WO2021093756A1 - 基于神经网络的目标背景噪声抑制方法及设备 - Google Patents

基于神经网络的目标背景噪声抑制方法及设备 Download PDF

Info

Publication number
WO2021093756A1
WO2021093756A1 PCT/CN2020/128021 CN2020128021W WO2021093756A1 WO 2021093756 A1 WO2021093756 A1 WO 2021093756A1 CN 2020128021 W CN2020128021 W CN 2020128021W WO 2021093756 A1 WO2021093756 A1 WO 2021093756A1
Authority
WO
WIPO (PCT)
Prior art keywords
segmented
circumscribed
target
frame
targets
Prior art date
Application number
PCT/CN2020/128021
Other languages
English (en)
French (fr)
Inventor
欧阳瑶
周治尹
Original Assignee
中科智云科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中科智云科技有限公司 filed Critical 中科智云科技有限公司
Publication of WO2021093756A1 publication Critical patent/WO2021093756A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G06T5/70—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Definitions

  • This application relates to the computer field, and in particular to a method and equipment for suppressing target background noise based on neural networks.
  • this segmentation method has the following problems: when the bounding rectangle contains a large amount of background noise, the interfering target and the detected target are closely connected to each other, and the complete and accurate boundary of the target object cannot be clearly calibrated.
  • Figure 3 shows the result of segmentation of the target to be segmented in one of the circumscribed rectangles. It can be seen from Figure 3 that there are multiple closely connected unrelated interfering targets in the circumscribed rectangle, and the boundary of the target to be segmented in the circumscribed rectangle is submerged by a lot of noise , There is a problem that the target cannot be located directly and accurately. At present, there is still a lack of an effective method to suppress the dense interference target noise in the circumscribed rectangular frame.
  • One purpose of this application is to provide a method and equipment for suppressing background noise of dense targets based on neural networks.
  • a method for suppressing background noise of a dense target based on a neural network includes:
  • the target to be divided is output.
  • the above-mentioned multiple densely distributed targets to be segmented are respectively subjected to circumscribed frame segmentation to obtain the corresponding multiple densely distributed targets to be segmented.
  • the circumscribed frame includes: circumscribed rectangular frames are respectively performed on the plurality of densely distributed objects to be divided by a circumscribed rectangular segmentation method to obtain the circumscribed rectangular frames corresponding to the plurality of densely distributed objects to be divided.
  • the circumscribed frames are numbered, and an intersection is established for each of the circumscribed frames A list of frames, and record the intersection area between the circumscribed frame of one of the objects to be divided and the circumscribed frames of other objects to be divided.
  • the above-mentioned circumscribed frame corresponding to the plurality of densely distributed targets to be segmented is performed to perform the hierarchical sorting process of the targets to be segmented, including :
  • area i represents the area covered by any one of the circumscribed frames
  • area cover represents the area of the circumscribed frame covered by other interfering circumscribed frames
  • the threshold for defining the outer frame to be covered is T, and if the covered area ratio P i is less than the threshold T, then the outer frame that meets the threshold requirement is placed in parallel for processing at the same level.
  • the above-mentioned outer frame meeting the threshold requirement is removed from the remaining Delete the intersecting box list of the external frame, and obtain the intersecting frame list of the new external frame;
  • the targets of each layer are processed sequentially from top to bottom according to the hierarchical sorting graph.
  • the targets of each layer are processed sequentially from top to bottom according to the hierarchical sorting graph, including:
  • the target rotation angle detection method is used to calculate the rotation angle of each target to be divided in the same layer.
  • the result of the rotation angle of each target to be segmented in the same layer is calculated separately, and the circumscribed rectangle is again performed. Segmentation to obtain the precise position of each target to be segmented on the same layer.
  • each segment to be segmented in the same layer that has been segmented is After the target is removed, all to-be-segmented targets of the next layer in the hierarchical sorting diagram are segmented until all the to-be-segmented targets of all layers are segmented.
  • a computer-readable medium on which computer-readable instructions are stored.
  • the processor can realize the method described in any one of the above items. The method described.
  • a dense target background noise suppression device based on a neural network, the device including:
  • a memory and a processor where the memory stores a computer program, wherein the computer program implements any of the above-mentioned methods when the computer program is executed by the processor.
