WO2019237342A1 - 用于语义分割的分类神经网络的训练方法及装置、电子设备 - Google Patents

用于语义分割的分类神经网络的训练方法及装置、电子设备 Download PDF

Info

Publication number
WO2019237342A1
WO2019237342A1 PCT/CN2018/091542 CN2018091542W WO2019237342A1 WO 2019237342 A1 WO2019237342 A1 WO 2019237342A1 CN 2018091542 W CN2018091542 W CN 2018091542W WO 2019237342 A1 WO2019237342 A1 WO 2019237342A1
Authority
WO
WIPO (PCT)
Prior art keywords
classification
pixel
weight
neural network
loss
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/CN2018/091542
Other languages
English (en)
French (fr)
Inventor
石路
张聪
王琪
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.)
Fujitsu Ltd
Original Assignee
Fujitsu 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 Fujitsu Ltd filed Critical Fujitsu Ltd
Priority to CN201880092678.3A priority Critical patent/CN112020721A/zh
Priority to PCT/CN2018/091542 priority patent/WO2019237342A1/zh
Publication of WO2019237342A1 publication Critical patent/WO2019237342A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition

Definitions

  • the present invention relates to the field of information technology, and in particular, to a method and device for training a classified neural network for semantic segmentation, and an electronic device.
  • Semantic segmentation is one of the latest technologies combining classification neural networks and image coding and decoding technologies such as Fully Convolutional Networks (FCN). Similar to other neural network operations, the dataset used for semantic segmentation is very important. But different from the problem of object classification solved by ordinary neural networks, image-level segmentation needs to be performed at the pixel level. In other words, object detection only needs to identify the classification of pictures and add location information on this basis. However, semantic segmentation needs to classify all pixels and their position information. It has higher requirements for data annotation.
  • FCN Fully Convolutional Networks
  • the current semantic segmentation algorithms have defects in the processing of details. For example, in the boundary area of different objects, there is a large uncertainty in the annotation and classification results. Moreover, due to these uncertainties, the loss of the neural network during the training process continues to oscillate, so the neural network model is not easy to converge, resulting in a slow training speed and poor model stability.
  • Embodiments of the present invention provide a method and device for training a classified neural network for semantic segmentation, and an electronic device.
  • a weight of an uncertain pixel in a boundary region to be small when calculating a loss of the neural network, the weight can be reduced.
  • the influence of these pixels on the training speed and model stability effectively improves the training speed and model stability, and can ensure the accuracy of the training results of the main part of the training image.
  • a method for training a classified neural network for semantic segmentation includes: determining a boundary line between regions in a truth image of a training image; Expand the line to obtain a boundary extension area; set the weight of a specific pixel in the boundary extension area to be less than 1, the specific pixel is the same as the classification result of at least one of the adjacent areas of the boundary extension area
  • the weight of the pixels in the training image other than the specific pixel is set to 1; calculating the loss of the classification neural network according to the weight of each pixel in the training image; and according to the calculated The loss, adjusting parameters of the classification neural network.
  • a training device for a classified neural network for semantic segmentation includes a determination unit for determining a boundary between regions in a truth value map of a training image. Line; an expansion unit for expanding the boundary line to obtain a boundary expansion area; a first setting unit for setting a weight of a specific pixel in the boundary expansion area to be less than 1, the specific The pixel is the same pixel as the classification result of at least one of the adjacent areas of the boundary extension area; the second setting unit is configured to set weights of pixels other than the specific pixel in the training image Is 1; a calculation unit configured to calculate a loss of the classification neural network according to a weight value of each pixel in the training image; and an adjustment unit configured to adjust the classification nerve according to the calculated loss The parameters of the network.
  • an electronic device including the apparatus according to the second aspect of the embodiments of the present invention.
  • the beneficial effect of the present invention is that by setting the weights of pixels with uncertainty in the boundary region to be small when calculating the loss of the neural network, the influence of these pixels on the training speed and model stability can be reduced, thereby effectively improving The training speed and model stability can ensure the accuracy of the training results of the main part of the training image.
  • FIG. 1 is a schematic diagram of a training method of a classification neural network for semantic segmentation according to Embodiment 1 of the present invention
  • Embodiment 1 of the present invention is a training image of Embodiment 1 of the present invention
  • FIG. 3 is a truth map of the training image shown in FIG. 2;
  • FIG. 5 is an image obtained by performing boundary line expansion on the boundary line image shown in FIG. 4;
  • FIG. 6 is another schematic diagram of a training method of a classification neural network for semantic segmentation according to Embodiment 1 of the present invention.
  • FIG. 7 is a schematic diagram of a training apparatus for a classification neural network for semantic segmentation according to Embodiment 2 of the present invention.
  • FIG. 8 is a schematic diagram of an electronic device according to Embodiment 3 of the present invention.
  • FIG. 9 is a schematic block diagram of a system configuration of an electronic device according to Embodiment 3 of the present invention.
  • FIG. 1 is a schematic diagram of a training method of a classification neural network for semantic segmentation according to Embodiment 1 of the present invention. As shown in Figure 1, the method includes:
  • Step 101 Determine a boundary line between various regions in the truth image of the training image
  • Step 102 Expand the boundary line to obtain a boundary extension area.
  • Step 103 Set a weight of a specific pixel in the boundary extension area to be less than 1, the specific pixel is a pixel having the same classification result as at least one of adjacent areas of the boundary extension area;
  • Step 104 Set the weights of pixels other than the specific pixel in the training image to 1.
  • Step 105 calculating the loss of the classification neural network according to the weight of each pixel in the training image.
  • Step 106 Adjust the parameters of the classification neural network according to the calculated loss.
  • the classification neural network may be various types of classification neural networks, such as FCN and the like.
  • the training image may be various types of images, for example, a surveillance video image.
  • the training image may be obtained according to various methods, for example, the monitoring video image is cropped to obtain multiple training images.
  • an existing method may be used to obtain a truth map of the training image and determine a boundary line between various regions in the truth map.
  • a truth map is obtained by a manual labeling method
  • a boundary line is obtained by a line detection method.
  • FIG. 2 is a training image of Embodiment 1 of the present invention
  • FIG. 3 is a truth value map of the training image shown in FIG. 2
  • FIG. 4 is a boundary line image obtained after line detection is performed on the truth value map shown in FIG. 3.
  • the training image shown in FIG. 2 is manually labeled, and the main parts are roughly determined, for example, roads, trees, and isolation zones, but the boundary areas of roads and trees are blurred, for example.
  • FIG. 4 after line detection is performed on the truth map shown in FIG. 3, a boundary line marked with a line is obtained.
  • step 102 each of the obtained boundary lines is extended to obtain a plurality of boundary extension regions.
  • FIG. 5 is an image obtained by performing boundary line expansion on the boundary line image shown in FIG. 4. As shown in FIG. 5, the expanded boundary line becomes thicker, and the number of pixels included in the expanded boundary area is larger than that before the expansion. In this way, it is easy to ignore the uncertain details in the boundary area and focus on the overall training.
  • various methods can be used for the expansion of each boundary line, for example, pixels in a certain range near the boundary line are divided into pixels in a boundary extension area.
