WO2022174397A1 - 图像检测方法、装置及存储介质 - Google Patents

图像检测方法、装置及存储介质 Download PDF

Info

Publication number
WO2022174397A1
WO2022174397A1 PCT/CN2021/076937 CN2021076937W WO2022174397A1 WO 2022174397 A1 WO2022174397 A1 WO 2022174397A1 CN 2021076937 W CN2021076937 W CN 2021076937W WO 2022174397 A1 WO2022174397 A1 WO 2022174397A1
Authority
WO
WIPO (PCT)
Prior art keywords
pixel
image
image area
connection relationship
center point
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/076937
Other languages
English (en)
French (fr)
Inventor
匡海威
蒋卓键
陈晓智
许凌云
陈超
陈雄涛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
SZ DJI Technology Co Ltd
Original Assignee
SZ DJI Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by SZ DJI Technology Co Ltd filed Critical SZ DJI Technology Co Ltd
Priority to PCT/CN2021/076937 priority Critical patent/WO2022174397A1/zh
Publication of WO2022174397A1 publication Critical patent/WO2022174397A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection

Definitions

  • the present application relates to the technical field of image recognition, and in particular, to an image detection method, device and storage medium.
  • Lane line detection is one of the most complex and challenging tasks in autonomous driving. The accuracy, real-time and stability of the lane line detection algorithm will directly affect the performance of the entire assisted driving system. In practical applications, lane line detection is usually performed according to the image captured by the camera on the vehicle. Not only do we need to know which pixels in the image are lane lines, we also need to know which pixels belong to the same instance of lane lines.
  • the network can automatically extract useful feature information from the training data.
  • the existing lane line detection method usually adopts the "segmentation + clustering" method; however, this method takes a long time in the clustering step, and the accuracy is not high.
  • the present application provides an image detection method, device and storage medium.
  • the present application provides an image detection method, the method comprising:
  • the present application provides an image detection method, the method comprising:
  • the connected domain information and the indication information determine the first pixel and the second pixel located on different connected domains
  • the present application provides an image detection device, the device comprising: a memory and a processor;
  • the memory is used to store computer programs
  • the processor is configured to execute the computer program and implement the following steps when executing the computer program:
  • the present application provides an image detection device, the device comprising: a memory and a processor;
  • the memory is used to store computer programs
  • the processor is configured to execute the computer program and implement the following steps when executing the computer program:
  • the connected domain information and the indication information determine the first pixel and the second pixel located on different connected domains
  • the present application provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, the processor realizes the image according to the first aspect Detection method.
  • the present application provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor realizes the image according to the second aspect Detection method.
  • Embodiments of the present application provide an image detection method, device, and storage medium for acquiring an image; detecting the image to obtain indication information that a first pixel in the image points to a second pixel, wherein the first pixel Among the surrounding pixels of , the second pixel and the first pixel have the largest probability of connection relationship. Since the first pixel in the image points to the second pixel with the largest probability of connection relationship after the image is detected, the detection method is relatively fast and has high accuracy; if the first pixel and the second pixel in the image are connected If there are lane lines in the image, it is obvious that the image of the lane lines can be obtained after the connection.
  • This detection method is very useful for specific There are no requirements for application scenarios, and the performance is also stable for diversion scenarios and vehicle lane changing scenarios; therefore, the method of this embodiment provides technical support for fast and accurate detection of lane lines, and the method has no requirements for application scenarios.
  • FIG. 1 is a schematic flowchart of an embodiment of an image detection method of the present application.
  • FIG. 2 is a schematic flowchart of another embodiment of the image detection method of the present application.
  • FIG. 3 is a schematic flowchart of another embodiment of the image detection method of the present application.
  • FIG. 5 is a schematic flowchart of another embodiment of the image detection method of the present application.
  • FIG. 6 is a schematic flowchart of another embodiment of the image detection method of the present application.
  • FIG. 7 is a schematic flowchart of another embodiment of the image detection method of the present application.
  • FIG. 8 is a schematic structural diagram of an embodiment of an image detection apparatus of the present application.
  • Lane line detection is one of the most important tasks in autonomous driving. In practical applications, lane line detection is usually performed according to the image captured by the camera on the vehicle. It is necessary to know which pixels in the image are lane lines and which pixels belong to the same lane line instance.
  • the existing lane line detection method usually adopts the "segmentation + clustering" method; however, this method takes a long time in clustering, and the accuracy is not high.
  • detection methods that remove clusters and output lane lines directly through a convolutional neural network, but this method performs poorly in diversion scenarios and vehicle lane changing scenarios.
  • Embodiments of the present application provide an image detection method, device, and storage medium for acquiring an image; detecting the image to obtain indication information that a first pixel in the image points to a second pixel, wherein the first pixel Among the surrounding pixels of , the second pixel and the first pixel have the largest probability of connection relationship. Since the first pixel in the image points to the second pixel with the largest probability of connection relationship after the image is detected, the detection method is relatively fast and has high accuracy; if the first pixel and the second pixel in the image are connected If there are lane lines in the image, it is obvious that the image of the lane lines can be obtained after the connection.
  • This detection method is very useful for specific There are no requirements for application scenarios, and the performance is also stable for diversion scenarios and vehicle lane changing scenarios; therefore, the method of this embodiment provides technical support for fast and accurate detection of lane lines, and the method has no requirements for application scenarios.
  • FIG. 1 is a schematic flowchart of an embodiment of an image detection method of the present application.
  • the method includes: step S101 and step S102.
  • Step S101 acquire an image.
  • Step S102 Detecting the image to obtain indication information that the first pixel in the image points to the second pixel, wherein the second pixel and the first pixel have the largest value among the surrounding pixels of the first pixel.
  • the first pixel may refer to a pixel that is actively pointed to be connected to itself
  • the second pixel may refer to a pixel pointed to by the first pixel and most likely to be connected to the first pixel among the surrounding pixels.
  • the connection relationship probability may refer to the probability that each pixel in the surrounding pixels of the first pixel and the first pixel can be connected together. The pixel with the greatest probability of connection is the one most likely to be connected with the first pixel.
  • the specific calculation method of the connection relationship probability includes but is not limited to: it can be determined by the distance between pixels; or, determined by the gray value between pixels; or, by the distance between pixels and the gray value between pixels values are determined together; and so on.
  • the first pixel is pixel 1
  • the surrounding pixels of pixel 1 include pixel 2, pixel 3, pixel 4, pixel 5, and pixel 6, wherein the difference between pixel 2, pixel 3, pixel 4, pixel 5, pixel 6 and pixel 1
  • the connection relationship probabilities are 10%, 55%, 96%, 43%, and 30%, respectively.
  • the maximum connection relationship probability is 96%, and the corresponding pixel is pixel 4, so the second pixel is pixel 4.
  • Detect the image to obtain the indication information that the first pixel in the image points to the second pixel The specific implementation methods include but are not limited to: simple image segmentation can be performed to obtain image regions, and each image region can be obtained by detecting in each image region. Indication information that the first pixel points to the second pixel; or, pre-train with an image with a pixel connection label to obtain a trained model, and use the trained model to detect the image to obtain that the first pixel in the image points to the second pixel. pixel indication; etc.
  • the indication information obtained when the image is detected may be the indication information of an image area in the image (that is, the indication information that the first pixel in an image area points to the second pixel), or may be the indication information of multiple image areas in the image Corresponding indication information (that is, indication information that the first pixel of each image area points to the second pixel).
  • the embodiment of the present application acquires an image; detects the image to obtain indication information that a first pixel in the image points to a second pixel, wherein the second pixel and the first pixel in the surrounding pixels of the first pixel A pixel has the largest probability of connection relationship. Since the first pixel in the image points to the second pixel with the largest probability of connection relationship after the image is detected, the detection method is relatively fast and has high accuracy; if the first pixel and the second pixel in the image are connected If there are lane lines in the image, it is obvious that the image of the lane lines can be obtained after the connection.
  • This detection method is very useful for specific There are no requirements for application scenarios, and the performance is also stable for diversion scenarios and vehicle lane changing scenarios; therefore, the method of this embodiment provides technical support for fast and accurate detection of lane lines, and the method has no requirements for application scenarios.
  • the most common application is to connect the first pixel and the second pixel in the image to obtain the connected target image area .
  • the following possible applications are possible:
  • the method further includes: step S103A.
  • Step S103A Connect the first pixel and the second pixel in the image according to the indication information to obtain a connected target image area.
  • the first pixel and the second pixel in the image are directly connected according to the indication information to obtain a connected target image area.
  • pixel 1 points to pixel 2
  • pixel 2 points to pixel 3
  • pixel 3 points to pixel 4
  • pixel 4 points to pixel 5, ...
  • the method further includes: step S103B1, step S103B2 and step S103B3.
  • Step S103B1 Obtain the connected domain information where each pixel in the image is located.
  • Step S103B2 Determine the first pixel and the second pixel located on different connected domains according to the connected domain information and the indication information.
  • Step S103B3 Connect the first pixels located on different connected domains and the connected domains where the second pixels are located to obtain a connected target image area.
  • a connected region is also called a connected region, and may refer to a set (ie, an image region) composed of adjacent pixels with the same pixel value in the image.
  • Obtain the connected domain information of each pixel in the image determine the first pixel and the second pixel on different connected domains according to the connected domain information and the indication information, and locate the first pixel and the second pixel on different connected domains.
  • the connected domains are connected to obtain the connected target image area.