  • Figure 1 shows a schematic diagram of multiple densely distributed targets in the prior art
  • FIG. 2 shows a schematic diagram of the structure shown in FIG. 1 after the circumscribed rectangular frame is divided;
  • FIG. 3 shows a partial enlarged view of one of the rectangular circumscribed frames segmented based on the segmentation method shown in FIG. 2;
  • Fig. 4 shows a flowchart of one aspect of a method for suppressing background noise of dense targets based on a neural network according to the present application
  • FIG. 5 shows a block diagram of a circumscribed rectangle obtained by using a circumscribed rectangle segmentation method in one aspect of the method for suppressing background noise of dense targets based on a neural network of the present application;
  • FIG. 6 shows a schematic diagram of a hierarchical sorting processing structure of one aspect of a method for suppressing background noise of dense targets based on a neural network in the present application
  • FIG. 7 shows a schematic diagram of a circumscribed rectangular frame with A as the target to be divided according to an aspect of the present application
  • FIG. 8 shows a detection result diagram of the target rotation angle using A as the target to be segmented in one aspect of the present application
  • FIG. 9 shows a schematic diagram of rotating to a horizontal direction with A as the target to be divided in an aspect of the present application.
  • FIG. 10 shows a schematic diagram after the circumscribed rectangle is divided based on the structure shown in FIG. 9;
  • FIG. 11 shows a schematic diagram of cropping of a circumscribed rectangular frame with A as the target to be divided according to an aspect of the present application
  • FIG. 12 shows an accurate positioning and segmentation diagram of the target A to be segmented in one aspect of the present application
  • FIG. 13 shows an accurate positioning and segmentation diagram of the target to be segmented in the first layer according to an aspect of the present application
  • FIG. 14 shows the precise positioning and segmentation diagram of the target to be segmented obtained by the method for suppressing the background noise of the dense target based on the neural network of the present application.
  • the terminal, the equipment serving the network, and the trusted party all include one or more processors, such as a central processing unit (CPU), input/output interface, network interface, and memory.
  • processors such as a central processing unit (CPU), input/output interface, network interface, and memory.
  • Memory includes non-permanent memory in computer-readable media, random access memory (RAM) and/or non-volatile memory, such as read only memory (ROM) or flash memory (flash). RAM). Memory is an example of computer readable media.
  • RAM random access memory
  • ROM read only memory
  • flash flash memory
  • Computer-readable media include permanent and non-permanent, removable and non-removable media, and information storage can be realized by any method or technology.
  • the information can be computer-readable instructions, data structures, program modules, or other data.
  • Examples of computer memory include, but are not limited to, Phase-Change RAM (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), flash memory or other memory technologies, read-only Compact Disc Read-Only Memory (CD-ROM), Digital Versatile Disk (DVD) or other optical storage, magnetic cassette tape, magnetic tape disk storage or other magnetic storage devices or any other non- Transmission media can be used to store information that can be accessed by computing devices.
  • PRAM Phase-Change RAM
  • SRAM Static Random Access Memory
  • DRAM Dynamic Random Access Memory
  • RAM random access memory
  • a method for suppressing background noise of dense targets based on a neural network includes:
  • Step S11 acquiring an image and determining multiple densely distributed targets to be segmented
  • Step S12 performing circumscribed frame segmentation on the plurality of densely distributed objects to be divided, respectively, to obtain circumscribed frames corresponding to the plurality of densely distributed objects to be divided;
  • Step S13 based on the circumscribed frames corresponding to the multiple densely distributed objects to be segmented, perform a hierarchical sorting process of the objects to be segmented;
  • Step S14 based on the result of the above-mentioned hierarchical sorting of the target to be segmented, detecting the rotation angle of the target to be segmented;
  • Step S15 based on the result of detecting the rotation angle of the target to be segmented, perform circumscribed frame segmentation on the target to be segmented;
  • Step S16 based on the result of the circumscribed frame segmentation of the target to be segmented, output the target to be segmented.
  • this embodiment when the circumscribed frame contains a plurality of densely distributed objects to be segmented, this embodiment can effectively suppress background noise, and can clearly demarcate the complete and accurate boundaries of the objects to be segmented.
  • FIGS. 5 to 10 show schematic diagrams of the dense target background noise suppression process involved in the dense target background noise suppression method based on the neural network of this embodiment.
  • the structural examples (such as the structure of different bottles or food packaging bags) of the target A to be divided, the target B to be divided, the target C to be divided, the target to be divided and the target E to be divided (such as the structure of different bottles or food packaging bags) are only schematic and The explanation does not limit the scope of protection of this application; of course, the above-mentioned illustration can also use other existing common product structures as examples.
  • the determination of the target A to be segmented, the target to be segmented B, the target to be segmented C, the target to be segmented D, and the target to be segmented E only indicate one of the dense distributions.