  • a weight of a specific pixel in each boundary extension region is set to be less than 1, and the specific pixel is a pixel having the same classification result as at least one of adjacent regions of the boundary extension region.
  • the two adjacent areas are the first adjacent area and the second adjacent area, where the classification result of the first adjacent area is "first classification” and the second adjacent area is The classification result of the region is “second classification”.
  • a pixel whose classification result in the boundary extension region is “first classification” or “second classification” is taken as a specific pixel.
  • the weights of these specific pixels are set to 1 for calculating the loss of the classification neural network.
  • these specific pixels are actually pixels with uncertain classification results in the boundary extension area.
  • uncertain pixels by reducing the weight of their calculation loss, their influence on the convergence of the neural network model can be reduced, and , Will not affect the training results of the main part.
  • the classification results of these pixels and regions are determined by their true values, and their true values can be determined according to the output of the classifier. For example, it is determined based on the output of softmax.
  • the weight of the classification result in the specific pixel is the second classification and the pixels adjacent to the first adjacent area
  • the weight of the classification result in the specific pixel is the first classification and the pixels adjacent to the second adjacent area.
  • the value is set to a first weight value less than 1
  • the weight of other pixels in a specific pixel is set to a second weight value less than 1, the first weight value is greater than the second weight value.
  • the weight of the classification result in the specific pixel is the second classification and the pixel adjacent to the first adjacent area
  • the weight of the classification result in the specific pixel is the first classification and the pixel adjacent to the second adjacent area is set to 0.8
  • the weight of these pixels is improved Value can further improve the accuracy of the training results.
  • step 104 the weights of pixels other than the specific pixel in the training image are set to 1, that is, the initial weights of these pixels are maintained.
  • the loss of the classification neural network is calculated according to the weight of each pixel in the training image.
  • Various methods can be used to calculate the loss of the classification neural network. For example, according to the weight of each pixel in the training image, the cumulative value of the cross-entropy of each pixel is calculated to obtain the loss of the classification neural network.
  • the loss of the classification neural network can be calculated according to the following formula (1):
  • L (i) represents the ith training image set
  • B represents the number of training images in each training image set
  • h represents the height of the training image
  • w represents the width of the training image
  • Weight (j) represents the training image
  • H (j) represents the crossover between the j-th pixel in the training image.
  • the camera angle when calculating the loss of the classification neural network, can also be considered.
  • the cumulative value of the cross-entropy of each pixel is calculated according to the weight of each pixel in the training image and the imaging angle of each pixel, so as to obtain the loss of the classification neural network.
  • the loss of the classification neural network can also be calculated according to the following formula (2):
  • L (i) represents the ith training image set
  • B represents the number of training images in each training image set
  • h represents the height of the training image
  • w represents the width of the training image
  • Weight (j) represents the training image
  • H (j) represents the intersection of the j-th pixel in the training image
  • Angle (j) represents the camera angle of the j-th pixel in the training image.
  • the imaging angle of each pixel can be obtained according to an existing method. For example, it is obtained from data marked when the imaging device is shipped from the factory.
  • the factors of the camera angle of each pixel are considered when calculating the network loss, which further enhances the role of the main subject (the camera angle is larger) in the training process and ignores the secondary subject (the camera angle is smaller). Its role in the training process can further improve the speed of training and the stability of the model.
  • step 105 when calculating the loss of the classification neural network, the probability of belonging to a specific classification may be increased for the pixels in the boundary extension area.
  • the specific classification is a classification of interest to the user. For example, if the user is more concerned about the integrity of the lane, the probability that a pixel in the boundary extension area is determined to be a lane can be increased.
  • the method may further include: setting a weight of a pixel of the uninteresting region to be less than 1 for calculating a loss of the classification neural network. At this time, the weights of the specific pixels and pixels other than the pixels of the uninteresting area are set to 1.
  • the parameters of the classification neural network are adjusted according to the calculated loss. For example, according to the partial derivatives of the loss for the weight parameters and the bias parameters, the weight parameters and bias parameters of each layer of the classification neural network are adjusted.
  • FIG. 6 is another schematic diagram of a training method of a classification neural network for semantic segmentation according to Embodiment 1 of the present invention. As shown in Figure 6, the method includes:
  • Step 601 Determine a boundary line between various regions in the truth image of the training image
  • Step 602 Expand the boundary line to obtain a boundary extension area.
  • Step 603 Set a weight of a specific pixel in the boundary extension area to be less than 1, the specific pixel is a pixel having the same classification result as at least one of adjacent areas of the boundary extension area;
  • Step 604 Set the weights of pixels other than the specific pixel in the training image to 1.
  • Step 605 Calculate the loss of the classification neural network according to the weight of each pixel in the training image
  • Step 606 Adjust the parameters of the classification neural network according to the calculated loss.
  • Step 607 Determine whether the loss is less than a preset threshold. When the determination result is “Yes”, proceed to step 608, and when the determination result is "No”, proceed to step 609;
  • Step 608 increase the weight of the specific pixel and maintain the weight of other pixels
  • Step 609 Determine whether the convergence condition is satisfied. When the determination result is "Yes”, the training is terminated, and when the determination result is "No", proceed to Step 610;
  • Step 610 Select a new training image.
  • step 607 and step 608 it is determined whether the loss is less than a preset threshold, and the preset threshold may be set according to actual needs.
  • the weight of the specific pixel is increased to recalculate the loss and adjust parameters. In this way, when the classification neural network model is stable, gradually increasing the weight of a specific pixel can further improve the accuracy of the training result when the model is stable.
  • the value of the added weight may be determined according to the actual situation. For example, each time you increase the original weight by 20%, the maximum weight after the increase does not exceed 1.
  • the convergence condition may be set in advance, for example, the loop reaches a preset value this time, or the loss value of the network is reduced to the preset value.
  • FIG. 7 is a schematic diagram of a training apparatus for a classification neural network for semantic segmentation according to Embodiment 2 of the present invention. As shown in FIG. 7, the apparatus 700 includes:
  • a determining unit 701 configured to determine a boundary line between regions in a truth image of a training image
  • An expansion unit 702 configured to expand the boundary line to obtain a boundary extension area
  • a first setting unit 703, configured to set a weight of a specific pixel in the boundary extension area to be less than 1, the specific pixel being a pixel having the same classification result as at least one of adjacent areas of the boundary extension area;
  • a second setting unit 704 configured to set the weights of pixels other than specific pixels in the training image to 1;
  • a calculation unit 705, configured to calculate the loss of the classification neural network according to the weight of each pixel in the training image
  • An adjusting unit 706 is configured to adjust parameters of the classification neural network according to the calculated loss.
  • the apparatus 700 may further include:
