle damage, wherein the plurality of negative samples include at least one first negative sample, and the first negative sample is obtained by replacing the local image in the target frame of the real image with Unrealistic images obtained from other partial images; and
  • the second training unit is configured to use the plurality of positive samples and the plurality of negative samples to train a classification model as the discrimination model.
  • the first negative sample includes at least one of the following characteristics: inconsistent components inside and outside the target frame, inconsistent models inside and outside the target frame, inconsistent colors inside and outside the target frame, and inconsistent textures inside and outside the target frame.
  • the plurality of negative samples further include at least one second negative sample, and the second negative sample is a real image that does not contain a local image of the vehicle damage in the target frame.
  • the discrimination model includes a semantic recognition model, and the semantic recognition model is used to determine whether the sample target frame contains a local image of a vehicle damage.
  • the device further includes a using unit configured to use the vehicle damage image to train a vehicle damage recognition model after generating the vehicle damage image, wherein the vehicle damage recognition model is used for Vehicle damage is identified based on the vehicle damage image.
  • Another aspect of this specification provides a computing device including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, any one of the foregoing methods is implemented.
  • the image generated by the generated model can be used as a training sample with labels for training a vehicle damage recognition model, so that manual labeling is not required, and a large number of labels can be generated directly by generating the model.
  • the data can also be exhaustive of samples of various situations, such as vehicle models, lighting conditions, old and new, shooting angles, etc., so that the accuracy of the vehicle damage recognition model can be improved.
  • FIG. 1 shows a schematic diagram of a vehicle damage image generation system 100 according to an embodiment of the present specification
  • FIG. 2 shows a flowchart of a method for training a car damage image discrimination model according to an embodiment of the present specification
  • FIG. 3 shows a flowchart of a method for training an image filling model according to an embodiment of the present specification
  • FIG. 4 shows a flowchart of a computer-implemented method for generating a car damage image according to an embodiment of the present specification
  • FIG. 5 illustrates an apparatus 500 for training a car damage image discrimination model according to an embodiment of the present specification
  • FIG. 6 illustrates an apparatus 600 for training an image filling model according to an embodiment of the present specification.
  • FIG. 7 illustrates a computer-implemented apparatus 700 for generating a car damage image according to an embodiment of the present specification.
  • FIG. 1 shows a schematic diagram of a vehicle damage image generation system 100 according to an embodiment of the present specification.
  • the system 100 includes a mask 11, a generation model 12, and a discrimination model 13.
  • the generation model 12 and the discrimination model 13 constitute a generation adversarial model (GAN).
  • GAN generation adversarial model
  • the mask 11 is used to determine the position of the target frame on the actually captured vehicle image and cut out the partial image within the target frame.
  • the discrimination model 13 is first trained by at least one positive sample and / or at least one negative sample.
  • the positive sample is, for example, an image obtained by taking a photograph of a vehicle damage of an accident vehicle, and the image includes a target frame for marking the vehicle damage.
  • the negative sample is, for example, an image obtained by culling a partial image in a target frame in the above image and filling other car damage images.
  • the loss function of the discriminant model 13 is related to the semantic content included inside and outside the target frame, and to the smoothness inside and outside the target frame.
  • the determination network 13 determines whether the vehicle components inside and outside the target frame in the image are the same component, whether the image inside the target frame is a car damage image, and determines the color and texture continuity inside and outside the target frame.
  • the discriminative model 13 can be used to train and generate a model 12. Specifically, at least one actually captured vehicle image (ie, the first image) is input to the mask 11 to obtain at least one training sample (the second image), and the training sample is the vehicle actually captured through the mask 11 as described above An image obtained by determining the position of the target frame on the image and cutting out a partial image within the target frame.
  • the generating model 12 is trained using the at least one training sample, so that the loss function of the generating model 12 is reduced.
  • the loss function of the generating network 12 is obtained based on the discriminant model 13 for discriminating at least one output image, wherein the at least one output image is separately generated by filling the target frame of the at least one second image by the generating model. Image. That is, the training model 12 is trained so that the discrimination value of the discrimination model 13 for the car damage image generated by the generation model 12 is increased, that is, more like a real image.
  • FIG. 2 shows a flowchart of a method for training the discrimination model according to an embodiment of the present specification.
  • the discrimination model is a classification model including a convolutional neural network.
  • the method is a training process for training the discriminant model, and includes:
  • step S202 at least one positive sample and / or at least one negative sample are obtained, and the positive sample and the negative sample are both vehicle images including a target frame, wherein the positive sample is a real image and the positive sample
  • the local image in the target frame of is a vehicle damage local image
  • the at least one negative sample includes a non-real image obtained by replacing the local image in the target frame of the real image with another partial image
  • step S204 the discriminant model is trained using the at least one positive sample and / or at least one negative sample, so that the loss function of the discriminant model after training is reduced compared to before training, wherein the The loss function is related to discriminating the authenticity of each of the at least one positive sample and / or the at least one negative sample.
  • step S202 at least one positive sample and / or at least one negative sample are obtained, and the positive sample and the negative sample are both vehicle images including a target frame, wherein the positive sample is a real image and the The partial image in the target frame of the positive sample is a partial image of the vehicle damage, wherein the at least one negative sample includes a non-real image obtained by replacing the partial image in the target frame of the real image with another partial image.
  • the discriminant model discriminates the authenticity of the input image and the semantics in the image target frame, and outputs the probability score of the image that can be used as the image of the car damage.
  • the real image is an unprocessed image obtained directly by shooting.
  • a real vehicle image with labeled car damage when input, its output value should be close to 1, that is, the probability that the image can be used as a labeled vehicle damage image is close to 100%, and when the input is processed
  • the output value of a non-real image should be close to 0, that is, the probability that the image can be used as a labeled car damage image is close to 0.
  • the output value of the foregoing discriminant model is only exemplary, and it is not limited to the probability of taking a value between 0 and 1, but can be set according to the requirements of a specific scenario.
  • the output value of the discriminant model can be the sum of several probabilities, and so on.
  • the discrimination model is used to detect whether a vehicle damage image with a target frame (that is, a vehicle image with a vehicle damage) is true, and also used to detect whether the target frame is a vehicle damage image.
  • a vehicle damage image with a target frame that is, a vehicle image with a vehicle damage
  • positive and negative samples are obtained for training the discriminant model. Therefore, the positive sample is a real image of the accident vehicle, and a target frame is marked on the real image, and the target frame marks the vehicle damage.
  • the at least one negative sample includes a first negative sample, and the first negative sample is a non-real image obtained by replacing a partial image in a target frame of a real image with another partial image.
  • a first negative sample is generated by pasting other car damage images into the target frame.
  • the first negative sample includes at least one of the following characteristics: inconsistent components inside and outside the target frame, inconsistent models inside and outside the target frame, incoherent colors inside and outside the target frame, and inconsistent textures inside and outside the target frame.
  • step S204 the discriminant model is trained using the at least one positive sample and / or at least one negative sample, so that the loss function of the discriminant model after training is reduced compared to before training, wherein the The loss function is related to discriminating the authenticity of each of the at least one positive sample and / or the at least one negative sample.
  • the loss function L D (x, ⁇ ) of the discriminant model can be shown in the following formula (1):
  • ⁇ D represents the parameters of the discriminant model, Indicates a positive sample, Represents a negative sample, and the sum of i and j is m
  • the prediction formula corresponding to the discriminant value of the model. It can be seen from the loss function that the larger the discrimination value for positive samples, the smaller the loss function, the smaller the discrimination value for negative samples, and the smaller the loss function, that is, the loss function reflects the model's discrimination between positive and negative samples. Accuracy.
  • the ⁇ can be adjusted by, for example, the gradient descent method, so that the value of the loss function is reduced, and the model is more accurate.
  • the discrimination model according to the embodiment of the present specification is used to discriminate the authenticity of an image and whether a vehicle damage is included in an image target frame. Therefore, the loss function used to train the discriminant model is also determined based on the purpose of the discriminant model. Therefore, the loss function includes a loss term related to the authenticity of the discrimination sample and a loss term related to the semantic content in the discrimination target frame.
  • the authenticity of the sample may include: whether the components inside and outside the target frame are consistent, whether the models inside and outside the target frame are consistent, whether the colors inside and outside the target frame are coherent, and whether the textures inside and outside the target frame are consistent.
  • the determination of color continuity may also include the determination of continuity of brightness and contrast. Therefore, based on the specific discrimination content, the specific form of the loss function may include multiple forms.
  • the at least one negative sample further includes a second negative sample, and the second negative sample is a real image in the target frame that does not include a local image of the vehicle damage, and the loss
  • the function is also related to the following one of determining each of the at least one positive sample and / or the at least one negative sample: whether the sample target frame contains a local image of the vehicle damage.
  • the loss function represented by the above formula (1) can be adjusted according to the model configuration.