  • the densely distributed arrangement does not limit the scope of protection of this application.
  • this embodiment only illustrates that there are only 5 cases when the target to be segmented is densely distributed.
  • the specific set number of the target to be segmented may be at least two, and the target to be segmented may be at least two.
  • the quantitative disclosure does not limit the scope of protection of this application.
  • the target to be segmented A, the target to be segmented B, the target to be segmented C, the target to be segmented D, and the target to be segmented E are preferably displayed in a structure similar to the outer packaging of food, and the circumscribed frame involved is preferably adopted An example is illustrated by circumscribing a rectangular frame, as shown in FIG. 5, to further understand this embodiment clearly.
  • step S12 the above-mentioned circumscribed frame segmentation of the plurality of densely distributed objects to be divided respectively to obtain the circumscribed frame corresponding to the plurality of densely distributed objects to be divided includes: through the circumscribed rectangle
  • the segmentation method respectively performs circumscribed rectangular frame segmentation on the plurality of densely distributed objects to be divided to obtain circumscribed rectangular frames corresponding to the multiple densely distributed objects to be divided.
  • step S12 the circumscribed rectangular frame is numbered based on the target to be divided and the circumscribed rectangular frame corresponding to the target.
  • each of the above-mentioned circumscribed rectangular boxes is numbered, which are respectively circumscribed rectangle A', circumscribed rectangle B’, circumscribed rectangle C’, circumscribed rectangle D’, circumscribed rectangle E’, in order to further combine different
  • the circumscribed rectangular frame is better or more clearly distinguished, and the above circumscribed rectangular frame can also be marked in different colors while numbering, such as circumscribed rectangle A'(yellow), circumscribed rectangle B'(green), circumscribed rectangle Frame C'(blue), circumscribed rectangular frame D'(red), and circumscribed rectangular frame E'(purple).
  • the above-mentioned labels are in different colors, which are merely illustrative and do not limit the protection scope of the present application.
  • the circumscribed frames are numbered, and a list of intersecting frames is established for each circumscribed frame, and one of the circumscribed frames of the target to be divided is recorded The intersection area between the circumscribed frame and the circumscribed frame of other targets to be divided.
  • circumscribed rectangle A' circumscribed rectangle B’, circumscribed rectangle C’, circumscribed rectangle D’ and circumscribed rectangle E’
  • the intersecting box list established, as shown in Figure 5 and circumscribed rectangle A 'Intersecting areas are: circumscribed rectangular frame B', circumscribed rectangular frame C'and circumscribed rectangular frame D', expressed as A':[B',C',D'].
  • the remaining targets to be segmented are expressed as B':[A',C',D',E'], C':[A',B',D'], D':[A',B ',C',E'] and E':[B',D'].
  • This embodiment only exemplifies one of the densely distributed multiple targets to be segmented.
  • step S13 the above-mentioned circumscribed frame corresponding to the plurality of densely distributed targets to be segmented is performed to perform the hierarchical sorting process of the targets to be segmented, including:
  • area i represents the area covered by any one of the circumscribed frames
  • area cover represents the area of the circumscribed frame covered by other interfering circumscribed frames
  • the threshold for defining the outer frame to be covered is T, and if the covered area ratio P i is less than the threshold T, then the outer frame that meets the threshold requirement is placed in parallel for processing at the same level.
  • the area ratio P of the circumscribed rectangle A', circumscribed rectangle B’, circumscribed rectangle C’, circumscribed rectangle D’, and circumscribed rectangle E’ respectively covered by other circumscribed rectangles is P i is compared with the above threshold T.
  • the circumscribed rectangle A', the circumscribed rectangle B’, and the circumscribed rectangle E’ meet the requirements, then they will be compared with the circumscribed rectangle A’, circumscribed rectangle B’, and circumscribed rectangle respectively.
  • the target A to be segmented, the target B to be segmented, and the target E to be segmented corresponding to the rectangular frame E′ are treated as the first layer of optimal processing.
  • step S13 based on the above-mentioned processing result of juxtaposing the circumscribed frame that meets the threshold requirement at the same level, the circumscribed frame that meets the threshold requirement is deleted from the intersecting frame list of the remaining circumscribed frames , Get the intersecting frame list of the new external frame;
  • circumscribed rectangular frame A' since the above-mentioned circumscribed rectangular frame A', circumscribed rectangular frame B'and circumscribed rectangular frame E'meet the requirements, they serve as the target circumscribed rectangular frame for the parallel processing of the first layer. Then, delete the circumscribed rectangle A’, circumscribed rectangle B’, and circumscribed rectangle E’ in the intersection box list of the circumscribed rectangle C', circumscribed rectangle D’, and get a new intersection box list C’:[D '], D':[C'].