  • a third setting unit 707 is configured to increase the weight of a specific pixel when the loss of the classification neural network is less than a preset threshold, to recalculate the loss and adjust the parameters of the classification neural network according to the loss.
  • the apparatus 700 may further include:
  • a fourth setting unit 708 is configured to set the weight of the pixels of the uninteresting region to be less than 1 for calculating the loss of the classification neural network.
  • FIG. 8 is a schematic diagram of the electronic device according to Embodiment 3 of the present invention.
  • the electronic device 800 includes a classification neural network training device 801 for semantic segmentation.
  • the structure and function of the classification neural network training device 801 for semantic segmentation are the same as those described in Embodiment 2. I won't repeat them here.
  • FIG. 9 is a schematic block diagram of a system configuration of an electronic device according to Embodiment 3 of the present invention.
  • the electronic device 900 may include a central processing unit 901 and a memory 902; the memory 902 is coupled to the central processing unit 901.
  • the figure is exemplary; other types of structures can also be used to supplement or replace the structure to implement telecommunication functions or other functions.
  • the electronic device 900 may further include an input unit 903, a display 904, and a power supply 905.
  • the function of the training device for the classified neural network for semantic segmentation described in Embodiment 1 may be integrated into the central processor 901.
  • the central processing unit 901 may be configured to: determine a boundary line between various regions in a truth image of a training image; expand the boundary line to obtain a boundary extension region; and expand a specific pixel in the boundary extension region
  • the weight is set to less than 1, the specific pixel is the same pixel as the classification result of at least one of the adjacent areas of the boundary extension area; and the other pixels other than the specific pixel in the training image are
  • the weight is set to 1; the loss of the classification neural network is calculated according to the weight of each pixel in the training image; and the parameters of the classification neural network are adjusted according to the calculated loss.
  • the adjacent regions of the boundary extension region include a first adjacent region and a second adjacent region.
  • the classification result of the first adjacent region is a first classification
  • the classification result of the second adjacent region is A second classification
  • the setting a weight of a specific pixel in the boundary extension region to be less than 1 including: setting a classification result in the specific pixel to a pixel of the second classification and adjacent to the first adjacent region ,
  • the classification result in the specific pixel is the first classification and the weight of pixels adjacent to the second adjacent region is set to a first weight less than 1, and the weights of other pixels in the specific pixel are set It is set to a second weight that is less than 1, and the first weight is greater than the second weight.
  • calculating the loss of the classification neural network according to the weights of all pixels in the training image includes: calculating a cumulative value of the cross-entropy of each pixel according to the weights of each pixel in the training image, Thus, the loss of the classification neural network is obtained.
  • the calculating the loss of the classification neural network according to the weights of all pixels in the training image includes: calculating the weight of each pixel according to the weight of each pixel in the training image and the camera angle of each pixel. The cumulative value of the cross entropy is obtained to obtain the loss of the classification neural network.
  • the central processing unit 901 may be further configured to: when the loss of the classification neural network is less than a preset threshold, increase the weight of the specific pixel to recalculate the loss and perform the calculation based on the loss. The adjustment of the parameters of the classification neural network is described.
  • the probability of belonging to a specific classification is increased.
  • the central processing unit 901 may be further configured to set a weight of a pixel of an uninteresting region to be less than 1 for calculating a loss of the classification neural network.
  • the training device for the classification neural network for semantic segmentation described in Embodiment 1 may be configured separately from the central processor 901.
  • the training device for the classification neural network for semantic segmentation may be configured as The chip connected to the central processing unit 901 implements the function of a training device for a classification neural network for semantic segmentation through the control of the central processing unit 901.
  • the central processing unit 901 is sometimes also called a controller or an operation control, and may include a microprocessor or other processor devices and / or logic devices.
  • the central processing unit 901 receives input and controls various components of the electronic device 900. Operation.
  • the memory 902 may be, for example, one or more of a buffer, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable devices.
  • the central processing unit 901 may execute the program stored in the memory 902 to implement information storage or processing.
  • the functions of other components are similar to the existing ones and will not be repeated here.
  • the components of the electronic device 900 may be implemented by dedicated hardware, firmware, software, or a combination thereof without departing from the scope of the present invention.
  • An embodiment of the present invention also provides a computer-readable program, wherein when the program is executed in a training device or an electronic device for a classification neural network for semantic segmentation, the program causes a computer to The training device or the electronic device of the segmented classification neural network executes the training method of the classification neural network for semantic segmentation described in Embodiment 1.
  • An embodiment of the present invention also provides a storage medium storing a computer-readable program, wherein the computer-readable program causes a computer to execute the classification neural network training apparatus or electronic device for semantic segmentation in Embodiment 1 A classification neural network training method for semantic segmentation.
  • the method for training in a classification neural network training device for semantic segmentation described in combination with the embodiments of the present invention may be directly embodied as hardware, a software module executed by a processor, or a combination of the two.
  • one or more of the functional block diagrams and / or one or more combinations of the functional block diagrams shown in FIG. 7 may correspond to each software module of a computer program flow or each hardware module.
  • These software modules can respectively correspond to the steps shown in FIG. 1.
  • These hardware modules can be implemented by using a field programmable gate array (FPGA) to cure these software modules.
  • FPGA field programmable gate array
  • a software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
  • a storage medium may be coupled to the processor so that the processor can read information from and write information to the storage medium; or the storage medium may be a component of the processor.
  • the processor and the storage medium may reside in an ASIC.
  • the software module can be stored in the memory of the mobile terminal or in a memory card that can be inserted into the mobile terminal.
  • the software module may be stored in the MEGA-SIM card or a large-capacity flash memory device.
  • One or more of the functional block diagrams described in FIG. 7 and / or one or more combinations of the functional block diagrams may be implemented as a general-purpose processor, a digital signal processor (DSP), or a dedicated processor for performing the functions described in this application.
  • DSP digital signal processor
  • ASICs Integrated circuits
  • FPGAs field programmable gate arrays
  • FIG. 7 One or more of the functional block diagrams described in FIG. 7 and / or one or more combinations of the functional block diagrams may be implemented as a general-purpose processor, a digital signal processor (DSP), or a dedicated processor for performing the functions described in this application.
  • ASICs Integrated circuits
  • FPGAs field programmable gate arrays
  • FIG. 7 One or more of the functional block diagrams described in FIG.
  • FIG. 7 and / or one or more combinations of the functional block diagrams may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, One or more microprocessors in conjunction with DSP communications or any other such configuration.