  • the vehicle damage image discrimination model includes a vehicle damage recognition model, and the vehicle damage recognition model is used to determine whether a sample target frame contains a vehicle damage local image.
  • the output value of the discriminant model is obtained based on the discrimination of the authenticity of the image and the discriminant value of the vehicle damage recognition model. Therefore, the existing semantic recognition models (such as imagenet-based various existing semantic recognition models, etc.) can be used for vehicle damage recognition without the need for additional training steps.
  • FIG. 3 shows a flowchart of a method for training the foregoing generation model according to an embodiment of the present specification.
  • the generating model includes a convolutional neural network, and the method is to train one training in the generating model, including:
  • step S302 at least one vehicle image is acquired, and the vehicle image is a real image of the vehicle;
  • step S304 obtaining at least one intermediate image based on the at least one vehicle image, the intermediate image is obtained by marking a target frame on the corresponding vehicle image and pruning a partial image within the target frame;
  • step S306 the generation model is trained using the at least one intermediate image, so that the loss function of the generation model after training is reduced compared to before training, where the loss function is based on
  • the discriminative model trained by the method shows discriminative acquisition of at least one output image, wherein the at least one output image is an image generated by filling the target frame of the at least one intermediate image by the generation model.
  • step S302 at least one vehicle image is acquired, and the vehicle image is a real image of the vehicle.
  • the vehicle image may be a photo of an accident vehicle or a photo of a non-damaged vehicle. That is, the vehicle image may or may not include a partial image of the vehicle damage.
  • step S304 at least one intermediate image is obtained based on the at least one vehicle image, and the intermediate image is obtained by marking a target frame on the corresponding vehicle image and pruning a partial image within the target frame.
  • the intermediate image is obtained by ripping out the image in a region from the vehicle image.
  • the position of the target frame can be randomly determined.
  • the vehicle image includes a local image of the vehicle damage.
  • the position of the target frame may be determined as the position of the local image of the damage in the figure.
  • the vehicle image includes a plurality of local images of vehicle damage.
  • the position of the target frame may be randomly determined as one of the positions of the plurality of local images of vehicle damage.
  • multiple locations of the vehicle that are prone to (high probability) vehicle damage can be determined by using multiple vehicle damage samples, and the positions of the target frames are randomly determined among the multiple locations.
  • the original image in the target frame of the vehicle image can be cut out through a mask.
  • the mask is a matrix of the same size as the image, where the matrix values at the positions corresponding to the target frame are set to 0, the matrix values at other positions are set to 1, and the mask is spotted with the vehicle image Product operation. Therefore, the pixels in the target frame can be erased, and the pixels outside the target frame are retained, so that the original image in the target frame is cut out.
  • step S306 the generation model is trained using the at least one intermediate image, so that the loss function of the generation model after training is reduced compared to before training, where the loss function is based on
  • the discriminative model trained by the method shows discriminative acquisition of at least one output image, wherein the at least one output image is an image generated by filling the target frame of the at least one intermediate image by the generation model.
  • the loss function L G (z, ⁇ G ) of the generated model can be shown in the following formula (2):
  • ⁇ G is the parameter of the generated model
  • z i is the input data of the generated model, that is, the intermediate image
  • G (z i , ⁇ G ) is the output of the generated model, that is, the target is filled in the intermediate image.
  • the output image obtained by the image in the frame, D (G (z i , ⁇ G )) is the discriminant value output from the input and output images of the discriminant model trained by the method shown in FIG. 2.
  • the training of the generative model may be ended, and the trained generative model is an image filling model.
  • the output image generated by the training model trained as above can be used as an image of a real accident vehicle with false and real.
  • FIG. 4 shows a flowchart of a computer-implemented method for generating a car damage image according to an embodiment of the present specification.
  • step S402 a first image is obtained, where the first image is a real image of a vehicle;
  • step S404 a second image is obtained based on the first image, and the second image is obtained by labeling a target frame on the first image and pruning a partial image within the target frame;
  • step S406 the second image is input to an image filling model to obtain a car damage image with a target frame from an output of the image filling model, wherein the image filling model passes the target of the second image
  • the frame is filled with a partial image of the vehicle damage and the vehicle damage image with the target frame is output.
  • steps S402 and S404 in this method reference may be made to steps S302 and S304 in FIG. 3, and details are not described herein again.
  • step S406 the second image is input to an image filling model to obtain a car damage image with a target frame from an output of the image filling model, wherein the image filling model passes the target of the second image
  • the frame is filled with a partial image of the vehicle damage and the vehicle damage image with the target frame is output.
  • the image filling model trained in the method shown in FIG. 3 the generated image is close to a real vehicle image, and the generated image includes a target frame, and the target frame is labeled with a local image of the vehicle damage.
  • the image generated by the image filling model can be used as a labeled training sample for training a vehicle damage recognition model, and the vehicle damage recognition model is used to identify a vehicle damage based on a vehicle damage image.
  • massive labeling data can be generated directly by generating models, and samples of various situations can be exhausted, such as vehicle models, lighting conditions, old and new, shooting angles, etc., which can improve the accuracy of vehicle damage identification models. .
  • FIG. 5 illustrates an apparatus 500 for training a discriminant model according to an embodiment of the present specification.
  • the discrimination model is a classification model including a convolutional neural network, and the device includes:
  • the obtaining unit 51 is configured to obtain at least one positive sample and / or at least one negative sample, and the positive sample and the negative sample are both vehicle images including a target frame, where the positive samples are real images and all The local image in the target frame of the positive sample is a vehicle damage local image, wherein the at least one negative sample includes a non-real image obtained by replacing the local image in the target frame of the real image with another local image; and
  • the training unit 52 is configured to use the at least one positive sample and / or at least one negative sample to train the discriminant model, so that a loss function of the discriminant model after training is reduced compared to before training, where The loss function is related to discriminating the authenticity of each of the at least one positive sample and / or the at least one negative sample.
  • the loss function includes a loss function related to discriminating at least one of the following: whether the components inside and outside the target frame are consistent, whether the vehicle models inside and outside the target frame are consistent, and the target frame Whether the colors inside and outside are consistent, and whether the texture inside and outside the target frame is consistent.
  • FIG. 6 illustrates an apparatus 600 for training a generation model according to an embodiment of the present specification.
  • the generation model includes a convolutional neural network, and the device includes:
  • the first obtaining unit 61 is configured to obtain at least one vehicle image, where the vehicle image is a real image of the vehicle;
  • the second obtaining unit 62 is configured to obtain at least one intermediate image based on the at least one vehicle image, where the intermediate image is marked with a target frame on the corresponding first image, and a partial image in the target frame is cut out. Obtained;
  • the training unit 63 is configured to train the generating model using the at least one intermediate image, so that a loss function of the generated model after training is reduced compared to that before training, where the loss function is based on
  • the vehicle damage image discriminant model trained by the device for training a discriminant model separately discriminates and obtains at least one output image, wherein the at least one output image is filled in the target frame of the at least one intermediate image by the generation model. Generated images separately.
  • the vehicle image includes or does not include a partial image of the vehicle damage.
  • the vehicle image includes at least one vehicle damage local image
  • the second acquisition unit 62 is further configured to randomly mark the target frame at the at least one vehicle damage local image.
  • the second obtaining unit 62 is further configured to mark a target frame at a first position, wherein the first position is a position randomly determined from a plurality of second positions, and the second position This is a high probability location for vehicle damage that is obtained through statistics.
  • the second obtaining unit 62 is further configured to perform a dot product operation on the vehicle image by using a mask to cut out a partial image within the target frame.
  • FIG. 7 illustrates a computer-implemented apparatus 700 for generating a car damage image according to an embodiment of the present specification, including:
  • the first obtaining unit 71 is configured to obtain a first image, where the first image is a real image of a vehicle;
  • the second obtaining unit 72 is configured to obtain a second image based on the first image, where the second image is obtained by marking a target frame on the first image and pruning a partial image within the target frame; as well as
  • the generating unit 73 is configured to input the second image into an image filling model to obtain a vehicle damage image with a target frame from an output of the image filling model, wherein the image filling model The target frame of the image is filled with a partial image of the vehicle damage and the vehicle damage image with the target frame is output.
  • the device for generating a car damage image further includes a using unit 74 configured to use the car damage image to train a vehicle damage recognition model after generating the car damage image, wherein the The vehicle damage recognition model is used to identify the vehicle damage based on the vehicle damage image.
  • Another aspect of this specification provides a computing device including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, any one of the foregoing methods is implemented.
  • the image generated by the generated model can be used as a training sample with labels for training a vehicle damage recognition model, so that manual labeling is not required, and a large number of labels can be generated directly by generating the model.
  • the data can also be exhaustive of samples of various situations, such as vehicle models, lighting conditions, old and new, shooting angles, etc., so that the accuracy of the vehicle damage recognition model can be improved.
  • RAM random access memory
  • ROM read-only memory
  • electrically programmable ROM electrically erasable programmable ROM
  • registers hard disks, removable disks, CD-ROMs, or in technical fields Any other form of storage medium known in the art.