  • this embodiment only illustrates the case where only the first and second layers are sorted according to the hierarchical sorting diagram. If there are more objects to be segmented Or multiple targets to be segmented arranged in other densely distributed manners, then the third layer, the fourth layer...the Nth layer may also appear. In this case, each object will be processed in order from top to bottom according to the hierarchical sorting graph. Targets in one layer until all targets to be segmented are segmented.
  • step S14 based on the above-mentioned hierarchical sorting processing result of the target to be segmented, the targets of each layer are sequentially processed from top to bottom according to the hierarchical sorting graph, including: adopting a target rotation angle detection method, respectively Calculate the rotation angle of each target to be divided in the same layer.
  • the targets of each layer are processed from top to bottom according to the hierarchical sorting graph.
  • the target rotation angle detection method is used to detect the rotation angle of the target A to be segmented.
  • the result is shown in Figure 8.
  • the included angle ⁇ A where ⁇ A is the rotation angle of the target A to be divided; in the same way, the rotation angle ⁇ B of the target B to be divided and the rotation angle ⁇ E of the target E to be divided are sequentially calculated.
  • step S15 based on the above-mentioned target rotation angle detection method, the rotation angle results of each target to be segmented in the same layer are calculated respectively, and combined with FIG. 5, the circumscribed rectangle segmentation is performed again to obtain The precise location of each target to be segmented on the same layer.
  • the circumscribed rectangle division is performed again to obtain the precise position of the target B to be divided; based on the rotation angle ⁇ of the target E to be divided E and Figure 5, the circumscribed rectangle segmentation is performed again, and the precise position of the target E to be segmented can be obtained.
  • step S15 based on the result of obtaining the precise position of each target to be segmented in the same layer, after removing each target to be segmented in the same layer that has been segmented, the hierarchical sorting map All to-be-segmented targets in the next layer are segmented until all the to-be-segmented targets in all layers are segmented.
  • the target A to be segmented, the target B to be segmented, and the target E to be segmented in the first layer are processed, and the final result obtained is as shown in FIG. 13.
  • the level All to-be-segmented targets in the second layer in the sorting diagram are segmented, that is, the to-be-segmented target C and the to-be-segmented target D in the second layer in this embodiment are segmented.
  • the specific segmentation steps of the target C to be segmented and the target D to be segmented are as follows:
  • the circumscribed rectangle segmentation is performed again, and the precise position of the target C to be segmented can be finally obtained.
  • the segmentation process please refer to the specific segmentation process of the target A to be segmented;
  • the rotation angle ⁇ D of the target D to be divided is as shown in Fig. 5, and the circumscribed rectangle is divided again, and the precise position of the target D to be divided can be finally obtained.
  • the division process please refer to the specific division process of the target A to be divided.
  • a computer-readable medium on which computer-readable instructions are stored.
  • the processor realizes the above-mentioned neural network-based A method of suppressing the background noise of dense targets.
  • a dense target background noise suppression device based on a neural network including:
  • One or more processors are One or more processors;
  • a computer-readable medium for storing one or more computer-readable instructions
  • the one or more processors When the one or more computer-readable instructions are executed by the one or more processors, the one or more processors implement the above-mentioned neural network-based dense target background noise suppression method.
  • this application when the circumscribed frame contains background targets, this application can effectively suppress the noise formed by the background targets; when the circumscribed frame contains densely distributed background targets, this application can accurately locate the boundary of the target to be segmented And the target position; when the background target and the target to be segmented are closely connected to each other, even if the external frame of this application contains a large amount of background noise, this application can still clearly demarcate the complete and accurate boundary of the target to be segmented.
  • this application can be implemented in software and/or a combination of software and hardware.
  • it can be implemented using an application specific integrated circuit (ASIC), a general purpose computer or any other similar hardware device.
  • ASIC application specific integrated circuit
  • the software program of the present application may be executed by a processor to realize the steps or functions described above.
  • the software program (including related data structure) of the present application can be stored in a computer-readable recording medium, such as RAM memory, magnetic or optical drive, or floppy disk and similar devices.
  • some steps or functions of the present application may be implemented by hardware, for example, as a circuit that cooperates with a processor to execute each step or function.
  • a part of this application can be applied as a computer program product, such as a computer program instruction, when it is executed by a computer, through the operation of the computer, the method and/or technical solution according to this application can be invoked or provided.