Landscapes

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

Abstract

一种用于语义分割的分类神经网络的训练装置及方法、电子设备。通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。

Description

用于语义分割的分类神经网络的训练方法及装置、电子设备 技术领域
本发明涉及信息技术领域，尤其涉及一种语义分割的分类神经网络的训练方法及装置、电子设备。
背景技术
语义分割是结合例如全卷积网络(Fully Convolutional Networks,FCN)等分类神经网络和图像编解码技术的最新技术之一。与其他神经网络操作类似，用于语义分割的数据集非常重要。但与普通神经网络解决的对象分类的课题不同，图像语义分割需要进行的是像素级分割。换句话说，对象检测只需要识别图片的分类，并在此基础上添加位置信息。但是，语义分割需要将所有像素与其位置信息进行分类。它对数据注释有更高的要求。
但是，在数据注释和对于分割效果的要求方面存在如下的问题：某些物体的边界模糊，并且没有准确的边界标记标准；场景的次要部分无法从主要部分分离出来；缺少后处理。目前，出现了一些语义分割算法，针对主要部分能够提供较为准确的分类结果。
应该注意，上面对技术背景的介绍只是为了方便对本发明的技术方案进行清楚、完整的说明，并方便本领域技术人员的理解而阐述的。不能仅仅因为这些方案在本发明的背景技术部分进行了阐述而认为上述技术方案为本领域技术人员所公知。
发明内容
但是，目前的语义分割算法在对细节的处理上存在缺陷，例如，在不同物体的边界区域，存在较大的注释和分类结果的不确定性。并且，由于这些不确定性，导致训练过程中的神经网络损失持续振荡，因此神经网络模型不易收敛，导致训练速度较慢且模型的稳定性较差。
本发明实施例提供一种语义分割的分类神经网络的训练方法及装置、电子设备，通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型 稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
根据本发明实施例的第一方面，提供一种用于语义分割的分类神经网络的训练方法，所述方法包括：确定训练图像的真值图中各个区域之间的边界线；对所述边界线进行扩展，获得边界扩展区域；将所述边界扩展区域中的特定像素的权值设置为小于1，所述特定像素是与所述边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；将所述训练图像中的所述特定像素以外的其他像素的权值设置为1；根据所述训练图像中各个像素的权值，计算所述分类神经网络的损失；以及根据计算出的所述损失，调整所述分类神经网络的参数。
根据本发明实施例的第二方面，提供一种用于语义分割的分类神经网络的训练装置，所述装置包括：确定单元，其用于确定训练图像的真值图中各个区域之间的边界线；扩展单元，其用于对所述边界线进行扩展，获得边界扩展区域；第一设置单元，其用于将所述边界扩展区域中的特定像素的权值设置为小于1，所述特定像素是与所述边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；第二设置单元，其用于将所述训练图像中的所述特定像素以外的其他像素的权值设置为1；计算单元，其用于根据所述训练图像中各个像素的权值，计算所述分类神经网络的损失；以及调整单元，其用于根据计算出的所述损失，调整所述分类神经网络的参数。
根据本发明实施例的第三方面，提供一种电子设备，所述电子设备包括根据本发明实施例的第二方面所述的装置。
本发明的有益效果在于：通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
参照后文的说明和附图，详细公开了本发明的特定实施方式，指明了本发明的原理可以被采用的方式。应该理解，本发明的实施方式在范围上并不因而受到限制。在所附权利要求的精神和条款的范围内，本发明的实施方式包括许多改变、修改和等同。
针对一种实施方式描述和/或示出的特征可以以相同或类似的方式在一个或更多个其它实施方式中使用，与其它实施方式中的特征相组合，或替代其它实施方式中的特征。
应该强调，术语“包括/包含”在本文使用时指特征、整件、步骤或组件的存在， 但并不排除一个或更多个其它特征、整件、步骤或组件的存在或附加。
附图说明
所包括的附图用来提供对本发明实施例的进一步的理解，其构成了说明书的一部分，用于例示本发明的实施方式，并与文字描述一起来阐释本发明的原理。显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。在附图中：
图1是本发明实施例1的用于语义分割的分类神经网络的训练方法的一示意图；
图2是本发明实施例1的训练图像；
图3是图2所示的训练图像的真值图；
图4是对图3所示的真值图进行线检测后获得的边界线图像；
图5是对图4所示的边界线图像实施边界线扩展后获得的图像；
图6是本发明实施例1的用于语义分割的分类神经网络的训练方法的另一示意图；
图7是本发明实施例2的用于语义分割的分类神经网络的训练装置的一示意图；
图8是本发明实施例3的电子设备的一示意图；
图9是本发明实施例3的电子设备的系统构成的一示意框图。
具体实施方式
参照附图，通过下面的说明书，本发明的前述以及其它特征将变得明显。在说明书和附图中，具体公开了本发明的特定实施方式，其表明了其中可以采用本发明的原则的部分实施方式，应了解的是，本发明不限于所描述的实施方式，相反，本发明包括落入所附权利要求的范围内的全部修改、变型以及等同物。
实施例1
本发明实施例提供一种用于语义分割的分类神经网络的训练方法。图1是本发明实施例1的用于语义分割的分类神经网络的训练方法的一示意图。如图1所示，该方法包括：
步骤101：确定训练图像的真值图中各个区域之间的边界线；
步骤102：对该边界线进行扩展，获得边界扩展区域；
步骤103：将该边界扩展区域中的特定像素的权值设置为小于1，该特定像素是 与该边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；
步骤104：将该训练图像中的该特定像素以外的其他像素的权值设置为1；
步骤105：根据该训练图像中各个像素的权值，计算该分类神经网络的损失；以及
步骤106：根据计算出的损失，调整该分类神经网络的参数。
由上述实施例可知，通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
在本实施例中，该分类神经网络可以是各种类型的分类神经网络，例如，FCN等。
在本实施例中，训练图像可以是各种类型的图像，例如，监控视频图像。
在本实施例中，训练图像可以根据各种方式而获得，例如，对监控视频图像进行裁剪而获得多个训练图像。
在步骤101中，可以使用现有方法来获得训练图像的真值图并确定真值图中各个区域之间的边界线。例如，通过人工标记的方法获得真值图，并通过线检测的方法获得边界线。
图2是本发明实施例1的训练图像；图3是图2所示的训练图像的真值图；图4是对图3所示的真值图进行线检测后获得的边界线图像。
如图3所示，对图2所示的训练图像进行人工标记，大致确定了各个主要部分，例如，道路，树木，隔离带等，但是，例如道路和树木的边界区域较为模糊。如图4所示，对图3所示的真值图进行线检测后，获得了以线条标记的边界线。