Landscapes

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

Abstract

本说明书实施例公开了一种基于GAN模型的生成车损图像的方法和装置，所述方法包括：获取第一图像，所述第一图像为车辆的真实图像；基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。

Description

基于GAN网络的车损图像生成方法和装置 技术领域
本说明书实施例涉及图像处理技术领域，更具体地，涉及一种训练车损图像判别模型的方法和装置、一种训练图像填充模型的方法和装置、以及一种生成车损图像的方法和装置。
背景技术
在常规的车险理赔场景中,保险公司需要派出专业的查勘定损人员到事故现场进行现场查勘定损，给出车辆的维修方案和赔偿金额，拍摄现场照片，并将定损照片留档以供核查人员核损核价。由于需要人工查勘定损，保险公司需要投入大量的人力成本，和专业知识的培训成本。从普通用户的体验来说，理赔流程由于等待人工查勘员现场拍照、定损员在维修地点定损、核损人员在后台核损，理赔周期较长。
随着互联网的发展，出现一种理赔方案，其中，通过用户在现场拍摄车损照片，并将所述照片上传至服务器，从而通过算法或人工基于所述车损照片进行定损和理赔。然而，在该方案中，在训练用于基于车损照片进行定损的算法模型时，需要对大量车损照片进行人工标注，消耗大量人力，并且数据难以穷举各种车型、光照条件、新旧程度以及拍摄角度等。
因此，需要一种更有效的获取用作训练样本的车损图像。
发明内容
本说明书实施例旨在提供一种更有效的获取车损图像的方案，以解决现有技术中的不足。
为实现上述目的，本说明书一个方面提供一种计算机执行的生成车损图像的方法，包括：
获取第一图像，所述第一图像为车辆的真实图像；
基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框 的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
在一个实施例中，所述第一图像中包括或不包括车辆损伤局部图像。
在一个实施例中，所述第一图像中包括至少一个车辆损伤局部图像，在所述第一图像上标注目标框包括，在所述至少一个车辆损伤局部图像处随机标注所述目标框。
在一个实施例中，在所述第一图像上标注目标框包括，在第一位置标注目标框，其中，所述第一位置是从多个第二位置中随机确定的位置，所述第二位置为通过统计获取的车辆出现车辆损伤的大概率位置。
在一个实施例中，抠除所述目标框内局部图像包括，通过使用掩膜对所述第一图像进行点积运算，抠除所述目标框内局部图像。
在一个实施例中，所述图像填充模型通过训练GAN模型中的生成模型获取，所述GAN模型中还包括判别模型，其中所述判别模型用于判别所述生成模型的输出图像是否为真实的图像、以及所述输出图像的目标框内是否为车辆损伤局部图像，其中，训练GAN模型中的生成模型包括：
获取多个第三图像，所述第三图像为车辆的真实图像；
基于所述多个第三图像获取多个第四图像，所述第四图像通过在对应的所述第三图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
使用至少所述多个第四图像，基于所述判别模型，训练所述生成模型，作为所述图像填充模型。
在一个实施例中，所述判别模型通过以下方式训练：
获取多个正样本和多个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述多个负样本包括至少一个第一负样本，所述第一负样本通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
使用所述多个正样本和多个负样本训练分类模型，作为所述判别模型。
在一个实施例中，所述第一负样本包括以下至少一项特征：目标框内外的部件不一致、目标框内外的车型不一致、目标框内外的颜色不连贯、以及目标框内外的纹理不连贯。
在一个实施例中，所述多个负样本还包括至少一个第二负样本，所述第二负样本为目标框内不含车辆损伤局部图像的真实图像。
在一个实施例中，所述判别模型中包括语义识别模型，所述语义识别模型用于判别样本目标框内是否包含车辆损伤局部图像。
在一个实施例中，所述方法还包括，在生成所述车损图像之后，将所述车损图像用于训练车辆损伤识别模型，其中，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。
本说明书另一方面提供一种计算机执行的生成车损图像的装置，包括：
第一获取单元，配置为，获取第一图像，所述第一图像为车辆的真实图像；
第二获取单元，配置为，基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
输入单元，配置为，将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
在一个实施例中，所述第一图像中包括或不包括车辆损伤局部图像。
在一个实施例中，所述第一图像中包括至少一个车辆损伤局部图像，所述第二获取单元还配置为，在所述至少一个车辆损伤局部图像处随机标注所述目标框。
在一个实施例中，所述第二获取单元还配置为，在第一位置标注目标框，其中，所述第一位置是从多个第二位置中随机确定的位置，所述第二位置为通过统计获取的车辆出现车辆损伤的大概率位置。
在一个实施例中，所述第二获取单元还配置为，通过使用掩膜对所述第一图像进行点积运算，抠除所述目标框内局部图像。
在一个实施例中，所述图像填充模型通过第一训练装置训练，所述第一训练装置用于训练GAN模型中的生成模型，所述GAN模型中还包括判别模型，其中所述判别模型用于判别所述生成模型的输出图像是否为真实的图像、以及所述输出图像的目标框内是否为车辆损伤局部图像，其中，所述第一训练装置包括：
第三获取单元，配置为，获取多个第三图像，所述第三图像为车辆的真实图像；
第四获取单元，配置为，基于所述多个第三图像获取多个第四图像，所述第四图像 通过在对应的所述第三图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
第一训练单元，配置为，使用至少所述多个第四图像，基于所述判别模型，训练所述生成模型，作为所述图像填充模型。
在一个实施例中，所述判别模型通过第二训练装置训练，所述第二训练装置包括：
第五获取单元，配置为，获取多个正样本和多个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述多个负样本包括至少一个第一负样本，所述第一负样本通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
第二训练单元，配置为，使用所述多个正样本和多个负样本训练分类模型，作为所述判别模型。
在一个实施例中，所述第一负样本包括以下至少一项特征：目标框内外的部件不一致、目标框内外的车型不一致、目标框内外的颜色不连贯、以及目标框内外的纹理不连贯。
在一个实施例中，所述多个负样本还包括至少一个第二负样本，所述第二负样本为目标框内不含车辆损伤局部图像的真实图像。
在一个实施例中，所述判别模型中包括语义识别模型，所述语义识别模型用于判别样本目标框内是否包含车辆损伤局部图像。
在一个实施例中，所述装置还包括使用单元，配置为，在生成所述车损图像之后，将所述车损图像用于训练车辆损伤识别模型，其中，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。
本说明书另一方面提供一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现上述任一项方法。
通过根据本说明书实施例的基于GAN模型的方案，通过生成模型生成的图像可作为带标注的训练样本，用于训练车辆损伤识别模型，从而不需要人工标注，可直接通过生成模型生成海量的标注数据，并且还可以穷尽各种情况的样本，如车型、光照条件、新旧程度、拍摄角度等，从而可以提升车辆损伤识别模型的准确率。
附图说明
通过结合附图描述本说明书实施例，可以使得本说明书实施例更加清楚：
图1示出了根据本说明书实施例的车损图像生成系统100的示意图；
图2示出根据本说明书实施例的一种训练车损图像判别模型的方法的流程图；
图3示出根据本说明书实施例的一种训练图像填充模型的方法的流程图；
图4示出根据本说明书实施例的一种计算机执行的生成车损图像的方法的流程图；
图5示出根据本说明书实施例的一种训练车损图像判别模型的装置500；
图6示出根据本说明书实施例的一种训练图像填充模型的装置600；以及
图7示出根据本说明书实施例的一种计算机执行的生成车损图像的装置700。
具体实施方式
下面将结合附图描述本说明书实施例。
图1示出了根据本说明书实施例的车损图像生成系统100的示意图。如图1所示，系统100包括掩膜11、生成模型12和判别模型13。其中，生成模型12和判别模型13构成了生成对抗模型(GAN)。其中掩膜11用于在实际拍摄的车辆图像上确定目标框的位置并抠除目标框内的局部图像。
在模型训练阶段，首先通过至少一个正样本和/或至少一个负样本训练判别模型13。正样本例如为通过拍摄事故车辆的车损照片所获取的图像，并且在该图像中包括用于标注车辆损伤的目标框。负样本例如为通过将上述图像中的目标框中的局部图像抠除并填充其它车损图像所获取的图像。该判别模型13的损失函数与目标框内外包括的语义内容相关、以及与目标框内外的平滑性等相关。具体是，判别网络13判别图像中目标框内外的车辆部件是否为同一个部件、目标框内的图像是否为车损图像，判别目标框内外的颜色、纹理的连贯性等等。判别网络13对样本的判别越准确，即损失函数越小。因此，通过所述至少一个正样本和/或至少一个负样本训练判别模型13，以使得所述判别模型13的损失函数减小。
在训练好判别模型13之后，可通过判别模型13训练生成模型12。具体是，将至少一个实际拍摄的车辆图像(即第一图像)输入掩膜11，以获取至少一个训练样本(第二 图像)，该训练样本如上所述是通过掩膜11在实际拍摄的车辆图像上确定目标框的位置并抠除目标框内的局部图像所获取的图像。使用该至少一个训练样本训练生成模型12，使得生成模型12的损失函数减小。生成网络12的损失函数基于判别模型13分别对至少一个输出图像的判别获取，其中所述至少一个输出图像为通过所述生成模型在所述至少一个第二图像的目标框内进行填充而分别生成的图像。也就是说，训练生成模型12，使得判别模型13对生成模型12生成的车损图像的判别值增大，即，更像真实图像。
图2示出根据本说明书实施例的一种训练上述判别模型的方法的流程图。所述判别模型为包括卷积神经网络的分类模型。所述方法为训练所述判别模型的一次训练过程，包括：
在步骤S202，获取至少一个正样本和/或至少一个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述至少一个负样本包括通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
在步骤S204，使用所述至少一个正样本和/或至少一个负样本训练所述判别模型，以使得，相比于训练前，训练后的所述判别模型的损失函数减小，其中，所述损失函数与判别所述至少一个正样本和/或至少一个负样本各自的样本真实性相关。
首先，在步骤S202，获取至少一个正样本和/或至少一个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述至少一个负样本包括通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像。
所述判别模型对输入图像的真实性和图像目标框中的语义进行判别，输出的是图像可用作标注车损图像的概率分值。所述真实图像即通过拍摄直接获取的未经处理的图像。例如，对于训练好的判别模型，当输入真实的标注车损的车辆图像时，其输出值应接近1，即该图像可用作的标注车损图像的概率接近100％，而当输入经过处理的非真实图像时，其输出值应接近0，即该图像可用作的标注车损图像的概率接近0。可以理解，上述的判别模型的输出值只是示例性的，其不限于是取值在0到1之间的概率，而是可以根据具体场景的需求自行设定。例如，判别模型的输出值可以是几种概率的和，等等。
在本说明书实施例中，判别模型用于检测带有目标框的车损图像(即带有车辆损伤的车辆图像)是否真实，另外，还用于检测目标框中是否是车辆损伤图像。基于该训练 目的，获取用于训练判别模型的正样本和负样本。因此，正样本为事故车辆的真实图像，并且在该真实图像上标注有目标框，该目标框标出车辆损伤。所述至少一个负样本包括第一负样本，所述第一负样本通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像。例如，通过在真实车辆图像中随机确定目标框位置，抠除该目标框中的原图，使用如图1所示的初始的生成模型填充该目标框，从而生成一个第一负样本，或者可通过对该目标框内贴其它车损图像，从而生成第一负样本。所述第一负样本包括以下至少一项特征：目标框内外的部件不一致、目标框内外的车型不一致、目标框内外的颜色不连贯、以及目标框内外的纹理不连贯。