  • the image includes 5 connected domains, namely connected domain 1, connected domain 2, connected domain 3, connected domain 4, and connected domain 5; the number of pixels included in each connected domain may not be equal, for the convenience of description , assuming that the number of pixels included in each connected domain is equal: connected domain 1 includes pixel A1, pixel A2, pixel A3, ..., pixel A10; connected domain 2 includes pixel B1, pixel B2, pixel B3, ..., Pixel B10; connected domain 3 includes pixel C1, pixel C2, pixel C3, ..., pixel C10; connected domain 4 includes pixel D1, pixel D2, pixel D3, ..., pixel D10; connected domain 5 includes pixel E1, pixel E2 , pixel E3, ..., pixel E10; wherein, pixel A2 (located in connected domain 1) points to pixel B2 (located in connected domain 2), pixel B2 (located in connected domain 2) points to pixel D2 (located in connected domain 4), pixel C3 (located in connected domain 3) points to
  • the type of the target image area can also be determined, that is, the method further includes:
  • the type information of pixels may refer to classification information at the pixel level, and pixels belonging to the same class may be classified into one class.
  • pixels belonging to people can be classified into one class
  • pixels belonging to motorcycles can also be classified into one class
  • pixels belonging to linear objects can be classified into one class
  • background pixels can also be classified into one class.
  • the pixels classified into one category can be further instance segmented, that is, further refined classification information.
  • the pixels of different people are classified into different classes; the pixels of linear objects can be further divided: the pixels of solid line objects are classified into one class, the pixels of dotted line objects are classified into one class, and so on.
  • Obtaining the type information of each pixel in an image can be achieved by various means of semantic segmentation.
  • the type information of each pixel can be realized by using a semantic segmentation prediction model with many applications, that is, the acquiring the type information of each pixel in the image may also include: using the trained semantic segmentation prediction model to The image is detected to obtain the type information of each pixel in the image.
  • a trained semantic segmentation prediction model can be obtained through training in advance, and the type information of each pixel in the image can be obtained by detecting the image through the model.
  • the target image area includes a linear image area.
  • the linear image area includes a lane line image area.
  • the lane line image area may be a lane line instance image area.
  • the image area of the lane line instance can be obtained through the pixel connection relationship of the lane line and the type information of each pixel.
  • the segmentation result of the lane line instance can be obtained by directly connecting the associated pixels together, which is simple and fast. And it can detect the lane line instance of the whole image, there is no limit on the number, and there is no limit on the application scene.
  • the detection efficiency can be improved, so the image is detected by using the trained pixel connection relationship prediction model. That is, in step S102, the detection of the image to obtain the indication information that the first pixel in the image points to the second pixel may further include: using the trained pixel connection relationship prediction model to detect the image, to obtain In the image, the first pixel points to the indication information of the second pixel and a plurality of connected domains, and the first pixel and the second pixel in the connected domain are connected.
  • the trained pixel connection relationship prediction model is used to detect the image, in addition to obtaining the indication information that the first pixel in the image points to the second pixel, multiple connected domains can also be obtained.
  • the first pixel and the second pixel of are already connected, so the efficiency of connecting different connected domains into the target image area can be improved.
  • the method may further include: step S103C1, step S103C2 and step S103C3.
  • Step S103C1 According to the indication information that the first pixel points to the second pixel and the plurality of connected domains in the image, determine the indication information of the first pixel at the center point of each connected domain.
  • Step S103C2 According to the indication information of the first pixel at the center point of each connected domain, determine the connected domain where the second pixel pointed to by the first pixel at the center point of each connected domain is located.
  • Step S103C3 Connect the connected domain where the first pixel of the center point of each connected domain is located and the connected domain where the second pixel pointed to by the first pixel of the center point of each connected domain is located to obtain the connected target image area .
  • the connected domain has a center point, which can indicate that the shape of the connected domain is relatively regular. For a relatively regular connected domain, it can be determined that the first pixel of the center point of the connected domain points to the second pixel of the center point of another connected domain. Indication information, Accordingly, the connected domain where the first pixel of the center point of each connected domain is located and the connected domain where the second pixel pointed to by the first pixel is located can be connected to obtain a connected target image area.
  • step S102 the image is detected by using the trained pixel connection relationship prediction model to obtain indication information that the first pixel in the image points to the second pixel Before connecting to multiple connected domains, it may further include: step S104 and step S105, as shown in FIG. 5 .
  • Step S104 Acquire a training image including a pixel connection relationship label.
  • Step S105 The model is trained by using the training image including the label of the pixel connection relationship to obtain a pixel connection relationship prediction model.
  • the training of the model by using the training image including the pixel connection relationship label to obtain the pixel connection relationship prediction model may further include: using the training image including the pixel connection relationship label to train the model, and obtaining a loss function satisfying Preset required pixel connection relationship prediction model.
  • the loss function also known as the cost function, is a function that maps the value of a random event or its related random variables to non-negative real numbers to represent the "risk" or "loss" of the random event.
  • loss functions are often associated with optimization problems as learning criteria, i.e. solving and evaluating models by minimizing the loss function. It can be used for parametric estimation of models in machine learning. Loss functions include, but are not limited to: hinge loss function, cross-entropy loss function, and exponential loss function.
  • the cross entropy loss function is a smooth function, and its essence is the application of cross entropy in information theory in classification problems; according to the definition of cross entropy, minimizing cross entropy is equivalent to minimizing The relative entropy of the observed value and the estimated value, that is, the Kullback-Leibler divergence of the probability distribution of the two, so it is a surrogate loss that provides an unbiased estimate; the cross-entropy loss function is the most widely used surrogate loss.
  • the training image includes a linear image area; at this time, in step S104, the acquiring a training image including a pixel connection relationship label may further include: sub-step S1041 and sub-step S1042.
  • Sub-step S1041 According to the center point of each line image region in each line of the training image, determine the connection relationship indicating label of the pixel at the center point of each line of each line image region to the pixel of the center point of the previous line.
  • Sub-step S1042 According to the connection relationship indication label of the pixels of the center point of each line of each line image area and the width of each line image area, determine the points on the width of each line center point of each line image area on both sides of the center point A label indicating the connection relationship of the pixels of , and then an image of the label indicating the connection relationship of all the pixels in the linear image area is obtained.
  • the linear image area has a center point, and also has a regular width and length, according to these characteristics, you can first determine the connection relationship between the pixels at the center point of each row of the linear image area and the pixels at the center point of the previous row.
  • the width of the linear image area is determined, and the connection relationship indication labels of the pixels on the width of both sides of the center point of each line are determined, so that the connection relationship indication labels of all pixels in the linear image area can be obtained. In this way, training images including pixel connection relationship labels can be obtained simply, efficiently and conveniently.
  • FIG. 7 is a schematic flowchart of another embodiment of the image detection method of the present application. It should be noted that the method of this embodiment has the same content as the above-mentioned image detection method. For details, please refer to the above-mentioned image detection method. The related content will not be repeated here.
  • the method includes: step S201, step S202, step S203 and step S204.
  • Step S201 Obtain the connected domain information where each pixel in the image is located.
  • Step S202 Acquire indication information that a first pixel in the image points to a second pixel, wherein among the surrounding pixels of the first pixel, the second pixel and the first pixel have the largest probability of a connection relationship.
  • Step S203 Determine the first pixel and the second pixel located on different connected domains according to the connected domain information and the indication information.
  • Step S204 Connect the first pixels located on different connected domains and the connected domains where the second pixels are located to obtain a connected target image area.
  • determining the first pixel and the second pixel located on different connected domains according to the connected domain information and the indication information includes: determining each connected domain according to the connected domain information and the indication information The second pixel of another connected domain pointed to by the first pixel of the central point of the : Connect the connected domain where the first pixel of the center point of each connected domain is located and another connected domain where the second pixel pointed to by the first pixel of the center point of each connected domain is located to obtain the connected target image area .
  • the method further includes: acquiring type information of each pixel in the image; and determining the type of the connected target image area according to the type information of each pixel.
  • the acquiring the type information of each pixel in the image includes: using the trained semantic segmentation prediction model to detect the image to obtain the type information of each pixel in the image.
  • the target image area includes a linear image area.
  • the linear image area includes a lane line image area.
  • acquiring the connected domain information where each pixel in the image is located, and acquiring the indication information that the first pixel in the image points to the second pixel includes: detecting the image by using a trained pixel connection relationship prediction model, Obtaining information indicating that the first pixel in the image points to the second pixel and a plurality of connected domains, where the first pixel and the second pixel in the connected domain are connected.
  • the method before detecting the image by using the trained pixel connection relationship prediction model, and obtaining the indication information that the first pixel in the image points to the second pixel and multiple connected domains, the method includes: acquiring a label including a pixel connection relationship The training image is used to train the model with the training image including the pixel connection relationship label, and the pixel connection relationship prediction model is obtained.