  • the program instructions for calling the method of the present application may be stored in a fixed or removable recording medium, and/or be transmitted through a data stream in a broadcast or other signal-bearing medium, and/or be stored in accordance with the Said program instructions run in the working memory of the computer equipment.
  • an embodiment according to the present application includes a device that includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device triggers
  • the operation of the device is based on the aforementioned methods and/or technical solutions according to multiple embodiments of the present application.

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

一种基于神经网络的稠密目标背景噪声抑制方法及设备，该方法包括：获取图像并确定多个稠密分布的待分割目标；分别对上述待分割目标进行外接框分割，获得上述待分割目标所对应的外接框；进行所述待分割目标的层级排序处理；对待分割目标的旋转角度进行检测；对所述待分割目标进行外接框分割；输出待分割目标。计算机可读介质，其上存储有计算机可读指令，计算机可读指令可被处理器执行时，使处理器实现如上述的方法。该设备包括：处理器；计算机可读介质，用于存储计算机可读指令，当计算机可读指令被处理器执行，使得处理器实现上述的方法。

Description

基于神经网络的目标背景噪声抑制方法及设备 技术领域
本申请涉及计算机领域，尤其涉及基于神经网络的目标背景噪声抑制方法及设备。
背景技术
当使用一种基于神经网络的分割方法对如图1所示的稠密分布的多个目标用外接矩形进行分割，得到的结果如图2所示。但是，该分割方法存在以下问题：当外接矩形框中包含了大量的背景噪声，干扰目标与被检测目标相互紧密连接，无法清晰标定目标物体的完整准确边界。
当图像中包含多个待分割目标时，目标间相互干扰程度不同，存在无法直接精确定位分割目标物体的现象。图3为其中一个外接矩形框中的待分割目标分割的结果，从图3可知，该外接矩形框中存在多个紧密相连无关干扰目标，该外接矩形框中的待分割目标边界被大量噪声淹没，存在无法直接精确定位目标的问题。目前尚缺少一种对外接矩形框内稠密干扰目标噪声的有效抑制方法。
发明内容
本申请的一个目的是提供基于神经网络的稠密目标背景噪声抑制方法及设备。
根据本申请的一个方面，提供了基于神经网络的稠密目标背景噪声抑制方法，所述方法包括：
获取图像并确定多个稠密分布的待分割目标；
分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框；
基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理；
基于上述待分割目标的层级排序处理的结果，对待分割目标的旋转角度进行检测；
基于上述对待分割目标的旋转角度进行检测的结果，对所述待分割目标进行外接框分割；
基于待分割目标进行外接框分割结果，输出待分割目标。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，上述的分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框包括：通过外接矩形分割方法分别对上述多个稠密分布的待分割目标进行外接矩形框分割，获得所述多个稠密分布的待分割目标所对应的外接矩形框。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述多个待分割目标的外接框，对所述外接框进行编号，并为每个所述外接框建立一个相交框的列表，并记录其中一个所述待分割目标的外接框与其他待分割目标的外接框的相交区域。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，上述的基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理，包括：
基于每个上述外接框所建立一个相交框的列表，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，其中，
P i＝area cover/area i
其中，area i表示其中任意一个外接框所覆盖的面积，area cover表示该外接框被其他干扰的外接框所覆盖的面积；