在步骤102中，对获得的各个边界线进行了扩展，获得了多个边界扩展区域。
图5是对图4所示的边界线图像实施边界线扩展后获得的图像。如图5所示，经过扩展后的边界线变粗，边界扩展区域包含的像素的数量相比于扩展之前增大。这样，便于忽略边界区域中不确定的细节而注重整体的训练。
在本实施例中，对各个边界线扩展可以使用各种方法，例如，将边界线附近一定范围内的像素划分为边界扩展区域内的像素。
在步骤103中，将各个边界扩展区域中的特定像素的权值设置为小于1，该特定像素是与该边界扩展区域的相邻区域中的至少一个的分类结果相同的像素。
例如，对于某个边界扩展区域，其相邻的两个区域为第一相邻区域和第二相邻区域，其中，第一相邻区域的分类结果为“第一分类”，第二相邻区域的分类结果为“第二分类”，此时，将该边界扩展区域中的分类结果是“第一分类”或者“第二分类”的像素作为特定像素。
在本实施例中，在确定了该边界扩展区域中的所有特定像素之后，将这些特定像素的权值设置为1，以用于计算分类神经网络的损失。
这样，这些特定像素实际上是边界扩展区域中分类结果存在不确定性的像素，对于这些不确定的像素，通过减少其计算损失时的权值，能够降低其对神经网络模型收敛的影响，并且，不会影响主要部分的训练结果。
在本实施例中，这些像素以及区域的分类结果是通过其真值而确定的，其真值可以根据分类器的输出而确定。例如，根据softmax的输出结果而确定。
在本实施例中，可以将特定像素中的分类结果为第二分类且邻近第一相邻区域的像素、以及特定像素中的分类结果为第一分类且邻近第二相邻区域的像素的权值设置为小于1的第一权值，将特定像素中的其他像素的权值设置为小于1的第二权值，该第一权值大于第二权值。
例如，将特定像素中的分类结果为第二分类且邻近第一相邻区域的像素，以及特定像素中的分类结果为第一分类且邻近第二相邻区域的像素的权值设置为0.8，将特定像素中的其他像素的权值设置为0.5。
这样，由于分类结果为第二分类且邻近第一相邻区域的像素分类结果为第一分类且邻近第二相邻区域的像素的分类结果是错误的可能性较大，通过提高这些像素的权值，能够进一步提高训练结果的准确性。
在步骤104中，将该训练图像中的该特定像素以外的其他像素的权值设置为1，也就是说，保持这些像素的初始权值。
在步骤105中，根据该训练图像中各个像素的权值，计算该分类神经网络的损失。计算该分类神经网络的损失可以使用各种方法，例如，根据该训练图像中的各个像素的权值，计算各个像素的交叉熵的累计值，从而获得该分类神经网络的损失。
例如，可以根据以下的公式(1)计算该分类神经网络的损失：
Figure PCTCN2018091542-appb-000001
其中，L(i)表示第i个训练图像集，B表示每个训练图像集中的训练图像的数量， h表示训练图像的高度，w表示该训练图像的宽度，Weight(j)表示该训练图像中第j个像素的权值，H(j)表示该训练图像中第j个像素的交叉墒。
在本实施例中，在计算分类神经网络的损失时，还可以考虑摄像角度。例如，根据训练图像中的各个像素的权值以及各个像素的摄像角度，计算各个像素的交叉熵的累计值，从而获得该分类神经网络的损失。
例如，还可以根据以下的公式(2)计算该分类神经网络的损失：
Figure PCTCN2018091542-appb-000002
其中，L(i)表示第i个训练图像集，B表示每个训练图像集中的训练图像的数量，h表示训练图像的高度，w表示该训练图像的宽度，Weight(j)表示该训练图像中第j个像素的权值，H(j)表示该训练图像中第j个像素的交叉墒，Angle(j)表示该训练图像中第j个像素的摄像角度。
在本实施例中，各个像素的摄像角度可以根据现有方法而获得。例如，根据摄像装置出厂时标记的数据而获得。
这样，在计算网络损失时考虑了各个像素的摄像角度的因素，进一步提升了主要被摄体(摄像角度较大)在训练过程中的作用，忽略了次要被摄体(摄像角度较小)在训练过程中的作用，从而能够进一步提高训练的速度以及模型的稳定性。
在步骤105中，还可以在计算该分类神经网络的损失时，针对边界扩展区域中的像素，提高属于特定分类的概率。
例如，该特定分类是用户感兴趣的分类。例如，如果用户比较关注车道的完整性，则可以提高边界扩展区域中的像素被判定为车道的概率。
这样，能够保持用户感兴趣的分类或者重视的分类的完整性。
在本实施例中，该方法还可以包括：将不感兴趣区域的像素的权值设置为小于1，以用于计算该分类神经网络的损失。此时，将特定像素以及不感兴趣区域的像素之外的其他像素的权值设置为1。
这样，通过将用户不感兴趣区域的像素的权值设置为小于1，能够降低不感兴趣区域对于训练结果的影响，进一步加快训练速度。
在步骤106中，根据计算出的损失，调整该分类神经网络的参数。例如，根据损失对于权重参数和偏置参数的偏导数，调整该分类神经网络的各层的权重参数和偏置参数。
图6是本发明实施例1的用于语义分割的分类神经网络的训练方法的另一示意图。如图6所示，该方法包括：
步骤601：确定训练图像的真值图中各个区域之间的边界线；
步骤602：对该边界线进行扩展，获得边界扩展区域；
步骤603：将该边界扩展区域中的特定像素的权值设置为小于1，该特定像素是与该边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；
步骤604：将该训练图像中的该特定像素以外的其他像素的权值设置为1；
步骤605：根据该训练图像中各个像素的权值，计算该分类神经网络的损失；
步骤606：根据计算出的损失，调整该分类神经网络的参数；
步骤607：判断该损失是否小于预设阈值，当判断结果为“是”时，进入步骤608，当判断结果为“否”时，进入步骤609；
步骤608：增加该特定像素的权值，并保持其他像素的权值；
步骤609：判断是否满足收敛条件，当判断结果为“是”时，结束训练，当判断结果为“否”时，进入步骤610；
步骤610：选择新的训练图像。
在步骤607和步骤608中，判断该损失是否小于预设阈值，该预设阈值可以根据实际需要而设置。当判断为该损失小于预设阈值时，增加所述特定像素的权值以重新计算该损失并进行参数的调整。这样，在分类神经网络模型趋于稳定的情况下，逐渐增加特定像素的权值，能够在模型稳定的情况下进一步提高训练结果的准确性。
在步骤608中，增加的权值的数值可以根据实际情况而确定。例如，每次增加原来权值的20％，增加后的权值最大不超过1。
在步骤609中，该收敛条件可以是预先设定的，例如，循环的此次达到了预设值，或者，网络的损失值降低到预设值。
由上述实施例可知，通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
实施例2
本发明实施例还提供一种用于语义分割的分类神经网络的训练装置，其对应于实 施例1的训练方法。图7是本发明实施例2的用于语义分割的分类神经网络的训练装置的一示意图。如图7所示，装置700包括：
确定单元701，其用于确定训练图像的真值图中各个区域之间的边界线；
扩展单元702，其用于对该边界线进行扩展，获得边界扩展区域；
第一设置单元703，其用于将边界扩展区域中的特定像素的权值设置为小于1，该特定像素是与该边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；
第二设置单元704，其用于将训练图像中的特定像素以外的其他像素的权值设置为1；