  • the training the model by using the training images including the pixel connection relationship label to obtain the pixel connection relationship prediction model includes: using the training image including the pixel connection relationship label to train the model, and obtaining the pixels whose loss function meets the preset requirements Connection relationship prediction model.
  • the training image includes a linear image area
  • the acquiring the training image including the pixel connection relationship label includes: determining each linear image according to the center point of each linear image area in each row of the training image The connection relationship between the pixels at the center point of each line of the region points to the pixels at the center point of the previous line indicating the label; according to the connection relationship between the pixels at the center point of each line in each line image region indicating the label and the width of each line image region, Determine the connection relationship indicating label of the pixels on the width of each line center point of each line image area, and then obtain the image of the connection relationship indicating label of all pixels in the linear image area.
  • the target image area as an example image area of a lane line as an example.
  • the lane line image can be projected onto the bird's-eye view (the front view can also be used directly), and for each lane line instance, find the lane line instance in each line of the image. center point.
  • step 1.2 Repeat step 1.2 to obtain a pixel connection relationship indication label of the center point of each row of a lane line instance.
  • the network structure can be a multi-task fully convolutional neural network composed of a backbone network, a first branch network (ie, a semantic segmentation prediction model) and a second branch network (ie, a pixel connection relationship prediction model).
  • the backbone network can be used to perform feature extraction on the input image.
  • the first branch network can be used for lane line semantic segmentation prediction, process the data after feature extraction, and obtain the type information of each pixel of the image, such as background, solid line, dashed line, etc.
  • the second branch network can be used to predict the pixel connection relationship of the lane line, process the data after feature extraction, and obtain the position of the center point of the lane line in the next row that each pixel in the image points to and belongs to the same instance.
  • the outputs of the first branch network and the second branch network can be constrained by the cross-entropy loss function:
  • L is the final loss function of the network, where ⁇ represents the scale coefficient of the semantic segmentation loss of the first branch network in the overall loss, and ⁇ represents the scale coefficient of the pixel connection loss of the second branch network in the overall loss.
  • the input training image is forwarded to obtain the prediction result, the loss is calculated according to the corresponding label data and the loss function, and the network parameters are updated by back-propagation to obtain the final lane line detection network model.
  • step 4.3 Repeat step 4.3 until the entire image is traversed, and the instance segmentation result after pixel connection is obtained.
  • the embodiment of the present application obtains the result of segmentation of the lane line instance by predicting the pixel connection relationship of the lane line instance.
  • the detection method is simple, time-consuming, and can support obtaining the results of multiple lane line instances at the same time, and there is no limitation on the number of them; Using a bird's-eye view of the input image, you can also get more accurate results on distant lane lines.
  • FIG. 8 is a schematic structural diagram of an embodiment of an image detection apparatus of the present application. It should be noted that the image detection apparatus of this embodiment can perform the operations in the first aspect and the related image detection method. For details, please refer to the above-mentioned related content, which will not be repeated here.
  • the apparatus 100 includes: a memory 1 and a processor 2; the processor 2 and the memory 1 are connected through a bus.
  • the processor 2 may be a microcontroller unit, a central processing unit or a digital signal processor, and so on.
  • the memory 1 may be a Flash chip, a read-only memory, a magnetic disk, an optical disk, a U disk, a mobile hard disk, and the like.
  • the memory 1 is used to store a computer program; the processor 2 is used to execute the computer program and implement the following steps when executing the computer program:
  • the processor when executing the computer program, implements the following steps: acquiring the connected domain information where each pixel in the image is located; A pixel and a second pixel; connect the first pixel located on different connected domains and the connected domain where the second pixel is located to obtain a connected target image area.
  • the processor when executing the computer program, implements the following steps: acquiring type information of each pixel in the image; and determining the type of the connected target image area according to the type information of each pixel.
  • the processor when executing the computer program, implements the following steps: using the trained semantic segmentation prediction model to detect the image to obtain type information of each pixel in the image.
  • the target image area includes a linear image area.
  • the linear image area includes a lane line image area.
  • the processor when executing the computer program, implements the following steps: using the trained pixel connection relationship prediction model to detect the image, to obtain the indication information that the first pixel in the image points to the second pixel and a plurality of connected domains, the first pixel and the second pixel in the connected domain are connected; according to the indication information that the first pixel points to the second pixel in the image and the plurality of connected domains, the center point of each connected domain is determined.
  • the indication information of the first pixel of the The connected domain where the first pixel of the center point of the domain is located is connected with the connected domain where the second pixel pointed to by the first pixel of the center point of each connected domain is located to obtain a connected target image area.
  • the processor when executing the computer program, implements the following steps: acquiring a training image including pixel connection relationship labels; using the training images including pixel connection relationship labels to train a model to obtain a pixel connection relationship prediction model.
  • the processor when executing the computer program, implements the following steps: using training images including pixel connection relationship labels to train the model, to obtain a pixel connection relationship prediction model whose loss function meets preset requirements.
  • the training image includes a linear image area; when the processor executes the computer program, the processor implements the following steps: determining each line according to the center point of each linear image area in each row of the training image The pixel at the center point of each line of the linear image area points to the connection relationship indication label of the pixel at the center point of the previous line; according to the connection relationship indication label of the pixel at the center point of each line in each line image area, the label and the The width of each line image area is determined, and the connection relationship indicating label of the pixels on the width of each line center point is determined, and then the image of the connection relationship indicating label of all pixels in the linear image area is obtained.
  • the present application further provides another image detection apparatus.
  • the image detection apparatus of this embodiment can perform the operations in the second aspect and the related image detection method.
  • the related content please refer to the above related content , will not be repeated here.
  • the device includes: a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the following steps when executing the computer program:
  • the processor when executing the computer program, implements the following steps: according to the connected domain information and the indication information, determining the first pixel of another connected domain pointed to by the first pixel of the center point of each connected domain. two pixels; the connecting the first pixels located on different connected domains and the connected domains where the second pixels are located to obtain the connected target image area, including: connecting the first pixel of the center point of each connected domain where the first pixel is located.
  • the connected domain is connected with another connected domain where the second pixel pointed to by the first pixel of the center point of each connected domain is located, and the connected target image area is obtained.
  • the processor when executing the computer program, implements the following steps: acquiring type information of each pixel in the image; and determining the type of the connected target image area according to the type information of each pixel.
  • the processor when executing the computer program, implements the following steps: using the trained semantic segmentation prediction model to detect the image to obtain type information of each pixel in the image.
  • the target image area includes a linear image area.
  • the linear image area includes a lane line image area.
  • the processor when executing the computer program, implements the following steps: using the trained pixel connection relationship prediction model to detect the image, to obtain the indication information that the first pixel in the image points to the second pixel and A plurality of connected domains in which the first pixel and the second pixel are connected.
  • the processor when executing the computer program, implements the following steps: acquiring a training image including pixel connection relationship labels; using the training images including pixel connection relationship labels to train a model to obtain a pixel connection relationship prediction model.
  • the processor when executing the computer program, implements the following steps: using training images including pixel connection relationship labels to train the model, to obtain a pixel connection relationship prediction model whose loss function meets preset requirements.
  • the training image includes a linear image area; when the processor executes the computer program, the processor implements the following steps: determining each line according to the center point of each linear image area in each row of the training image The pixel at the center point of each line of the linear image area points to the connection relationship indication label of the pixel at the center point of the previous line; according to the connection relationship indication label of the pixel at the center point of each line in each line image area, the label and the The width of each line image area is determined, and the connection relationship indicating label of the pixels on the width of each line center point is determined, and then the image of the connection relationship indicating label of all pixels in the linear image area is obtained.
  • the present application also provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, causes the processor to implement the above-mentioned first aspect and any related items.
  • the described image detection method is also provided.
  • the computer-readable storage medium may be an internal storage unit of the above-mentioned image detection apparatus, such as a hard disk or a memory.
  • the computer-readable storage medium may also be an external storage device, such as an equipped plug-in hard disk, smart memory card, secure digital card, flash memory card, and the like.
  • the present application also provides another computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor enables the processor to implement any one of the second aspect and related items above The described image detection method.
  • the computer-readable storage medium may be an internal storage unit of the above-mentioned image detection apparatus, such as a hard disk or a memory.
  • the computer-readable storage medium may also be an external storage device, such as an equipped plug-in hard disk, smart memory card, secure digital card, flash memory card, and the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种图像检测方法、装置及存储介质,所述方法包括:获取图像(S101);对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率(S102)。