其中，定义外接框被覆盖的阈值为T，如果所覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述的将符合该阈值要求的所述外接框并列在同一层级处理结果，将上述符合阈值要求的外接框从余下外接框的相交框列表中删除，获得新的外接框的相交框列表；
基于上述获得的新的外接框的相交框列表结果，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，并将该覆盖的面积比P i与阈值T比较，如果该覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理；
重复上述步骤，直至完成所有的待分割目标的层级排序处理。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述的待分割目标的层级排序处理结果，按照层级排序图从上到下依次处理每一层的目标。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述的待分割目标的层级排序处理结果，按照层级排序图从上到下依次处理每一层的目标，包括：采用目标旋转角度检测方法，分别计算出同一层的每个待分割目标的旋转角度。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述的采用目标旋转角度检测方法，分别计算出同一层的每个待分割目 标的旋转角度结果，再次进行外接矩形分割，获得同一层的每个待分割目标的精确位置。
在一实施例中，上述的基于神经网络的稠密目标背景噪声抑制方法中，基于上述的获得同一层的每个待分割目标的精确位置结果，将上述已经完成分割的同一层的每个待分割目标去除后，对层级排序图中的下一层的所有待分割目标进行分割，直至所有层的待分割目标均被分割完成。
根据本申请的另一方面，还提供了计算机可读介质，其上存储有计算机可读指令，所述计算机可读指令可被处理器执行时，使所述处理器实现如上述任一项所述的方法。根据本申请的另一方面，还提供了基于神经网络的稠密目标背景噪声抑制设备，该设备包括：
存储器和处理器，所述存储器存储计算机程序，其特征在于，所述计算机程序被处理器执行时实现上述任一项所述的方法。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本申请的其它特征、目的和优点将会变得更明显：
图1示出现有技术中多个稠密分布的目标的示意图；
图2示出将如图1所示结构进行外接矩形框分割后的示意图；
图3示出基于如图2所示的分割方法所分割出的其中一个矩形外接框的局部放大图；
图4示出根据本申请基于神经网络的稠密目标背景噪声抑制方法的一个方面的流程图；
图5示出本申请基于神经网络的稠密目标背景噪声抑制方法的一个方面的采用外接矩形分割方法获得的外接矩形框图；
图6示出本申请中基于神经网络的稠密目标背景噪声抑制方法的一个方面的层级排序处理结构示意图；
图7示出本申请中一个方面的以A为待分割目标的外接矩形框示意图；
图8示出本申请中一个方面的以A为待分割目标采用目标旋转角度检测结果图；
图9示出本申请中一个方面的以A为待分割目标旋转至水平方向的示意图；
图10示出基于如图9所示结构进行外接矩形分割后的示意图；
图11示出本申请中一个方面的以A为待分割目标的外接矩形框的裁剪示意图；
图12示出本申请中一个方面的待分割目标A的精确定位分割图；
图13示出本申请中一个方面的第一层待分割目标的精确定位分割图；
图14示出本申请基于神经网络的稠密目标背景噪声抑制方法获得的待分割目标的精确定位分割图。
附图中相同或相似的附图标记代表相同或相似的部件。
具体实施方式
下面结合附图对本申请作进一步详细描述。
在本申请一个典型的配置中，终端、服务网络的设备和可信方均包括一个或多个处理器，例如中央处理器(Central Processing Unit，CPU)、输入/输出接口、网络接口和存储器。
存储器包括计算机可读介质中的非永久性存储器，随机存取存储器(Random Access Memory，RAM)和/或非易失性内存等形式，如只读存储器(Read Only Memory，ROM)或闪存(flash RAM)。内存是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储器的例子包括，但不限于相变内存(Phase-Change RAM，PRAM)、静态随机存取存储器(Static Random Access Memory，SRAM)、动态随机存取存储器(Dynamic Random Access Memory，DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(Electrically Erasable Programmable Read-Only Memory，EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(Compact Disc Read-Only Memory，CD-ROM)、数字多功能光盘(Digital Versatile Disk，DVD)或其他光学存储、磁盒式磁带，磁带磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。