计算单元705，其用于根据该训练图像中各个像素的权值，计算该分类神经网络的损失；以及
调整单元706，其用于根据计算出的损失，调整分类神经网络的参数。
在本实施例中，该装置700还可以包括：
第三设置单元707，其用于当该分类神经网络的损失小于预设阈值时，增加特定像素的权值，以重新计算损失并根据该损失进行分类神经网络的参数的调整。
在本实施例中，该装置700还可以包括：
第四设置单元708，其用于将不感兴趣区域的像素的权值设置为小于1，以用于计算分类神经网络的损失。
在本实施例中，上述各个单元的功能的实现可以参见实施例1中的训练方法的各个步骤的实施，此处不再赘述。
由上述实施例可知，通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
实施例3
本发明实施例还提供了一种电子设备，图8是本发明实施例3的电子设备的一示意图。如图8所示，电子设备800包括用于语义分割的分类神经网络的训练装置801，其中，用于语义分割的分类神经网络的训练装置801的结构和功能与实施例2中的记载相同，此处不再赘述。
图9是本发明实施例3的电子设备的系统构成的一示意框图。如图9所示，电子 设备900可以包括中央处理器901和存储器902；存储器902耦合到中央处理器901。该图是示例性的；还可以使用其它类型的结构，来补充或代替该结构，以实现电信功能或其它功能。
如图9所示，该电子设备900还可以包括：输入单元903、显示器904、电源905。
在一个实施方式中，实施例1所述的用于语义分割的分类神经网络的训练装置的功能可以被集成到中央处理器901中。其中，中央处理器901可以被配置为：确定训练图像的真值图中各个区域之间的边界线；对所述边界线进行扩展，获得边界扩展区域；将所述边界扩展区域中的特定像素的权值设置为小于1，所述特定像素是与所述边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；将所述训练图像中的所述特定像素以外的其他像素的权值设置为1；根据所述训练图像中各个像素的权值，计算所述分类神经网络的损失；以及根据计算出的所述损失，调整所述分类神经网络的参数。
例如，所述边界扩展区域的相邻区域包括第一相邻区域和第二相邻区域，所述第一相邻区域的分类结果为第一分类，所述第二相邻区域的分类结果为第二分类，所述将所述边界扩展区域中的特定像素的权值设置为小于1，包括：将所述特定像素中的分类结果为第二分类且邻近所述第一相邻区域的像素、以及所述特定像素中的分类结果为第一分类且邻近所述第二相邻区域的像素的权值设置为小于1的第一权值，将所述特定像素中的其他像素的权值设置为小于1的第二权值，所述第一权值大于所述第二权值。
例如，所述根据所述训练图像中所有像素的权值，计算所述分类神经网络的损失，包括：根据所述训练图像中的各个像素的权值，计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
例如，所述根据所述训练图像中所有像素的权值，计算所述分类神经网络的损失，包括：根据所述训练图像中的各个像素的权值以及各个像素的摄像角度，计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
例如，中央处理器901还可以被配置为：当所述分类神经网络的所述损失小于预设阈值时，增加所述特定像素的权值，以重新计算所述损失并根据所述损失进行所述分类神经网络的参数的调整。
例如，在计算所述分类神经网络的损失时，针对所述边界扩展区域中的像素，提 高属于特定分类的概率。
例如，中央处理器901还可以被配置为：将不感兴趣区域的像素的权值设置为小于1，以用于计算所述分类神经网络的损失。
在另一个实施方式中，实施例1所述的用于语义分割的分类神经网络的训练装置可以与中央处理器901分开配置，例如可以将用于语义分割的分类神经网络的训练装置配置为与中央处理器901连接的芯片，通过中央处理器901的控制来实现用于语义分割的分类神经网络的训练装置的功能。
在本实施例中电子设备900也并不是必须要包括图9中所示的所有部件。
如图9所示，中央处理器901有时也称为控制器或操作控件，可以包括微处理器或其它处理器装置和/或逻辑装置，中央处理器901接收输入并控制电子设备900的各个部件的操作。
存储器902，例如可以是缓存器、闪存、硬驱、可移动介质、易失性存储器、非易失性存储器或其它合适装置中的一种或更多种。并且中央处理器901可执行该存储器902存储的该程序，以实现信息存储或处理等。其它部件的功能与现有类似，此处不再赘述。电子设备900的各部件可以通过专用硬件、固件、软件或其结合来实现，而不偏离本发明的范围。
由上述实施例可知，通过在计算神经网络的损失时将边界区域中具有不确定性的像素的权值设置为较小，能够降低这些像素对于训练速度和模型稳定性的影响，从而有效提高训练速度和模型稳定性，并且，能够保证训练图像中主要部分的训练结果的准确性。
本发明实施例还提供一种计算机可读程序，其中当在用于用于语义分割的分类神经网络的训练装置或电子设备中执行所述程序时，所述程序使得计算机在所述用于语义分割的分类神经网络的训练装置或电子设备中执行实施例1所述的用于语义分割的分类神经网络的训练方法。
本发明实施例还提供一种存储有计算机可读程序的存储介质，其中所述计算机可读程序使得计算机在用于语义分割的分类神经网络的训练装置或电子设备中执行实施例1所述的用于语义分割的分类神经网络的训练方法。
结合本发明实施例描述的在用于语义分割的分类神经网络的训练装置中进行训练的方法可直接体现为硬件、由处理器执行的软件模块或二者组合。例如，图7中所 示的功能框图中的一个或多个和/或功能框图的一个或多个组合，既可以对应于计算机程序流程的各个软件模块，亦可以对应于各个硬件模块。这些软件模块，可以分别对应于图1所示的各个步骤。这些硬件模块例如可利用现场可编程门阵列(FPGA)将这些软件模块固化而实现。
软件模块可以位于RAM存储器、闪存、ROM存储器、EPROM存储器、EEPROM存储器、寄存器、硬盘、移动磁盘、CD-ROM或者本领域已知的任何其它形式的存储介质。可以将一种存储介质耦接至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息；或者该存储介质可以是处理器的组成部分。处理器和存储介质可以位于ASIC中。该软件模块可以存储在移动终端的存储器中，也可以存储在可插入移动终端的存储卡中。例如，若设备(例如移动终端)采用的是较大容量的MEGA-SIM卡或者大容量的闪存装置，则该软件模块可存储在该MEGA-SIM卡或者大容量的闪存装置中。
针对图7描述的功能框图中的一个或多个和/或功能框图的一个或多个组合，可以实现为用于执行本申请所描述功能的通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现场可编程门阵列(FPGA)或其它可编程逻辑器件、分立门或晶体管逻辑器件、分立硬件组件、或者其任意适当组合。针对图7描述的功能框图中的一个或多个和/或功能框图的一个或多个组合，还可以实现为计算设备的组合，例如，DSP和微处理器的组合、多个微处理器、与DSP通信结合的一个或多个微处理器或者任何其它这种配置。
以上结合具体的实施方式对本发明进行了描述，但本领域技术人员应该清楚，这些描述都是示例性的，并不是对本发明保护范围的限制。本领域技术人员可以根据本发明的精神和原理对本发明做出各种变型和修改，这些变型和修改也在本发明的范围内。