Description

图像检测方法、装置及存储介质 技术领域
本申请涉及图像识别技术领域,尤其涉及一种图像检测方法、装置及存储介质。
背景技术
车道线检测是自动驾驶诸多复杂且具有挑战性的任务中,最受重视的任务之一。车道线检测算法的准确性、实时性、稳定性会直接影响整个辅助驾驶系统的性能。在实际应用中,通常是根据车辆上摄像头拍摄到的图像进行车道线检测。不仅需要知道图像中哪些像素是车道线,还需要知道哪些像素属于同一个车道线实例。
近年来,深度学习技术得到快速发展,通过对图像进行标注,将标注好的图像作为训练集输入到深度学习中的卷积神经网络,网络能够自动从训练数据中提取出有用的特征信息。现有的车道线检测方法通常采用“分割+聚类”的方法;但是该方法在聚类这步耗时会比较长,且准确度不高。也有的车道线检测方法去除聚类的步骤,直接通过卷积神经网络输出固定数量的车道线,但是该方法对于分流场景、车辆变道场景表现都很差。
发明内容
基于此,本申请提供一种图像检测方法、装置及存储介质。
第一方面,本申请提供一种图像检测方法,所述方法包括:
获取图像;
对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
第二方面,本申请提供一种图像检测方法,所述方法包括:
获取图像中每个像素所在的连通域信息;
获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率;
根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
第三方面,本申请提供一种图像检测装置,所述装置包括:存储器和处理器;
所述存储器用于存储计算机程序;
所述处理器用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
获取图像;
对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
第四方面,本申请提供一种图像检测装置,所述装置包括:存储器和处理器;
所述存储器用于存储计算机程序;
所述处理器用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
获取图像中每个像素所在的连通域信息;
获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率;
根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
第五方面,本申请提供一种计算机可读存储介质,所述计算机可读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使所述处理器实现如第一方面所述的图像检测方法。
第六方面,本申请提供一种计算机可读存储介质,所述计算机可读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使所述处理器实现如第二方面所述的图像检测方法。
本申请实施例提供了一种图像检测方法、装置及存储介质,获取图像;对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。由于对图像进行检测后得到图像中第一像素指向具有最大的连接关系概率的第二像素的指示信息,该检测方法比较快速、准确度高;如果将图像中的第一像素和第二像素连接起来,就可以得到具有连接关系的图像区域,例如可以将具有连接关系的像素快速、准确地连接起来,如果图像中有车道线,很显然连接后能够得到车道线图像,该检测方法对具体的应用场景没有要求,对于分流场景、车辆变道场景表现也很稳定;因此,本实施例的方法为快速、准确检测车道线提供技术支持,且该方法对应用场景没有要求。
应当理解的是,以上的一般描述和后文的细节描述仅是示例性和解释性的,并不能限制本申请。
附图说明
为了更清楚地说明本申请实施例技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1是本申请图像检测方法一实施例的流程示意图;
图2是本申请图像检测方法另一实施例的流程示意图;
图3是本申请图像检测方法又一实施例的流程示意图;
图4是本申请图像检测方法又一实施例的流程示意图;
图5是本申请图像检测方法又一实施例的流程示意图;
图6是本申请图像检测方法又一实施例的流程示意图;
图7是本申请图像检测方法又一实施例的流程示意图;
图8是本申请图像检测装置一实施例的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
附图中所示的流程图仅是示例说明,不是必须包括所有的内容和操作/步骤,也不是必须按所描述的顺序执行。例如,有的操作/步骤还可以分解、组合或部分合并,因此实际执行的顺序有可能根据实际情况改变。
车道线检测是自动驾驶中最受重视的任务之一。在实际应用中,通常是根据车辆上摄像头拍摄到的图像进行车道线检测,需要知道图像中哪些像素是车道线,哪些像素属于同一个车道线实例。现有的车道线检测方法通常采用“分割+聚类”的方法;但是该方法在聚类耗时较长,准确度不高。也有检测方法去除聚类,直接通过卷积神经网络输出车道线,但是该方法对于分流场景、车辆变道场景表现很差。
本申请实施例提供了一种图像检测方法、装置及存储介质,获取图像;对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。由于对图像进行检测后得到图像中第一像素指向具有最大的连接关系概率的第二像素的指示信息,该检测方法比较快速、准确度高;如果将图像中的第一像素和第二像素连接起来,就可以得到具有连接关系的图像区域,例如可以将具有连接关系的像素快速、准确地连接起来,如果图像中有车道线,很显然连接后能够得到车道线图像,该检测方法对具体的应用场景没有要求,对于分流场景、车辆变道场景表现也很稳定;因此,本实施例的方法为快速、准确检测车道线提供技术支持,且该方法对应用场景没有要求。
下面结合附图,对本申请的一些实施方式作详细说明。在不冲突的情况下,下述的实施例及实施例中的特征可以相互组合。
参见图1,图1是本申请图像检测方法一实施例的流程示意图,所述方法包括:步骤S101和步骤S102。
步骤S101:获取图像。
步骤S102:对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
本实施例中,第一像素可以是指主动指向与自身连接的像素,第二像素可以是指第一像素指向的、周围像素中最有可能与第一像素连接在一起的像素。连接关系概率可以是指所述第一像素的周围像素中每个像素与第一像素可以连接在一起的概率。具有最大的连接关系概率的像素是最有可能与第一像素连接在一起的像素。连接关系概率的具体计算方式包括但不限于:可以通过像素之间的距离来确定;或者,通过像素之间的灰度值来确定;或者,通过像素之间的距离和像素之间的灰度值一起来确定;等等。
例如:第一像素为像素1,像素1的周围像素包括像素2、像素3、像素4、像素5、像素6,其中,像素2、像素3、像素4、像素5、像素6与像素1的连接关系概率分别是10%、55%、96%、43%、30%,很显然最大的连接关系概率是96%,对应的像素是像素4,因此第二像素为像素4。
对图像进行检测,得到图像中第一像素指向第二像素的指示信息,具体实现方式包括但不限于:可以进行简单的图像分割得到图像区域,在每个图像区域中进行检测得到每个图像区域中第一像素指向第二像素的指示信息;或者,预先用具有像素连接关系标签的图像进行训练,得到训练好的模型,利用训练好的模型对图像进行检测得到图像中第一像素指向第二像素的指示信息;等等。
对图像进行检测,得到图像中第一像素指向第二像素的指示信息,在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率,即在第一像素的周围像素中第二像素与第一像素具有连接关系的概率是最大的。在对图像进行检测时得到的指示信息可以是图像中某个图像区域的指示信息(即某个图像区域中的第一像素指向第二像素的指示信息),也可以是图 像中多个图像区域对应的指示信息(即每个图像区域的第一像素指向第二像素的指示信息)。
本申请实施例获取图像;对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。由于对图像进行检测后得到图像中第一像素指向具有最大的连接关系概率的第二像素的指示信息,该检测方法比较快速、准确度高;如果将图像中的第一像素和第二像素连接起来,就可以得到具有连接关系的图像区域,例如可以将具有连接关系的像素快速、准确地连接起来,如果图像中有车道线,很显然连接后能够得到车道线图像,该检测方法对具体的应用场景没有要求,对于分流场景、车辆变道场景表现也很稳定;因此,本实施例的方法为快速、准确检测车道线提供技术支持,且该方法对应用场景没有要求。
在实际应用中,得到所述图像中第一像素指向第二像素的指示信息后,最常见的应用是将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。具体可以有以下可能的应用方式:
结合参见图2,在一实施例中,所述方法还包括:步骤S103A。
步骤S103A:根据所述指示信息将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。
本实施例根据所述指示信息直接将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。
例如:像素1指向像素2,像素2指向像素3,像素3指向像素4,像素4指向像素5,……,将像素1、像素2、像素3、像素4、像素5、……,顺序连接,得到连接后的目标图像区域。
参见图3,在另一实施例中,所述方法还包括:步骤S103B1、步骤S103B2以及步骤S103B3。
步骤S103B1:获取图像中每个像素所在的连通域信息。
步骤S103B2:根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素。