如图4所示，根据本实施例的一个方面，提供了基于神经网络的稠密目标背景噪声抑制方法，所述方法包括：
步骤S11，获取图像并确定多个稠密分布的待分割目标；
步骤S12，分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框；
步骤S13，基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理；
步骤S14，基于上述待分割目标的层级排序处理的结果，对待分割目标的旋转角度进行检测；
步骤S15，基于上述对待分割目标的旋转角度进行检测的结果，对所述待分割目标进行外接框分割；
步骤S16，基于待分割目标进行外接框分割结果，输出待分割目标。
上述步骤S11至步骤S16，当外接框中包含多个稠密分布的待分割目 标时，本实施例可以有效抑制背景噪声，还可以清晰地标定待分割目标的完整准确边界。
如图5至图10示出了本实施例基于神经网络的稠密目标背景噪声抑制方法中所涉及的稠密目标背景噪声抑制过程示意图。但是，附图中所涉及的待分割目标A、待分割目标B、待分割目标C、待分割目标D以及待分割目标E的结构示例(如不同瓶子或者食品包装袋的结构)仅仅是示意和解释说明，其并不对本申请的保护范围造成限定；当然，上述示意亦可有采用其他现有的常见的产品结构进行举例说明。
在一实施例中，待分割目标A、待分割目标B、待分割目标C、待分割目标D以及待分割目标E的确定仅仅示意了其中一种稠密分布情况进行了列举和说明，上述具体的稠密分布排列情况并不对本申请的保护范围造成限定。
再者，本实施例仅仅示意了与所述待分割目标稠密分布时仅仅设有5个情况，但在实际应用过程中，上述待分割目标的具体设置数量可以为至少两个，上述对待分割目标数量的公开并不对本申请的保护范围造成限定。
当然，在本实施例中，待分割目标A、待分割目标B、待分割目标C、待分割目标D以及待分割目标E优选采用类似食品外包装的结构进行显示，所涉及的外接框优选采用外接矩形框进行实例说明，如图5所示，以进一步清楚地了解本实施例。
接着本申请的上述实施例，步骤S12中，上述的分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框包括：通过外接矩形分割方法分别对上述多个稠密分布的待分割目标进行外接矩形框分割，获得所述多个稠密分布的待分割目标所对应的 外接矩形框。
接着本申请的上述实施例，步骤S12中，基于所述待分割目标以及与其所对应的外接矩形框，对所述外接矩形框进行编号。
例如，对每个上述的外接矩形框进行编号，分别为外接矩形框A’，外接矩形框B’，外接矩形框C’，外接矩形框D’，外接矩形框E’，为进一步将不同的外接矩形框更好或者更清楚地区分开来，亦可将上述外接矩形框在编号的同时标注为不同的颜色，如外接矩形框A’(黄)，外接矩形框B’(绿)，外接矩形框C’(蓝)，外接矩形框D’(红)以及外接矩形框E’(紫)。其中，上述的标注为不同的颜色，仅仅为列举示意并不对本申请的保护范围进行限定。
在一实施例中，基于上述多个待分割目标的外接框，对所述外接框进行编号，并为每个所述外接框建立一个相交框的列表，并记录其中一个所述待分割目标的外接框与其他待分割目标的外接框的相交区域。
在上述的外接矩形框A’，外接矩形框B’，外接矩形框C’，外接矩形框D’以及外接矩形框E’所建立的相交框列表，如图5所示，与外接矩形框A’相交的区域有：外接矩形框B’，外接矩形框C’以及外接矩形框D’，表示为A’:[B’,C’,D’]。那么依次类推，其余待分割目标依次表示为B’:[A’,C’,D’,E’]，C’:[A’,B’,D’]，D’:[A’,B’,C’,E’]以及E’:[B’,D’]。本实施例仅仅对其中一种稠密分布的多个待分割目标进行举例说明。
接着本申请的上述实施例，步骤S13中，上述的基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理，包括：
基于每个上述外接框所建立一个相交框的列表，计算每个所述外接框被 其他干扰的外接框所覆盖的面积比P i，其中，
P i＝area cover/area i
其中，area i表示其中任意一个外接框所覆盖的面积，area cover表示该外接框被其他干扰的外接框所覆盖的面积；
其中，定义外接框被覆盖的阈值为T，如果所覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理。
那么，按照上述公式，分别对外接矩形框A’，外接矩形框B’，外接矩形框C’，外接矩形框D’以及外接矩形框E’各自被其他的外接矩形框所覆盖的面积比P i和上述的阈值T进行比较，经计算，外接矩形框A’，外接矩形框B’以及外接矩形框E’符合要求，那么，分别将与外接矩形框A’，外接矩形框B’以及外接矩形框E’对应的待分割目标A、待分割目标B以及待分割目标E作为第一层优选处理。
接着本申请的上述实施例，步骤S13中，基于上述的将符合该阈值要求的所述外接框并列在同一层级处理结果，将上述符合阈值要求的外接框从余下外接框的相交框列表中删除，获得新的外接框的相交框列表；
基于上述获得的新的外接框的相交框列表结果，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，并将该覆盖的面积比P i与阈值T比较，如果该覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理；