在步骤S204，使用所述至少一个正样本和/或至少一个负样本训练所述判别模型，以使得，相比于训练前，训练后的所述判别模型的损失函数减小，其中，所述损失函数与判别所述至少一个正样本和/或至少一个负样本各自的样本真实性相关。
在获取了正、负样本之后，可通过将正、负样本代入模型的损失函数，以通过各种优化算法调整模型的参数。该判别模型的损失函数L D(x，θ)可如以下公式(1)所示：
Figure PCTCN2019096526-appb-000001
其中，θ D表示判别模型的参数，
Figure PCTCN2019096526-appb-000002
表示正样本，
Figure PCTCN2019096526-appb-000003
表示负样本，i与j之和为m，
Figure PCTCN2019096526-appb-000004
对应于模型判别值的预测公式。从该损失函数可见，对正样本的判别值越大，损失函数越小，对负样本的判别值越小，损失函数越小，也即，该损失函数体现了模型对正、负样本的判别的准确性。可通过例如梯度下降法调整θ，从而使得损失函数值减小，模型更准确。
如上文所述，根据本说明书实施例的判别模型用于判别图像的真实性、以及图像目标框中是否包括车损。因此，用于训练该判别模型的损失函数也基于该判别模型的目的而确定。因此，损失函数中包括与判别样本真实性相关的损失项、以及与判别目标框中语义内容相关的损失项。其中，对于样本真实性判别，可包括判别：目标框内外的部件是否一致、目标框内外的车型是否一致、目标框内外的颜色是否连贯、以及目标框内外的纹理是否连贯等等。另外，对于颜色连贯性的判别，还可以包括对亮度、对比度的连贯性的判别等等。因此，基于具体判别内容，损失函数的具体形式可包括多种形式。
在一个实施例中，为了强化对目标框内语义内容的判别，上述至少一个负样本还包括第二负样本，第二负样本为目标框内不含车辆损伤局部图像的真实图像，所述损失函数还与判别所述至少一个正样本和/或至少一个负样本各自的以下一项相关：样本目标框 内是否包含车辆损伤局部图像。在该情况中，可根据模型配置，调整上述公式(1)表示的损失函数。
在一个实施例中，所述车损图像判别模型中包括车损识别模型，所述车损识别模型用于判别样本目标框内是否包含车辆损伤局部图像。所述判别模型的输出值基于上述对图像真实性的判别和该车损识别模型的判别值综合得出。从而可使用已有的语义识别模型(如基于imagenet的各种已有语义识别模型等)进行车损识别，而省去了另外训练的步骤。
图3示出根据本说明书实施例的一种训练上述生成模型的方法的流程图。所述生成模型包括卷积神经网络，所述方法为训练所述生成模型中的一次训练，包括：
在步骤S302，获取至少一个车辆图像，所述车辆图像为车辆的真实图像；
在步骤S304，基于所述至少一个车辆图像获取至少一个中间图像，所述中间图像通过在对应的所述车辆图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
在步骤S306，使用所述至少一个中间图像训练所述生成模型，以使得，相比于训练前，训练后的所述生成模型的损失函数减小，其中，所述损失函数基于通过图2所示方法训练的判别模型分别对至少一个输出图像的判别获取，其中所述至少一个输出图像为通过所述生成模型在所述至少一个中间图像的目标框内进行填充而分别生成的图像。
首先，在步骤S302，获取至少一个车辆图像，所述车辆图像为车辆的真实图像。例如，该车辆图像可以是事故车辆的照片，也可以是无损伤车辆的照片。即，该车辆图像中可包括或不包括车辆损伤的局部图像。
在步骤S304，基于所述至少一个车辆图像获取至少一个中间图像，所述中间图像通过在对应的所述车辆图像上标注目标框、并抠除所述目标框内局部图像所获取。
也就是说，通过对车辆图像抠除一块区域内的图像，获取中间图像。首先，确定将要抠除的区域(即目标框)的位置。这里，可随机确定目标框的位置。在一个实施例中，车辆图像中包括车辆损伤局部图像，在该情况中，可将目标框的位置确定为图中损伤局部图像所在的位置。在一个实施例中，车辆图像中包括多个车辆损伤局部图像，在该情况中，可将目标框的位置随机确定为多个车辆损伤局部图像所在位置中的一个处。在一个实施例中，可通过已有多个车损样本确定车辆容易(大概率)出现车辆损伤的多个位置，并在该多个位置中随机确定目标框的位置。
在确定目标框的位置之后，可通过掩膜抠除车辆图像目标框内的原图。例如，所述 掩膜为与图像大小相同的矩阵，其中将与目标框对应的位置的矩阵值设置为0，将其它位置处的矩阵值设置为1，并将该掩膜与车辆图像进行点积运算。从而可以抹去目标框内的像素，保留目标框外的像素，从而抠除了目标框内的原图。
在步骤S306，使用所述至少一个中间图像训练所述生成模型，以使得，相比于训练前，训练后的所述生成模型的损失函数减小，其中，所述损失函数基于通过图2所示方法训练的判别模型分别对至少一个输出图像的判别获取，其中所述至少一个输出图像为通过所述生成模型在所述至少一个中间图像的目标框内进行填充而分别生成的图像。
所述生成模型的损失函数L G(z，θ G)可如下文公式(2)所示：
Figure PCTCN2019096526-appb-000005
如公式(2)所示，θ G为所述生成模型的参数，z i为生成模型输入数据，即中间图像，G(z i，θ G)为生成模型的输出，即对中间图像填充目标框内图像所获取的输出图像，D(G(z i，θ G))为对通过图2所示方法训练的判别模型输入输出图像所输出的判别值。由公式(2)可见，判别模型输出的判别值越大，生成模型的损失函数越小。也就是说，判别值越大，说明生成模型生成的输出图像越符合标准，因而模型的损失越小。从而，通过例如梯度下降法的各种优化算法调整θ G，使得损失函数越小，也即使得判别模型的判别值越大，从而优化生成模型。
当判别模型对由该生成模型生成的输出图像无法判断真假(例如，判别模型的输出值为0.5)时，可结束对该生成模型的训练，训练好的生成模型即图像填充模型。通过如上训练好的生成模型生成的输出图像可以以假乱真，作为真实的事故车辆的图像进行使用。
图4示出根据本说明书实施例的一种计算机执行的生成车损图像的方法的流程图。包括：
在步骤S402，获取第一图像，所述第一图像为车辆的真实图像；
在步骤S404,基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
在步骤S406，将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
该方法中的步骤S402和S404的具体实施可参考图3中的步骤S302和S304，在此不再赘述。
在步骤S406，将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。参考上文所述，通过使用以图3所示方法训练的图像填充模型，生成的图像接近真实的车辆图像，并且该生成图像中包括目标框，目标框中标注的是车辆损伤局部图像。通过该图像填充模型生成的图像可作为带标注的训练样本，用于训练车辆损伤识别模型，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。从而不需要人工标注，可直接通过生成模型生成海量的标注数据，并且还可以穷尽各种情况的样本，如车型、光照条件、新旧程度、拍摄角度等，从而可以提升车辆损伤识别模型的准确率。
图5示出根据本说明书实施例的一种训练判别模型的装置500。所述判别模型为包括卷积神经网络的分类模型，所述装置包括：
获取单元51，配置为，获取至少一个正样本和/或至少一个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述至少一个负样本包括通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
训练单元52，配置为，使用所述至少一个正样本和/或至少一个负样本训练所述判别模型，以使得，相比于训练前，训练后的所述判别模型的损失函数减小，其中，所述损失函数与判别所述至少一个正样本和/或至少一个负样本各自的样本真实性相关。
在一个实施例中，在所述训练判别模型的装置中，所述损失函数包括与判别以下至少一项相关的损失函数：目标框内外的部件是否一致、目标框内外的车型是否一致、目标框内外的颜色是否连贯、以及目标框内外的纹理是否连贯。
图6示出根据本说明书实施例的一种训练生成模型的装置600。所述生成模型包括卷积神经网络，所述装置包括：
第一获取单元61，配置为，获取至少一个车辆图像，所述车辆图像为车辆的真实图像；
第二获取单元62，配置为，基于所述至少一个车辆图像获取至少一个中间图像，所述中间图像通过在对应的所述第一图像上标注目标框、并抠除所述目标框内局部图像所 获取；以及
训练单元63，配置为，使用所述至少一个中间图像训练所述生成模型，以使得，相比于训练前，训练后的所述生成模型的损失函数减小，其中，所述损失函数基于通过上述训练判别模型的装置训练的车损图像判别模型分别对至少一个输出图像的判别获取，其中所述至少一个输出图像为通过所述生成模型在所述至少一个中间图像的目标框内进行填充而分别生成的图像。
在一个实施例中，所述车辆图像中包括或不包括车辆损伤的局部图像。
在一个实施例中，所述车辆图像中包括至少一个车辆损伤局部图像，所述第二获取单元62还配置为，在所述至少一个车辆损伤局部图像处随机标注所述目标框。
在一个实施例中，所述第二获取单元62还配置为，在第一位置标注目标框，其中，所述第一位置是从多个第二位置中随机确定的位置，所述第二位置为通过统计获取的车辆出现车辆损伤的大概率位置。
在一个实施例中，所述第二获取单元62还配置为，通过使用掩膜对所述车辆图像进行点积运算，抠除所述目标框内局部图像。
图7示出根据本说明书实施例的一种计算机执行的生成车损图像的装置700，包括：
第一获取单元71，配置为，获取第一图像，所述第一图像为车辆的真实图像；
第二获取单元72，配置为，基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
生成单元73，配置为，将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
在一个实施例中，所述生成车损图像的装置还包括使用单元74，配置为，在生成所述车损图像之后，将所述车损图像用于训练车辆损伤识别模型，其中，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。
本说明书另一方面提供一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现上述任一项方法。
通过根据本说明书实施例的基于GAN模型的方案，通过生成模型生成的图像可作为带标注的训练样本，用于训练车辆损伤识别模型，从而不需要人工标注，可直接通过生成模型生成海量的标注数据，并且还可以穷尽各种情况的样本，如车型、光照条件、新旧程度、拍摄角度等，从而可以提升车辆损伤识别模型的准确率。
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。
上述对本说明书特定实施例进行了描述。其它实施例在所附权利要求书的范围内。在一些情况下，在权利要求书中记载的动作或步骤可以按照不同于实施例中的顺序来执行并且仍然可以实现期望的结果。另外，在附图中描绘的过程不一定要求示出的特定顺序或者连续顺序才能实现期望的结果。在某些实施方式中，多任务处理和并行处理也是可以的或者可能是有利的。
本领域普通技术人员应该还可以进一步意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执轨道，取决于技术方案的特定应用和设计约束条件。本领域普通技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
结合本文中所公开的实施例描述的方法或算法的步骤可以用硬件、处理器执轨道的软件模块，或者二者的结合来实施。软件模块可以置于随机存储器(RAM)、内存、只读存储器(ROM)、电可编程ROM、电可擦除可编程ROM、寄存器、硬盘、可移动磁盘、CD-ROM、或技术领域内所公知的任意其它形式的存储介质中。
以上所述的具体实施方式，对本发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施方式而已，并不用于限定本发明的保护范围，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (23)