步骤S103B3:将位于不同连通域上的第一像素和第二像素所在的连通域 进行连接,得到连接后的目标图像区域。
本实施例中,连通域也称为连通区域,可以是指图像中具有相同像素值且位置相邻的像素点组成的集合(即图像区域)。获取图像中每个像素所在的连通域信息,根据连通域信息和指示信息,确定位于不同连通域上的第一像素和第二像素,将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
例如:图像中包括5个连通域,分别为连通域1、连通域2、连通域3、连通域4、连通域5;每个连通域中包括的像素的个数可以不相等,为了方便说明,假设每个连通域中包括的像素的个数相等:连通域1包括像素A1、像素A2、像素A3、……、像素A10;连通域2包括像素B1、像素B2、像素B3、……、像素B10;连通域3包括像素C1、像素C2、像素C3、……、像素C10;连通域4包括像素D1、像素D2、像素D3、……、像素D10;连通域5包括像素E1、像素E2、像素E3、……、像素E10;其中,像素A2(位于连通域1)指向像素B2(位于连通域2),像素B2(位于连通域2)指向像素D2(位于连通域4),像素C3(位于连通域3)指向像素E3(位于连通域5);因此,可以将连通域1、连通域2、连通域4连接在一起,组成目标图像区域1;可以将连通域3和连通域5连接在一起,组成目标图像区域2。
在一实施例中,还可以确定目标图像区域的类型,即所述方法还包括:
(1)获取图像中每个像素的类型信息。
(2)根据每个像素的类型信息确定连接后的目标图像区域的类型。
像素的类型信息可以是指像素级别上的分类信息,属于同一类的像素可以被归为一类。例如:属于人的像素可以分成一类,属于摩托车的像素也可以分成一类,属于线状物体的像素可以分成一类,除此之外还有背景像素也被分为一类。其中归为一类的像素还可以进一步进行实例分割,即进一步细化的分类信息。例如将不同人的像素归为不同的类;线状物体的像素还可以进一步划分:实线线状物体的像素分成一类,虚线线状物体的像素分成一类,等等。
获取图像中每个像素的类型信息可以通过语义分割的各种方式实现。
其中,每个像素的类型信息可以采用应用较多的语义分割预测模型来实现,即,所述获取图像中每个像素的类型信息,还可以包括:利用已训练的语义分 割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
本实施例可以预先通过训练得到已训练的语义分割预测模型,通过该模型对图像进行检测即可得到图像中每个像素的类型信息。
在一实施例中,所述目标图像区域包括线状图像区域。
其中,所述线状图像区域包括车道线图像区域。车道线图像区域可以是车道线实例图像区域。
本实施例通过车道线像素连接关系和每个像素的类型信息可以得到车道线实例图像区域,无需复杂后处理过程,直接将关联的像素连接在一起即可得到车道线实例分割结果,简单快速,并且可以将整张图像的车道线实例检测出来,没有数量的限制,也没有应用场景的限制。
在一实施例中,由于已训练的像素连接关系预测模型可以反复多次利用,能够提高检测效率,因此利用已训练的像素连接关系预测模型对所述图像进行检测。即步骤S102,所述对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,还可以包括:利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接。
本实施例中利用已训练的像素连接关系预测模型对所述图像进行检测,除了可以得到所述图像中第一像素指向第二像素的指示信息外,还可以得到多个连通域,连通域中的第一像素和第二像素已连接,因此还可以提高后续进一步将不同的连通域连接为目标图像区域的效率。
其中,参见图4,所述方法还可以包括:步骤S103C1、步骤S103C2以及步骤S103C3。
步骤S103C1:根据所述图像中第一像素指向第二像素的指示信息和多个连通域,确定每个连通域的中心点的第一像素的指示信息。
步骤S103C2:根据每个连通域的中心点的第一像素的指示信息,确定每个连通域的中心点的第一像素指向的第二像素所在的连通域。
步骤S103C3:将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的连通域进行连接,得到连接后的目标图像区域。
连通域具有中心点,可以说明该连通域的形状比较规则,对于比较规则的连通域,可以确定连通域的中心点的第一像素指向另一个连通域的中心点的第二像素的指示信息,据此可以将每个连通域的中心点的第一像素所在的连通域和该第一像素指向的第二像素所在的连通域进行连接,得到连接后的目标图像区域。
如果预先没有得到已训练的像素连接关系预测模型,则步骤S102,所述利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域之前,还可以包括:步骤S104和步骤S105,如图5所示。
步骤S104:获取包括像素连接关系标签的训练图像。
步骤S105:利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
其中,步骤S105,所述利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型,还可以包括:利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
损失函数(loss function)也称为代价函数(cost function),是将随机事件或其有关随机变量的取值映射为非负实数以表示该随机事件的“风险”或“损失”的函数。在应用中,损失函数通常作为学习准则与优化问题相联系,即通过最小化损失函数求解和评估模型。在机器学习中可以被用于模型的参数估计(parametric estimation)。损失函数包括但不限于:铰链损失函数、交叉熵损失函数、指数损失函数。其中交叉熵损失函数是一个平滑函数,其本质是信息理论(information theory)中的交叉熵(cross entropy)在分类问题中的应用;由交叉熵的定义可知,最小化交叉熵等价于最小化观测值和估计值的相对熵(relative entropy),即两者概率分布的Kullback-Leibler散度,因此其是一个提供无偏估计的代理损失;交叉熵损失函数是使用最广泛的代理损失。
在一实施例中,所述训练图像包括线状图像区域;此时步骤S104,所述获取包括像素连接关系标签的训练图像,还可以包括:子步骤S1041和子步骤S1042。
子步骤S1041:根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签。
子步骤S1042:根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
由于线状图像区域具有中心点,也具有规则的宽度和长度,根据这些特点可以先确定线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签,然后根据线状图像区域的宽度,确定每行中心点两边宽度上的点的像素的连接关系指示标签,从而可以得到线状图像区域所有像素的连接关系指示标签。通过这种方式,能够简单、高效、较为方便得到包括像素连接关系标签的训练图像。
参见图7,图7是本申请图像检测方法又一实施例的流程示意图,需要说明的是,本实施例的方法与上述的图像检测方法有相同的内容,具体请参见上述图像检测方法中的相关内容,在此不再赘叙。
所述方法包括:步骤S201、步骤S202、步骤S203以及步骤S204。
步骤S201:获取图像中每个像素所在的连通域信息。
步骤S202:获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
步骤S203:根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素。
步骤S204:将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
其中,所述根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素,包括:根据所述连通域信息和所述指示信息,确定每个连通域的中心点的第一像素指向的另一个连通域的第二像素;所述将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标 图像区域,包括:将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的另一个连通域进行连接,得到连接后的目标图像区域。
其中,所述方法还包括:获取图像中每个像素的类型信息;根据每个像素的类型信息确定连接后的目标图像区域的类型。
其中,所述获取图像中每个像素的类型信息,包括:利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
其中,所述目标图像区域包括线状图像区域。
其中,所述线状图像区域包括车道线图像区域。
其中,所述获取图像中每个像素所在的连通域信息,获取所述图像中第一像素指向第二像素的指示信息,包括:利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接。
其中,所述利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域之前,包括:获取包括像素连接关系标签的训练图像;利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
其中,所述利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型,包括:利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