Claims (15)

  1. 一种用于语义分割的分类神经网络的训练方法，所述方法包括：
    确定训练图像的真值图中各个区域之间的边界线；
    对所述边界线进行扩展，获得边界扩展区域；
    将所述边界扩展区域中的特定像素的权值设置为小于1，所述特定像素是与所述边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；
    将所述训练图像中的所述特定像素以外的其他像素的权值设置为1；
    根据所述训练图像中各个像素的权值，计算所述分类神经网络的损失；以及
    根据计算出的所述损失，调整所述分类神经网络的参数。
  2. 根据权利要求1所述的方法，其中，所述边界扩展区域的相邻区域包括第一相邻区域和第二相邻区域，所述第一相邻区域的分类结果为第一分类，所述第二相邻区域的分类结果为第二分类，
    所述将所述边界扩展区域中的特定像素的权值设置为小于1，包括：
    将所述特定像素中的分类结果为第二分类且邻近所述第一相邻区域的像素、以及所述特定像素中的分类结果为第一分类且邻近所述第二相邻区域的像素的权值设置为小于1的第一权值，将所述特定像素中的其他像素的权值设置为小于1的第二权值，所述第一权值大于所述第二权值。
  3. 根据权利要求1所述的方法，其中，所述根据所述训练图像中所有像素的权值，计算所述分类神经网络的损失，包括：
    根据所述训练图像中的各个像素的权值，计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
  4. 根据权利要求3所述的方法，其中，所述根据所述训练图像中所有像素的权值，计算所述分类神经网络的损失，包括：
    根据所述训练图像中的各个像素的权值以及各个像素的摄像角度，计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
  5. 根据权利要求1所述的方法，其中，所述方法还包括：
    当所述分类神经网络的所述损失小于预设阈值时，增加所述特定像素的权值，以重新计算所述损失并根据所述损失进行所述分类神经网络的参数的调整。
  6. 根据权利要求1所述的方法，其中，
    在计算所述分类神经网络的损失时，针对所述边界扩展区域中的像素，提高属于特定分类的概率。
  7. 根据权利要求1所述的方法，其中，该方法还包括：
    将不感兴趣区域的像素的权值设置为小于1，以用于计算所述分类神经网络的损失。
  8. 一种用于语义分割的分类神经网络的训练装置，所述装置包括：
    确定单元，其用于确定训练图像的真值图中各个区域之间的边界线；
    扩展单元，其用于对所述边界线进行扩展，获得边界扩展区域；
    第一设置单元，其用于将所述边界扩展区域中的特定像素的权值设置为小于1，所述特定像素是与所述边界扩展区域的相邻区域中的至少一个的分类结果相同的像素；
    第二设置单元，其用于将所述训练图像中的所述特定像素以外的其他像素的权值设置为1；
    计算单元，其用于根据所述训练图像中各个像素的权值，计算所述分类神经网络的损失；以及
    调整单元，其用于根据计算出的所述损失，调整所述分类神经网络的参数。
  9. 根据权利要求8所述的装置，其中，所述边界扩展区域的相邻区域包括第一相邻区域和第二相邻区域，所述第一相邻区域的分类结果为第一分类，所述第二相邻区域的分类结果为第二分类，
    所述第一设置单元将所述特定像素中的分类结果为第二分类且邻近所述第一相邻区域的像素、以及所述特定像素中的分类结果为第一分类且邻近所述第二相邻区域的像素的权值设置为小于1的第一权值，将所述特定像素中的其他像素的权值设置为小于1的第二权值，所述第一权值大于所述第二权值。
  10. 根据权利要求8所述的装置，其中，
    所述计算单元根据所述训练图像中的各个像素的权值，计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
  11. 根据权利要求10所述的装置，其中，
    所述计算单元根据所述训练图像中的各个像素的权值以及各个像素的摄像角度， 计算各个像素的交叉熵的累计值，从而获得所述分类神经网络的损失。
  12. 根据权利要求8所述的装置，其中，所述装置还包括：
    第三设置单元，其用于当所述分类神经网络的所述损失小于预设阈值时，增加所述特定像素的权值，以重新计算所述损失并根据所述损失进行所述分类神经网络的参数的调整。
  13. 根据权利要求8所述的装置，其中，
    所述计算单元在计算所述分类神经网络的损失时，针对所述边界扩展区域中的像素，提高属于特定分类的概率。
  14. 根据权利要求8所述的装置，其中，所述装置还包括：
    第四设置单元，其用于将不感兴趣区域的像素的权值设置为小于1，以用于计算所述分类神经网络的损失。
  15. 一种电子设备，所述电子设备包括根据权利要求8-14中的任一项所述的装置。
PCT/CN2018/091542 2018-06-15 2018-06-15 用于语义分割的分类神经网络的训练方法及装置、电子设备 Ceased WO2019237342A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201880092678.3A CN112020721A (zh) 2018-06-15 2018-06-15 用于语义分割的分类神经网络的训练方法及装置、电子设备
PCT/CN2018/091542 WO2019237342A1 (zh) 2018-06-15 2018-06-15 用于语义分割的分类神经网络的训练方法及装置、电子设备