  1. 一种计算机执行的生成车损图像的方法，包括：
    获取第一图像，所述第一图像为车辆的真实图像；
    基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
    将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
  2. 根据权利要求1所述的方法，其中，所述第一图像中包括或不包括车辆损伤局部图像。
  3. 根据权利要求1所述的方法，其中，所述第一图像中包括至少一个车辆损伤局部图像，在所述第一图像上标注目标框包括，在所述至少一个车辆损伤局部图像处随机标注所述目标框。
  4. 根据权利要求1所述的方法，其中，在所述第一图像上标注目标框包括，在第一位置标注目标框，其中，所述第一位置是从多个第二位置中随机确定的位置，所述第二位置为通过统计获取的车辆出现车辆损伤的大概率位置。
  5. 根据权利要求1所述的方法，其中，抠除所述目标框内局部图像包括，通过使用掩膜对所述第一图像进行点积运算，抠除所述目标框内局部图像。
  6. 根据权利要求1所述的方法，其中，所述图像填充模型通过训练GAN模型中的生成模型获取，所述GAN模型中还包括判别模型，其中所述判别模型用于判别所述生成模型的输出图像是否为真实的图像、以及所述输出图像的目标框内是否为车辆损伤局部图像，其中，训练GAN模型中的生成模型包括：
    获取多个第三图像，所述第三图像为车辆的真实图像；
    基于所述多个第三图像获取多个第四图像，所述第四图像通过在对应的所述第三图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
    使用至少所述多个第四图像，基于所述判别模型，训练所述生成模型，作为所述图像填充模型。
  7. 根据权利要求6所述的方法，其中，所述判别模型通过以下方式训练：
    获取多个正样本和多个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述多个负样本包括至少一个第一负样本，所述第一负样本通过将真 实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
    使用所述多个正样本和多个负样本训练分类模型，作为所述判别模型。
  8. 根据权利要求7所述的方法，其中，所述第一负样本包括以下至少一项特征：目标框内外的部件不一致、目标框内外的车型不一致、目标框内外的颜色不连贯、以及目标框内外的纹理不连贯。
  9. 根据权利要求7所述的方法，其中，所述多个负样本还包括至少一个第二负样本，所述第二负样本为目标框内不含车辆损伤局部图像的真实图像。
  10. 根据权利要求7所述的方法，其中，所述判别模型中包括语义识别模型，所述语义识别模型用于判别样本目标框内是否包含车辆损伤局部图像。
  11. 根据权利要求7所述的方法，还包括，在生成所述车损图像之后，将所述车损图像用于训练车辆损伤识别模型，其中，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。
  12. 一种计算机执行的生成车损图像的装置，包括：
    第一获取单元，配置为，获取第一图像，所述第一图像为车辆的真实图像；
    第二获取单元，配置为，基于所述第一图像获取第二图像，所述第二图像通过在所述第一图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
    输入单元，配置为，将所述第二图像输入图像填充模型，以从所述图像填充模型的输出获取带有目标框的车损图像，其中，所述图像填充模型通过对所述第二图像的目标框内填充车辆损伤局部图像而输出所述带有目标框的车损图像。
  13. 根据权利要求12所述的装置，其中，所述第一图像中包括或不包括车辆损伤局部图像。
  14. 根据权利要求12所述的装置，其中，所述第一图像中包括至少一个车辆损伤局部图像，所述第二获取单元还配置为，在所述至少一个车辆损伤局部图像处随机标注所述目标框。
  15. 根据权利要求12所述的装置，其中，所述第二获取单元还配置为，在第一位置标注目标框，其中，所述第一位置是从多个第二位置中随机确定的位置，所述第二位置为通过统计获取的车辆出现车辆损伤的大概率位置。
  16. 根据权利要求12所述的装置，其中，所述第二获取单元还配置为，通过使用掩膜对所述第一图像进行点积运算，抠除所述目标框内局部图像。
  17. 根据权利要求12所述的装置，其中，所述图像填充模型通过第一训练装置训练，所述第一训练装置用于训练GAN模型中的生成模型，所述GAN模型中还包括判别模 型，其中所述判别模型用于判别所述生成模型的输出图像是否为真实的图像、以及所述输出图像的目标框内是否为车辆损伤局部图像，其中，所述第一训练装置包括：
    第三获取单元，配置为，获取多个第三图像，所述第三图像为车辆的真实图像；
    第四获取单元，配置为，基于所述多个第三图像获取多个第四图像，所述第四图像通过在对应的所述第三图像上标注目标框、并抠除所述目标框内局部图像所获取；以及
    第一训练单元，配置为，使用至少所述多个第四图像，基于所述判别模型，训练所述生成模型，作为所述图像填充模型。
  18. 根据权利要求17所述的装置，其中，所述判别模型通过第二训练装置训练，所述第二训练装置包括：
    第五获取单元，配置为，获取多个正样本和多个负样本，所述正样本和所述负样本都是包括目标框的车辆图像，其中，所述正样本为真实图像，并且所述正样本的目标框内的局部图像为车辆损伤局部图像，其中，所述多个负样本包括至少一个第一负样本，所述第一负样本通过将真实图像的目标框中的局部图像替换为其它局部图像所获取的非真实图像；以及
    第二训练单元，配置为，使用所述多个正样本和多个负样本训练分类模型，作为所述判别模型。
  19. 根据权利要求18所述的装置，其中，所述第一负样本包括以下至少一项特征：目标框内外的部件不一致、目标框内外的车型不一致、目标框内外的颜色不连贯、以及目标框内外的纹理不连贯。
  20. 根据权利要求18所述的装置，其中，所述多个负样本还包括至少一个第二负样本，所述第二负样本为目标框内不含车辆损伤局部图像的真实图像。
  21. 根据权利要求18所述的装置，其中，所述判别模型中包括语义识别模型，所述语义识别模型用于判别样本目标框内是否包含车辆损伤局部图像。
  22. 根据权利要求18所述的装置，还包括使用单元，配置为，在生成所述车损图像之后，将所述车损图像用于训练车辆损伤识别模型，其中，所述车辆损伤识别模型用于基于车辆的车损图像识别车辆的损伤。
  23. 一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现权利要求1-11中任一项所述的方法。
PCT/CN2019/096526 2018-09-04 2019-07-18 基于gan网络的车损图像生成方法和装置 WO2020048242A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP19857070.7A EP3848888A4 (en) 2018-09-04 2019-07-18 METHOD AND DEVICE FOR GENERATION OF A VEHICLE DAMAGE IMAGE BASED ON A GAN NETWORK
SG11202100107QA SG11202100107QA (en) 2018-09-04 2019-07-18 Method and apparatus for generating vehicle damage image on the basis of gan network
US17/141,790 US11972599B2 (en) 2018-09-04 2021-01-05 Method and apparatus for generating vehicle damage image on the basis of GAN network