其中,所述训练图像包括线状图像区域;所述获取包括像素连接关系标签的训练图像,包括:根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
下面以目标图像区域为车道线实例图像区域为例,说明本申请实施例的具体应用。
1、车道线图像区域像素连接关系指示标签的生成:
1.1、在已有车道线实例级别的标注条件下,可以将车道线图像投影到鸟瞰图上(也可以直接使用前视图),对每个车道线实例,找到该车道线实例在图像每一行的中心点。
1.2、计算当前行车道线中心点(y,x)与上一行车道线中心点(y-1,x+b)的横向坐标的相对偏移b,确定当前行中心点的像素连接关系指示标签为b。若当前行车道线为车道线终点,上一行不存在车道线,则当前行中心点的像素连接关系指示标签为e。
1.3、重复1.2步骤,得到一个车道线实例的每行中心点的像素连接关系指示标签。
1.4、按固定车道线宽度w扩展车道线像素连接关系指示标签。如当前行车道线中心点为(y,x),该中心点的像素连接关系指示标签为b,点(y,x+1)的像素连接关系指示标签则是b+1;往中心点两边拓宽到点(y,x-w/2)和(y,x+w/2),其像素连接关系指示标签分别为b-w/2和b+w/2;最终得到一个车道线实例图像区域的所有像素的连接关系指示标签。
1.5、对每个车道线实例图像区域,重复1.2、1.3、1.4步骤,得到全图像的车道线实例图像区域的像素连接关系指示标签。
2、构建用于检测图像的网络结构:
该网络结构可以由主干网络、第一分支网络(即语义分割预测模型)和第二分支网络(即像素连接关系预测模型)组成的多任务全卷积神经网络。主干网络可以用于对输入图像进行特征提取。第一分支网络可以用于车道线语义分割预测,处理特征提取后的数据,得到图像的每个像素的类型信息,如背景、实线、虚线等。第二分支网络可以用于车道线像素连接关系预测,处理特征提取后的数据,得到图像中每个像素所指向下一行属于同一实例的车道线中心点的位置。
3、网络训练:
利用步骤1得到的训练数据,输入步骤2的卷积神经网络中,第一分支网络与第二分支网络的输出均可以采用交叉熵损失函数进行约束:
L=αL seg+βL link
L为网络的最终损失函数,其中α表示第一分支网络的语义分割损失在整体损失中的比例系数,β表示第二分支网络的像素连接损失在整体损失中的比例系数。
输入训练图像正向传播得到预测结果,根据相应的标签数据和损失函数计算损失,反向传播更新网络参数,得到最终的车道线检测网络模型。
4、后处理:
4.1、输入车道线像素连接关系预测图。
4.2、从图像最底行查找连通域,以连通域的中心点作为一个车道线实例的起点,并赋上车道线实例的ID,每个连通域的中心点根据该点的像素连接关系指示标签都会指向上一行的某个位置。
4.3、对图像上一行查找连通域,对于每个连通域,判断下一行的车道线实例是否指向该连通域,如果没有相关车道线实例,则该连通域为一个新的实车道线例,赋上新的ID;如果只有一个车道线实例指向该连通域,则该连通域继承这个车道线实例的ID;如果有多个车道线实例指向该连通域,则该连通域继承最早出现的ID。
4.4、重复步骤4.3,直到遍历完整张图像,得到像素连接后的实例分割结果。
4.5、将4.4的结果与第一分支网络的语义分割结果做“与”操作(即第一分支网络的语义分割结果提供像素的类型信息),可以滤除背景像素,得到最终的车道线实例分割结果,不仅有每个像素的类型信息,还有车道线实例的ID。
本申请实施例通过预测车道线实例的像素连接关系,得到车道线实例分割的结果,该检测方法简单,耗时短,并且能够支持同时得到多个车道线实例的结果,没有限定数量的局限;输入图像使用鸟瞰图,在远处的车道线也能得到更准确的效果。
参见图8,图8是本申请图像检测装置一实施例的结构示意图,需要说明的是,本实施例的图像检测装置能够执行上述第一方面及相关的图像检测方法中的操作,相关内容的详细说明,请参见上述相关内容,在此不再赘叙。
所述装置100包括:存储器1和处理器2;处理器2与存储器1通过总线连接。
其中,处理器2可以是微控制单元、中央处理单元或数字信号处理器,等等。
其中,存储器1可以是Flash芯片、只读存储器、磁盘、光盘、U盘或者移动硬盘等等。
所述存储器1用于存储计算机程序;所述处理器2用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
获取图像;对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:根据所述指示信息将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:获取图像中每个像素所在的连通域信息;根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:获取图像中每个像素的类型信息;根据每个像素的类型信息确定连接后的目标图像区域的类型。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
其中,所述目标图像区域包括线状图像区域。
其中,所述线状图像区域包括车道线图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连 接;根据所述图像中第一像素指向第二像素的指示信息和多个连通域,确定每个连通域的中心点的第一像素的指示信息;根据每个连通域的中心点的第一像素的指示信息,确定每个连通域的中心点的第一像素指向的第二像素所在的连通域;将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的连通域进行连接,得到连接后的目标图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:获取包括像素连接关系标签的训练图像;利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
其中,所述训练图像包括线状图像区域;所述处理器在执行所述计算机程序时,实现如下步骤:根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
本申请还提供另一种图像检测装置,需要说明的是,本实施例的图像检测装置能够执行上述第二方面及相关的图像检测方法中的操作,相关内容的详细说明,请参见上述相关内容,在此不再赘叙。
所述装置包括:存储器和处理器;所述存储器用于存储计算机程序;所述处理器用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
获取图像中每个像素所在的连通域信息;获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率;根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:根据所述连通域信息和所述指示信息,确定每个连通域的中心点的第一像素指向的另一个连通域的第二像素;所述将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域,包括:将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的另一个连通域进行连接,得到连接后的目标图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:获取图像中每个像素的类型信息;根据每个像素的类型信息确定连接后的目标图像区域的类型。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
其中,所述目标图像区域包括线状图像区域。
其中,所述线状图像区域包括车道线图像区域。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:获取包括像素连接关系标签的训练图像;利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
其中,所述处理器在执行所述计算机程序时,实现如下步骤:利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
其中,所述训练图像包括线状图像区域;所述处理器在执行所述计算机程序时,实现如下步骤:根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度 上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
本申请还提供一种计算机可读存储介质,所述计算机可读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使所述处理器实现如上第一方面及相关的任一项所述的图像检测方法。
其中,该计算机可读存储介质可以是上述图像检测装置的内部存储单元,例如硬盘或内存。该计算机可读存储介质也可以是外部存储设备,例如配备的插接式硬盘、智能存储卡、安全数字卡、闪存卡,等等。
本申请还提供另一种计算机可读存储介质,所述计算机可读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使所述处理器实现如上第二方面及相关的任一项所述的图像检测方法。
其中,该计算机可读存储介质可以是上述图像检测装置的内部存储单元,例如硬盘或内存。该计算机可读存储介质也可以是外部存储设备,例如配备的插接式硬盘、智能存储卡、安全数字卡、闪存卡,等等。
应当理解,在本申请说明书中所使用的术语仅仅是出于描述特定实施例的目的而并不意在限制本申请。
还应当理解,在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合,并且包括这些组合。
以上所述,仅为本申请的具体实施例,但本申请的保护范围并不局限于此,任何熟悉本技术领域的技术人员在本申请揭露的技术范围内,可轻易想到各种等效的修改或替换,这些修改或替换都应涵盖在本申请的保护范围之内。因此,本申请的保护范围应以权利要求的保护范围为准。