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/091542 WO2019237342A1 (zh) 2018-06-15 2018-06-15 用于语义分割的分类神经网络的训练方法及装置、电子设备

Publications (1)

Publication Number Publication Date
WO2019237342A1 true WO2019237342A1 (zh) 2019-12-19

Family

ID=68841773

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/091542 Ceased WO2019237342A1 (zh) 2018-06-15 2018-06-15 用于语义分割的分类神经网络的训练方法及装置、电子设备

Country Status (2)

Country Link
CN (1) CN112020721A (zh)
WO (1) WO2019237342A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112330640A (zh) * 2020-11-09 2021-02-05 平安科技（深圳）有限公司 医学图像中结节区域的分割方法、装置及设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070258648A1 (en) * 2006-05-05 2007-11-08 Xerox Corporation Generic visual classification with gradient components-based dimensionality enhancement
CN107220980A (zh) * 2017-05-25 2017-09-29 重庆理工大学 一种基于全卷积网络的mri图像脑肿瘤自动分割方法
CN107316307A (zh) * 2017-06-27 2017-11-03 北京工业大学 一种基于深度卷积神经网络的中医舌图像自动分割方法
CN107766794A (zh) * 2017-09-22 2018-03-06 天津大学 一种特征融合系数可学习的图像语义分割方法
CN107945204A (zh) * 2017-10-27 2018-04-20 西安电子科技大学 一种基于生成对抗网络的像素级人像抠图方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107392310A (zh) * 2016-05-16 2017-11-24 北京陌上花科技有限公司 神经网络模型训练方法及装置
US10402690B2 (en) * 2016-11-07 2019-09-03 Nec Corporation System and method for learning random-walk label propagation for weakly-supervised semantic segmentation
CN106709924B (zh) * 2016-11-18 2019-11-22 中国人民解放军信息工程大学 基于深度卷积神经网络和超像素的图像语义分割方法
CN106886801B (zh) * 2017-04-14 2021-12-17 北京图森智途科技有限公司 一种图像语义分割方法及装置
CN107092960A (zh) * 2017-04-17 2017-08-25 中国民航大学 一种改进的并行通道卷积神经网络训练方法
CN108108768B (zh) * 2017-12-29 2020-09-25 清华大学 基于卷积神经网络的光伏玻璃缺陷分类方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070258648A1 (en) * 2006-05-05 2007-11-08 Xerox Corporation Generic visual classification with gradient components-based dimensionality enhancement
CN107220980A (zh) * 2017-05-25 2017-09-29 重庆理工大学 一种基于全卷积网络的mri图像脑肿瘤自动分割方法
CN107316307A (zh) * 2017-06-27 2017-11-03 北京工业大学 一种基于深度卷积神经网络的中医舌图像自动分割方法
CN107766794A (zh) * 2017-09-22 2018-03-06 天津大学 一种特征融合系数可学习的图像语义分割方法
CN107945204A (zh) * 2017-10-27 2018-04-20 西安电子科技大学 一种基于生成对抗网络的像素级人像抠图方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112330640A (zh) * 2020-11-09 2021-02-05 平安科技（深圳）有限公司 医学图像中结节区域的分割方法、装置及设备
CN112330640B (zh) * 2020-11-09 2024-07-16 平安科技（深圳）有限公司 医学图像中结节区域的分割方法、装置及设备

Also Published As

Publication number Publication date
CN112020721A (zh) 2020-12-01

Similar Documents

Publication Publication Date Title
US10019823B2 (en) Combined composition and change-based models for image cropping
US9299004B2 (en) Image foreground detection
US9330334B2 (en) Iterative saliency map estimation
CN109145713B (zh) 一种结合目标检测的小目标语义分割方法
WO2020248957A1 (zh) 遮挡的目标物体的检测方法、系统、设备及存储介质
WO2020098606A1 (zh) 节点分类方法、模型训练方法、装置、设备及存储介质
CN113505261A (zh) 数据标注方法、装置以及数据标注模型训练方法、装置
CN113947766A (zh) 一种基于卷积神经网络的实时车牌检测方法
CN110189256A (zh) 一种全景图像拼接方法、计算机可读存储介质及全景相机
WO2020107326A1 (zh) 车道线检测方法、设备、计算机可读存储介质
CN113129335B (zh) 一种基于孪生网络的视觉跟踪算法及多模板更新策略
CN106845621A (zh) 基于深度卷积神经网络的密集人群人数计数方法及系统
CN114897684B (zh) 车辆图像的拼接方法、装置、计算机设备和存储介质
WO2022012425A1 (zh) 目标检测方法、装置及电子设备
CN111967313A (zh) 一种深度学习目标检测算法辅助的无人机图像标注方法
CN112257703B (zh) 图像识别方法、装置、设备和可读存储介质
CN116258722B (zh) 基于图像处理的桥梁建筑智能检测方法
CN109961013A (zh) 车道线的识别方法、装置、设备及计算机可读存储介质
CN110969145B (zh) 一种遥感图像匹配优化方法、装置、电子设备及存储介质
CN110766711A (zh) 一种视频镜头分割方法、系统、装置和存储介质
TWI497449B (zh) 非監督式調適方法與應用其之影像自動分類方法
CN110547803A (zh) 一种适用于鱼眼相机俯视拍摄的行人身高估计方法
WO2020087322A1 (zh) 车道线识别方法和装置、车辆
US20200288066A1 (en) Delivery of notifications for feedback over visual quality of images
CN111582296B (zh) 一种遥感图像综合匹配方法、装置、电子设备及存储介质

Legal Events

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

Ref document number: 18922433

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: 18922433

Country of ref document: EP

Kind code of ref document: A1