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811027110.1 2018-09-04
CN201811027110.1A CN110569864A (zh) 2018-09-04 2018-09-04 基于gan网络的车损图像生成方法和装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/141,790 Continuation US11972599B2 (en) 2018-09-04 2021-01-05 Method and apparatus for generating vehicle damage image on the basis of GAN network

Publications (1)

Publication Number Publication Date
WO2020048242A1 true WO2020048242A1 (zh) 2020-03-12

Family

ID=68772402

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/096526 WO2020048242A1 (zh) 2018-09-04 2019-07-18 基于gan网络的车损图像生成方法和装置

Country Status (6)

Country Link
US (1) US11972599B2 (zh)
EP (1) EP3848888A4 (zh)
CN (1) CN110569864A (zh)
SG (1) SG11202100107QA (zh)
TW (1) TWI716923B (zh)
WO (1) WO2020048242A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111597902A (zh) * 2020-04-16 2020-08-28 浙江工业大学 机动车违停监测方法
CN111681162A (zh) * 2020-06-09 2020-09-18 创新奇智(成都)科技有限公司 一种缺陷样本生成方法、装置、电子设备及存储介质
CN112950458A (zh) * 2021-03-19 2021-06-11 润联软件系统(深圳)有限公司 基于对抗生成网络的图像印章去除方法、装置及相关设备

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111507945B (zh) * 2020-03-31 2022-08-16 成都数之联科技股份有限公司 一种使用无缺陷图训练深度学习缺陷检测模型的方法
CN111652209B (zh) * 2020-04-30 2024-05-10 平安科技（深圳）有限公司 损伤检测方法、装置、电子设备及介质
US11657373B2 (en) * 2020-08-21 2023-05-23 Accenture Global Solutions Limited System and method for identifying structural asset features and damage
CN112560575B (zh) * 2020-11-09 2023-07-18 北京物资学院 一种红富士苹果外形数据增强装置和方法
CN112668462B (zh) * 2020-12-25 2024-05-07 平安科技（深圳）有限公司 车损检测模型训练、车损检测方法、装置、设备及介质
CN113706315A (zh) * 2021-07-28 2021-11-26 中国银行保险信息技术管理有限公司 车辆定损方法、装置及设备
US12002192B2 (en) 2021-11-16 2024-06-04 Solera Holdings, Llc Transfer of damage markers from images to 3D vehicle models for damage assessment
US11669937B2 (en) * 2022-02-28 2023-06-06 Huajie ZENG Method and system for enhancing image captured by on-board camera, and computing device

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107194398A (zh) * 2017-05-10 2017-09-22 平安科技（深圳）有限公司 车损部位的识别方法及系统
CN107358596A (zh) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置、电子设备及系统
CN107403424A (zh) * 2017-04-11 2017-11-28 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置及电子设备
WO2018053340A1 (en) * 2016-09-15 2018-03-22 Twitter, Inc. Super resolution using a generative adversarial network
CN108229434A (zh) * 2018-02-01 2018-06-29 福州大学 一种车辆识别与细致重构的方法
CN108256521A (zh) * 2017-12-29 2018-07-06 济南中维世纪科技有限公司 用于车身颜色识别的有效区域定位方法