Claims (44)

  1. 一种图像检测方法,其特征在于,所述方法包括:
    获取图像;
    对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
  2. 根据权利要求1所述的方法,其特征在于,所述方法还包括:
    根据所述指示信息将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。
  3. 根据权利要求1所述的方法,其特征在于,所述方法还包括:
    获取图像中每个像素所在的连通域信息;
    根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
    将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  4. 根据权利要求2或3所述的方法,其特征在于,所述方法还包括:
    获取图像中每个像素的类型信息;
    根据每个像素的类型信息确定连接后的目标图像区域的类型。
  5. 根据权利要求4所述的方法,其特征在于,所述获取图像中每个像素的类型信息,包括:
    利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
  6. 根据权利要求2或3所述的方法,其特征在于,所述目标图像区域包括线状图像区域。
  7. 根据权利要求6所述的方法,其特征在于,所述线状图像区域包括车道线图像区域。
  8. 根据权利要求1所述的方法,其特征在于,所述对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,包括:
    利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接;
    所述方法还包括:
    根据所述图像中第一像素指向第二像素的指示信息和多个连通域,确定每个连通域的中心点的第一像素的指示信息;
    根据每个连通域的中心点的第一像素的指示信息,确定每个连通域的中心点的第一像素指向的第二像素所在的连通域;
    将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  9. 根据权利要求8所述的方法,其特征在于,所述利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域之前,包括:
    获取包括像素连接关系标签的训练图像;
    利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
  10. 根据权利要求9所述的方法,其特征在于,所述利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型,包括:
    利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
  11. 根据权利要求9所述的方法,其特征在于,所述训练图像包括线状图像区域;
    所述获取包括像素连接关系标签的训练图像,包括:
    根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;
    根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指 示标签的图像。
  12. 一种图像检测方法,其特征在于,所述方法包括:
    获取图像中每个像素所在的连通域信息;
    获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率;
    根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
    将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  13. 根据权利要求12所述的方法,其特征在于,所述根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素,包括:
    根据所述连通域信息和所述指示信息,确定每个连通域的中心点的第一像素指向的另一个连通域的第二像素;
    所述将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域,包括:
    将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的另一个连通域进行连接,得到连接后的目标图像区域。
  14. 根据权利要求12所述的方法,其特征在于,所述方法还包括:
    获取图像中每个像素的类型信息;
    根据每个像素的类型信息确定连接后的目标图像区域的类型。
  15. 根据权利要求14所述的方法,其特征在于,所述获取图像中每个像素的类型信息,包括:
    利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
  16. 根据权利要求12所述的方法,其特征在于,所述目标图像区域包括线状图像区域。
  17. 根据权利要求16所述的方法,其特征在于,所述线状图像区域包括车道线图像区域。
  18. 根据权利要求12所述的方法,其特征在于,所述获取图像中每个像素所在的连通域信息,获取所述图像中第一像素指向第二像素的指示信息,包括:
    利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接。
  19. 根据权利要求18所述的方法,其特征在于,所述利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域之前,包括:
    获取包括像素连接关系标签的训练图像;
    利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
  20. 根据权利要求18所述的方法,其特征在于,所述利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型,包括:
    利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
  21. 根据权利要求19所述的方法,其特征在于,所述训练图像包括线状图像区域;
    所述获取包括像素连接关系标签的训练图像,包括:
    根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;
    根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
  22. 一种图像检测装置,其特征在于,所述装置包括:存储器和处理器;
    所述存储器用于存储计算机程序;
    所述处理器用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
    获取图像;
    对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率。
  23. 根据权利要求22所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    根据所述指示信息将所述图像中的第一像素和第二像素进行连接,得到连接后的目标图像区域。
  24. 根据权利要求22所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    获取图像中每个像素所在的连通域信息;
    根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
    将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  25. 根据权利要求23或24所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    获取图像中每个像素的类型信息;
    根据每个像素的类型信息确定连接后的目标图像区域的类型。
  26. 根据权利要求25所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
  27. 根据权利要求23或24所述的装置,其特征在于,所述目标图像区域包括线状图像区域。
  28. 根据权利要求27所述的装置,其特征在于,所述线状图像区域包括车道线图像区域。
  29. 根据权利要求22所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接;
    根据所述图像中第一像素指向第二像素的指示信息和多个连通域,确定每个连通域的中心点的第一像素的指示信息;
    根据每个连通域的中心点的第一像素的指示信息,确定每个连通域的中心点的第一像素指向的第二像素所在的连通域;
    将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  30. 根据权利要求29所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    获取包括像素连接关系标签的训练图像;
    利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
  31. 根据权利要求30所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
  32. 根据权利要求30所述的装置,其特征在于,所述训练图像包括线状图像区域;
    所述处理器在执行所述计算机程序时,实现如下步骤:
    根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;
    根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
  33. 一种图像检测装置,其特征在于,所述装置包括:存储器和处理器;
    所述存储器用于存储计算机程序;
    所述处理器用于执行所述计算机程序并在执行所述计算机程序时,实现如下步骤:
    获取图像中每个像素所在的连通域信息;
    获取所述图像中第一像素指向第二像素的指示信息,其中在所述第一像素的周围像素中所述第二像素与所述第一像素具有最大的连接关系概率;
    根据所述连通域信息和所述指示信息,确定位于不同连通域上的第一像素和第二像素;
    将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域。
  34. 根据权利要求33所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    根据所述连通域信息和所述指示信息,确定每个连通域的中心点的第一像素指向的另一个连通域的第二像素;
    所述将位于不同连通域上的第一像素和第二像素所在的连通域进行连接,得到连接后的目标图像区域,包括:
    将每个连通域的中心点的第一像素所在的连通域和每个连通域的中心点的第一像素指向的第二像素所在的另一个连通域进行连接,得到连接后的目标图像区域。
  35. 根据权利要求33所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    获取图像中每个像素的类型信息;
    根据每个像素的类型信息确定连接后的目标图像区域的类型。
  36. 根据权利要求35所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用已训练的语义分割预测模型对所述图像进行检测,得到所述图像中每个像素的类型信息。
  37. 根据权利要求33所述的装置,其特征在于,所述目标图像区域包括线状图像区域。
  38. 根据权利要求37所述的装置,其特征在于,所述线状图像区域包括车道线图像区域。
  39. 根据权利要求33所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用已训练的像素连接关系预测模型对所述图像进行检测,得到所述图像中第一像素指向第二像素的指示信息和多个连通域,所述连通域中的第一像素和第二像素已连接。
  40. 根据权利要求39所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    获取包括像素连接关系标签的训练图像;
    利用包括像素连接关系标签的训练图像对模型进行训练,得到像素连接关系预测模型。
  41. 根据权利要求39所述的装置,其特征在于,所述处理器在执行所述计算机程序时,实现如下步骤:
    利用包括像素连接关系标签的训练图像对模型进行训练,得到损失函数满足预设要求的像素连接关系预测模型。
  42. 根据权利要求40所述的装置,其特征在于,所述训练图像包括线状图像区域;
    所述处理器在执行所述计算机程序时,实现如下步骤:
    根据每个线状图像区域在所述训练图像的每行中心点,确定每个线状图像区域的每行中心点的像素指向上一行中心点的像素的连接关系指示标签;
    根据每个线状图像区域的每行中心点的像素的连接关系指示标签和每个线状图像区域的宽度,确定每个线状图像区域的每行中心点两边宽度上的点的像素的连接关系指示标签,进而得到所述线状图像区域所有像素的连接关系指示标签的图像。
  43. 一种计算机可读存储介质,其特征在于,所述计算机可读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使所述处理器实现如权利要求1-11任一项所述的图像检测方法。
  44. 一种计算机可读存储介质,其特征在于,所述计算机可读存储介质存 储有计算机程序,所述计算机程序被处理器执行时使所述处理器实现如权利要求12-21任一项所述的图像检测方法。
PCT/CN2021/076937 2021-02-19 2021-02-19 图像检测方法、装置及存储介质 Ceased WO2022174397A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/076937 WO2022174397A1 (zh) 2021-02-19 2021-02-19 图像检测方法、装置及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/076937 WO2022174397A1 (zh) 2021-02-19 2021-02-19 图像检测方法、装置及存储介质