Family Cites Families (98)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS60126413A (ja) * 1983-12-10 1985-07-05 Taiyo Kogyo Kk 埋立地における遮水シ−トの敷設方法
US7031901B2 (en) 1998-05-13 2006-04-18 Abu El Ata Nabil A System and method for improving predictive modeling of an information system
US6397334B1 (en) 1998-12-17 2002-05-28 International Business Machines Corporation Method and system for authenticating objects and object data
US6644973B2 (en) 2000-05-16 2003-11-11 William Oster System for improving reading and speaking
US7093129B1 (en) 2000-06-19 2006-08-15 International Business Machines Corporation Secured encrypted communications in a voice browser
JP3846851B2 (ja) 2001-02-01 2006-11-15 松下電器産業株式会社 画像のマッチング処理方法及びその装置
US7565537B2 (en) 2002-06-10 2009-07-21 Microsoft Corporation Secure key exchange with mutual authentication
US20040196363A1 (en) 2003-04-01 2004-10-07 Gary Diamond Video identification verification system
US7466824B2 (en) 2003-10-09 2008-12-16 Nortel Networks Limited Method and system for encryption of streamed data
US7401012B1 (en) 2005-04-20 2008-07-15 Sun Microsystems, Inc. Method and apparatus for characterizing computer system workloads
US8448226B2 (en) 2005-05-13 2013-05-21 Sarangan Narasimhan Coordinate based computer authentication system and methods
US7536304B2 (en) 2005-05-27 2009-05-19 Porticus, Inc. Method and system for bio-metric voice print authentication
KR101426870B1 (ko) 2007-03-06 2014-09-19 스미토모덴키고교가부시키가이샤 화상 가공 방법, 컴퓨터 프로그램을 기록한 컴퓨터 판독 가능한 기록 매체, 및 화상 검사 방법
US7872584B2 (en) 2007-04-09 2011-01-18 Honeywell International Inc. Analyzing smoke or other emissions with pattern recognition
US8280106B2 (en) 2007-09-29 2012-10-02 Samsung Electronics Co., Ltd. Shadow and highlight detection system and method of the same in surveillance camera and recording medium thereof
US9298979B2 (en) 2008-01-18 2016-03-29 Mitek Systems, Inc. Systems and methods for mobile image capture and content processing of driver's licenses
US8180629B2 (en) 2008-07-10 2012-05-15 Trigent Softward Ltd. Automatic pattern generation in natural language processing
DE102008046254A1 (de) 2008-09-08 2010-03-11 Giesecke & Devrient Gmbh Wertdokumentbearbeitungsvorrichtung und ein Verfahren zur Reduktion von Staub in der Wertdokumentbearbeitungsvorrichtung
KR101556654B1 (ko) 2008-11-28 2015-10-02 삼성전자주식회사 영상 통화 수행 방법 및 장치
US8121400B2 (en) 2009-09-24 2012-02-21 Huper Laboratories Co., Ltd. Method of comparing similarity of 3D visual objects
WO2011137411A1 (en) * 2010-04-30 2011-11-03 Vucomp, Inc. Probability density function estimator
US9253167B2 (en) 2011-04-19 2016-02-02 Apriva, Llc Device and system for facilitating communication and networking within a secure mobile environment
US9082235B2 (en) 2011-07-12 2015-07-14 Microsoft Technology Licensing, Llc Using facial data for device authentication or subject identification
US8966613B2 (en) 2011-09-30 2015-02-24 Microsoft Technology Licensing, Llc Multi-frame depth image information identification
US9514357B2 (en) 2012-01-12 2016-12-06 Kofax, Inc. Systems and methods for mobile image capture and processing
US9066125B2 (en) 2012-02-10 2015-06-23 Advanced Biometric Controls, Llc Secure display
JP6052657B2 (ja) 2012-03-13 2016-12-27 パナソニックＩｐマネジメント株式会社 対象物検証装置、対象物検証プログラム、及び対象物検証方法
US20140309876A1 (en) 2013-04-15 2014-10-16 Flextronics Ap, Llc Universal vehicle voice command system
US8705836B2 (en) 2012-08-06 2014-04-22 A2iA S.A. Systems and methods for recognizing information in objects using a mobile device
US9582843B2 (en) 2012-08-20 2017-02-28 Tautachrome, Inc. Authentication and validation of smartphone imagery
US9036943B1 (en) 2013-03-14 2015-05-19 Amazon Technologies, Inc. Cloud-based image improvement
US9147127B2 (en) 2013-03-15 2015-09-29 Facebook, Inc. Verification of user photo IDs
US10475014B1 (en) 2013-03-15 2019-11-12 Amazon Technologies, Inc. Payment device security
US9723251B2 (en) 2013-04-23 2017-08-01 Jaacob I. SLOTKY Technique for image acquisition and management
FR3007172B1 (fr) * 2013-06-12 2020-12-18 Renault Sas Procede et systeme d'identification d'un degat cause a un vehicule
US10319035B2 (en) 2013-10-11 2019-06-11 Ccc Information Services Image capturing and automatic labeling system
US9202119B2 (en) 2013-10-18 2015-12-01 Daon Holdings Limited Methods and systems for determining user liveness
JP6287047B2 (ja) 2013-10-22 2018-03-07 富士通株式会社 画像処理装置、画像処理方法および画像処理プログラム
US9607138B1 (en) 2013-12-18 2017-03-28 Amazon Technologies, Inc. User authentication and verification through video analysis
US10320807B2 (en) 2014-02-25 2019-06-11 Sal Khan Systems and methods relating to the authenticity and verification of photographic identity documents
JP6303090B2 (ja) 2014-03-24 2018-04-04 アルパイン株式会社 画像処理装置および画像処理プログラム
US9646227B2 (en) 2014-07-29 2017-05-09 Microsoft Technology Licensing, Llc Computerized machine learning of interesting video sections
US9258303B1 (en) 2014-08-08 2016-02-09 Cellcrypt Group Limited Method of providing real-time secure communication between end points in a network
CA3186147A1 (en) 2014-08-28 2016-02-28 Kevin Alan Tussy Facial recognition authentication system including path parameters
US9619696B2 (en) 2015-04-15 2017-04-11 Cisco Technology, Inc. Duplicate reduction for face detection
US9794260B2 (en) 2015-08-10 2017-10-17 Yoti Ltd Liveness detection
US20170060867A1 (en) 2015-08-31 2017-03-02 Adfamilies Publicidade, SA Video and image match searching
US10065441B2 (en) 2015-09-01 2018-09-04 Digimarc Corporation Counterfeiting detection using machine readable indicia
US10706266B2 (en) 2015-09-09 2020-07-07 Nec Corporation Guidance acquisition device, guidance acquisition method, and program
GB201517462D0 (en) 2015-10-02 2015-11-18 Tractable Ltd Semi-automatic labelling of datasets
WO2017059576A1 (en) 2015-10-09 2017-04-13 Beijing Sensetime Technology Development Co., Ltd Apparatus and method for pedestrian detection
US20170148102A1 (en) 2015-11-23 2017-05-25 CSI Holdings I LLC Damage assessment and repair based on objective surface data
CN105719188B (zh) 2016-01-22 2017-12-26 平安科技（深圳）有限公司 基于多张图片一致性实现保险理赔反欺诈的方法及服务器
CA2956780A1 (en) * 2016-02-01 2017-08-01 Mitchell International, Inc. Methods for improving automated damage appraisal and devices thereof
US11144889B2 (en) * 2016-04-06 2021-10-12 American International Group, Inc. Automatic assessment of damage and repair costs in vehicles
US10692050B2 (en) 2016-04-06 2020-06-23 American International Group, Inc. Automatic assessment of damage and repair costs in vehicles
US10055882B2 (en) 2016-08-15 2018-08-21 Aquifi, Inc. System and method for three-dimensional scanning and for capturing a bidirectional reflectance distribution function
US10019621B2 (en) * 2016-09-14 2018-07-10 General Motors Llc Identifying a vehicle using a mobile device
GB2554361B8 (en) 2016-09-21 2022-07-06 Emergent Network Intelligence Ltd Automatic image based object damage assessment
EP3336774B1 (en) * 2016-12-13 2020-11-25 Axis AB Method, computer program product and device for training a neural network
US20200005151A1 (en) 2016-12-30 2020-01-02 Nokia Technologies Oy Artificial neural network
US11113800B2 (en) * 2017-01-18 2021-09-07 Nvidia Corporation Filtering image data using a neural network
EP3596449A4 (en) 2017-03-14 2021-01-06 University of Manitoba DETECTION OF STRUCTURAL DEFECTS USING AUTOMATIC LEARNING ALGORITHMS
CN107392218B (zh) * 2017-04-11 2020-08-04 创新先进技术有限公司 一种基于图像的车辆定损方法、装置及电子设备
CN111914692B (zh) * 2017-04-28 2023-07-14 创新先进技术有限公司 车辆定损图像获取方法及装置
CN107610091A (zh) * 2017-07-31 2018-01-19 阿里巴巴集团控股有限公司 车险图像处理方法、装置、服务器及系统
KR102334575B1 (ko) 2017-07-31 2021-12-03 삼성디스플레이 주식회사 무라 검출 장치 및 무라 검출 장치의 검출 방법
CN107437077A (zh) * 2017-08-04 2017-12-05 深圳市唯特视科技有限公司 一种基于生成对抗网络的旋转面部表示学习的方法
US11087292B2 (en) 2017-09-01 2021-08-10 Allstate Insurance Company Analyzing images and videos of damaged vehicles to determine damaged vehicle parts and vehicle asymmetries
CN107578017B (zh) * 2017-09-08 2020-11-17 百度在线网络技术（北京）有限公司 用于生成图像的方法和装置
US20190095877A1 (en) * 2017-09-26 2019-03-28 Panton, Inc. Image recognition system for rental vehicle damage detection and management
CN107909621A (zh) * 2017-11-16 2018-04-13 深圳市唯特视科技有限公司 一种基于双生成对抗网络的医学图像合成方法
US11586875B2 (en) 2017-11-22 2023-02-21 Massachusetts Institute Of Technology Systems and methods for optimization of a data model network architecture for target deployment
CN108109137A (zh) * 2017-12-13 2018-06-01 重庆越畅汽车科技有限公司 车辆部件的机器视觉检测系统及方法
CN109919308B (zh) 2017-12-13 2022-11-11 腾讯科技（深圳）有限公司 一种神经网络模型部署方法、预测方法及相关设备
CN107945140A (zh) * 2017-12-20 2018-04-20 中国科学院深圳先进技术研究院 一种图像修复方法、装置及设备
CN108257100A (zh) * 2018-01-12 2018-07-06 北京奇安信科技有限公司 一种图像修复方法及服务器
US10942767B2 (en) 2018-02-27 2021-03-09 Microsoft Technology Licensing, Llc Deep neural network workload scheduling
US10445638B1 (en) * 2018-02-28 2019-10-15 Amazon Technologies, Inc. Restructuring a multi-dimensional array
US10554738B1 (en) 2018-03-02 2020-02-04 Syncsort Incorporated Methods and apparatus for load balance optimization based on machine learning
CN108416751A (zh) * 2018-03-08 2018-08-17 深圳市唯特视科技有限公司 一种基于深度辅助全分辨率网络的新视点图像合成方法
US10997413B2 (en) 2018-03-23 2021-05-04 NthGen Software Inc. Method and system for obtaining vehicle target views from a video stream
GB2573809B (en) 2018-05-18 2020-11-04 Emotech Ltd Speaker Recognition
US10832065B1 (en) 2018-06-15 2020-11-10 State Farm Mutual Automobile Insurance Company Methods and systems for automatically predicting the repair costs of a damaged vehicle from images
US11030735B2 (en) 2018-08-09 2021-06-08 Exxonmobil Upstream Research Company Subterranean drill bit management system
CN109145903A (zh) * 2018-08-22 2019-01-04 阿里巴巴集团控股有限公司 一种图像处理方法和装置
CN110569837B (zh) * 2018-08-31 2021-06-04 创新先进技术有限公司 优化损伤检测结果的方法及装置
CN110569695B (zh) * 2018-08-31 2021-07-09 创新先进技术有限公司 基于定损图像判定模型的图像处理方法和装置
CN110569700B (zh) * 2018-09-26 2020-11-03 创新先进技术有限公司 优化损伤识别结果的方法及装置
WO2020062392A1 (zh) * 2018-09-28 2020-04-02 上海寒武纪信息科技有限公司 信号处理装置、信号处理方法及相关产品
CN109543824B (zh) * 2018-11-30 2023-05-23 腾讯科技（深圳）有限公司 一种序列模型的处理方法和装置
CN110569701B (zh) * 2018-12-29 2020-08-07 阿里巴巴集团控股有限公司 计算机执行的车辆定损方法及装置
CN109816615B (zh) * 2019-03-06 2022-12-16 腾讯科技（深圳）有限公司 图像修复方法、装置、设备以及存储介质
CN110135437B (zh) * 2019-05-06 2022-04-05 北京百度网讯科技有限公司 用于车辆的定损方法、装置、电子设备和计算机存储介质
US11138292B1 (en) * 2019-05-16 2021-10-05 Facebook, Inc. Circuit and method for computing depthwise convolution
US11222464B2 (en) * 2019-08-22 2022-01-11 The Travelers Indemnity Company Intelligent imagery
US10846716B1 (en) * 2019-12-27 2020-11-24 Capital One Services, Llc System and method for facilitating training of a prediction model to estimate a user vehicle damage tolerance
US11562518B2 (en) * 2020-06-05 2023-01-24 Google Llc Image manipulation by text instruction