Publications (1)

Publication Number Publication Date
WO2022174397A1 true WO2022174397A1 (zh) 2022-08-25

Family

ID=82931909

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/076937 Ceased WO2022174397A1 (zh) 2021-02-19 2021-02-19 图像检测方法、装置及存储介质

Country Status (1)

Country Link
WO (1) WO2022174397A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106340020A (zh) * 2015-07-10 2017-01-18 阿里巴巴集团控股有限公司 超像素分割方法和装置
CN108496202A (zh) * 2017-08-31 2018-09-04 深圳市大疆创新科技有限公司 斑点检测方法、系统、存储介质及计算机程序产品
CN111460866A (zh) * 2019-01-22 2020-07-28 北京市商汤科技开发有限公司 车道线检测及驾驶控制方法、装置和电子设备

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106340020A (zh) * 2015-07-10 2017-01-18 阿里巴巴集团控股有限公司 超像素分割方法和装置
CN108496202A (zh) * 2017-08-31 2018-09-04 深圳市大疆创新科技有限公司 斑点检测方法、系统、存储介质及计算机程序产品
CN111460866A (zh) * 2019-01-22 2020-07-28 北京市商汤科技开发有限公司 车道线检测及驾驶控制方法、装置和电子设备

Similar Documents

Publication Publication Date Title
CN113191256B (zh) 车道线检测模型的训练方法、装置、电子设备及存储介质
CN112232293B (zh) 图像处理模型训练、图像处理方法及相关设备
CN109117848B (zh) 一种文本行字符识别方法、装置、介质和电子设备
CN108038474B (zh) 人脸检测方法、卷积神经网络参数的训练方法、装置及介质
CN105144239B (zh) 图像处理装置、图像处理方法
CN106951830B (zh) 一种基于先验条件约束的图像场景多对象标记方法
CN110533950A (zh) 车位使用状况的检测方法、装置、电子设备及存储介质
CN109086811A (zh) 多标签图像分类方法、装置及电子设备
CN113516144B (zh) 目标检测方法及装置、计算设备
CN113920158B (zh) 跟踪模型的训练与交通对象跟踪方法、装置
US10945888B2 (en) Intelligent blind guide method and apparatus
CN114373088B (zh) 一种图像检测模型的训练方法和相关产品
CN103383732A (zh) 图像处理方法和装置
CN112329605B (zh) 一种城市容貌乱贴乱画行为识别方法,存储装置及服务器
CN115063770B (zh) 车道线检测方法、模型训练方法、装置、设备及存储介质
WO2022141858A1 (zh) 行人检测方法、装置、电子设备及存储介质
CN114519853A (zh) 一种基于多模态融合的三维目标检测方法及系统
JP2025503714A (ja) センサデータを自動的にアノテーションする方法およびシステム
CN115830399B (zh) 分类模型训练方法、装置、设备、存储介质和程序产品
CN112329810B (zh) 一种基于显著性检测的图像识别模型训练方法及装置
CN114708426A (zh) 目标检测方法、模型训练方法、装置、设备及存储介质
CN111652181A (zh) 目标跟踪方法、装置及电子设备
CN111008576A (zh) 行人检测及其模型训练、更新方法、设备及可读存储介质
CN112989869B (zh) 人脸质量检测模型的优化方法、装置、设备及存储介质
CN114387600B (zh) 文本特征识别方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21926117

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: 21926117

Country of ref document: EP

Kind code of ref document: A1