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018053340A1 (en) * 2016-09-15 2018-03-22 Twitter, Inc. Super resolution using a generative adversarial network
CN107358596A (zh) * 2017-04-11 2017-11-17 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置、电子设备及系统
CN107403424A (zh) * 2017-04-11 2017-11-28 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置及电子设备
CN107194398A (zh) * 2017-05-10 2017-09-22 平安科技（深圳）有限公司 车损部位的识别方法及系统
CN108256521A (zh) * 2017-12-29 2018-07-06 济南中维世纪科技有限公司 用于车身颜色识别的有效区域定位方法
CN108229434A (zh) * 2018-02-01 2018-06-29 福州大学 一种车辆识别与细致重构的方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3848888A4 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111597902A (zh) * 2020-04-16 2020-08-28 浙江工业大学 机动车违停监测方法
CN111597902B (zh) * 2020-04-16 2023-08-11 浙江工业大学 机动车违停监测方法
CN111681162A (zh) * 2020-06-09 2020-09-18 创新奇智(成都)科技有限公司 一种缺陷样本生成方法、装置、电子设备及存储介质
CN111681162B (zh) * 2020-06-09 2023-09-01 创新奇智(成都)科技有限公司 一种缺陷样本生成方法、装置、电子设备及存储介质
CN112950458A (zh) * 2021-03-19 2021-06-11 润联软件系统(深圳)有限公司 基于对抗生成网络的图像印章去除方法、装置及相关设备

Also Published As

Publication number Publication date
SG11202100107QA (en) 2021-02-25
US20210133501A1 (en) 2021-05-06
TWI716923B (zh) 2021-01-21
EP3848888A4 (en) 2022-06-22
EP3848888A1 (en) 2021-07-14
CN110569864A (zh) 2019-12-13
TW202020803A (zh) 2020-06-01
US11972599B2 (en) 2024-04-30

Similar Documents

Publication Publication Date Title
WO2020048242A1 (zh) 基于gan网络的车损图像生成方法和装置
WO2020048271A1 (zh) 基于gan网络的车损图像增强方法和装置
US11276162B2 (en) Surface defect identification method and apparatus
CN110569878B (zh) 一种基于卷积神经网络的照片背景相似度聚类方法及计算机
US10692050B2 (en) Automatic assessment of damage and repair costs in vehicles
CN110570435B (zh) 用于对车辆损伤图像进行损伤分割的方法及装置
WO2020038138A1 (zh) 样本标注方法及装置、损伤类别的识别方法及装置
CN111639629B (zh) 一种基于图像处理的猪只体重测量方法、装置及存储介质
WO2021217940A1 (zh) 车辆部件识别方法、装置、计算机设备和存储介质
CN108323209B (zh) 信息处理方法、系统、云处理设备以及计算机存储介质
CN110942456B (zh) 篡改图像检测方法、装置、设备及存储介质
US20210390282A1 (en) Training data increment method, electronic apparatus and computer-readable medium
CN114444565B (zh) 一种图像篡改检测方法、终端设备及存储介质
CN111461101A (zh) 工服标志的识别方法、装置、设备及存储介质
CN110766077A (zh) 证据链图像中特写图筛选方法、装置和设备
CN111444911B (zh) 车牌识别模型的训练方法及装置、车牌识别方法及装置
CN114494751A (zh) 证照信息识别方法、装置、设备及介质
CN112287905A (zh) 车辆损伤识别方法、装置、设备及存储介质
CN117115823A (zh) 一种篡改识别方法、装置、计算机设备和存储介质
CN116258908A (zh) 一种基于无人机遥感影像数据的地灾预测评估分类方法
CN116597431A (zh) 商品外包装信息检测方法、装置、电子设备和存储介质
CN114283280A (zh) 一种基于改进卷积神经网络的水面漂浮垃圾识别方法
WO2024160201A1 (zh) 印刷缺陷检测方法、装置、设备及存储介质
US20240296661A1 (en) Method of analyzing a component, method of training a system, apparatus, computer program, and computer-readable storage medium
CN117911940A (zh) 管廊防火门开闭状态识别方法、装置及电子设备

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: 19857070

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2019857070

Country of ref document: EP

Effective date: 20210406