重复上述步骤，直至完成所有的待分割目标的层级排序处理。
在本实施例中，如图6所示，由于上述的外接矩形框A’，外接矩形框B’以及外接矩形框E’符合要求，作为第一层并行处理的目标外接矩形框。那么，将外接矩形框C’、外接矩形框D’的相交框列表中的外接矩形框A’， 外接矩形框B’以及外接矩形框E’删除，得到新的相交框列表C’:[D’]，D’:[C’]。
然后基于该新的相交框列表，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，并将该覆盖的面积比P i与阈值T比较，如果该覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理。经计算，外接矩形框C’和外接矩形框D’符合要求，作为第二层并行处理的目标外接矩形框，见图6所示。即，分别与外接矩形框C’、外接矩形框D’对应的待分割目标C和待分割目标D作为第二层进行处理。
基于上述，本实施例中的所有待分割目标均已被分割完成，本实施例仅仅示意了仅仅按照层级排序图排序为第一层和第二层的情况，如果还有更多的待分割目标或者按照其他稠密分布方式排列的多个待分割目标，那么，可能还会出现第三层、第四层……第N层的情况，这种情况则按照层级排序图从上到下依次处理每一层的目标，直至所有待分割目标被分割完成。
接着本申请的上述实施例，步骤S14中，基于上述的待分割目标的层级排序处理结果，按照层级排序图从上到下依次处理每一层的目标，包括：采用目标旋转角度检测方法，分别计算出同一层的每个待分割目标的旋转角度。
在本实施例中，按照层级排序图从上到下处理每一层的目标。
具体的方法如下所示：
以第一层的待分割目标A为例，如图7所示，采用目标旋转角度检测方法对待分割目标A的旋转角度进行检测，结果如图8所示，图中的线段G与水平方向的夹角θ A，其中，θ A即为待分割目标A的旋转角度；同理，依次计算待分割目标B的旋转角度θ B和待分割目标E的旋转角度θ E。
接着本申请的上述实施例，步骤S15中，基于上述的采用目标旋转角度检测方法，分别计算出同一层的每个待分割目标的旋转角度结果，并结合图5，再次进行外接矩形分割，获得同一层的每个待分割目标的精确位置。
具体步骤如下所示：
根据旋转角度θ A旋转待分割目标A至水平方向，旋转正以后的外接框如下图9所示，将图9所示的结构进行外接矩形框的分割，其分割结果如图10所示，其中，图中的矩形外接框为真实的目标框的位置。然后将图10中的矩形外接框区域还原至图11中，得到的目标精确定位区域如图12所示，其中的多边形区域为待分割的目标区域，也就是待分割目标A的精确分割位置。
在本实施例中，基于待分割目标A的旋转角度θ A和图5，再次进行外接矩形分割，最终获得的待分割目标A的精确位置如图12所示。
在本实施例中，按照上述方法，基于待分割目标B的旋转角度θ B和图5，再次进行外接矩形分割，可获得的待分割目标B的精确位置；基于待分割目标E的旋转角度θ E和图5，再次进行外接矩形分割，可获得的待分割目标E的精确位置。
接着本申请的上述实施例，步骤S15中，基于上述的获得同一层的每个待分割目标的精确位置结果，将上述已经完成分割的同一层的每个待分割目标去除后，对层级排序图中的下一层的所有待分割目标进行分割，直至所有层的待分割目标均被分割完成。
在本实施例中，首先将处于第一层的待分割目标A、待分割目标B以及待分割目标E处理完，最终获得的结果如图13所示。
在将处于第一层的待分割目标A、待分割目标B以及待分割目标E处理后，并将其第一层的待分割目标A、待分割目标B以及待分割目标E去除后， 对层级排序图中的第二层的所有待分割目标进行分割，即，对本实施例中的处于第二层的待分割目标C和待分割目标D进行分割。
本实施例中，待分割目标C和待分割目标D的具体分割步骤为：
采用目标旋转角度检测方法对待分割目标C的旋转角度进行检测，并计算出待分割目标C的旋转角度θ c；同理，依次计算待分割目标D的旋转角度θ D；
基于待分割目标C的旋转角度θ C和图5，再次进行外接矩形分割，最终可获得的待分割目标C的精确位置，该分割过程可参见上述待分割目标A的具体分割过程；同理，待分割目标D的旋转角度θ D和图5，再次进行外接矩形分割，最终可获得的待分割目标D的精确位置，该分割过程可参见上述待分割目标A的具体分割过程。
上述处于第一层和第二层的待分割目标分割完成后，本实施例中的所有待分割目标均被处理完，最终的处理结果如图14所述。
根据本申请的另一方面，还提供了计算机可读介质，其上存储有计算机可读指令，所述计算机可读指令可被处理器执行时，使所述处理器实现如上述的基于神经网络的稠密目标背景噪声抑制方法。
根据本申请的另一方面，还提供了基于神经网络的稠密目标背景噪声抑制设备，该设备包括：
一个或多个处理器；
计算机可读介质，用于存储一个或多个计算机可读指令，
当所述一个或多个计算机可读指令被所述一个或多个处理器执行，使得所述一个或多个处理器实现如上述的基于神经网络的稠密目标背景噪声抑制方法。
在此，所述设备的各实施例的详细内容，具体可参见上述设备端的基于神经网络的稠密目标背景噪声抑制方法实施例的对应部分，在此，不再赘述。
综上所述，在本申请中，当外接框中包含背景目标时，本申请可以有效抑制背景目标形成的噪声；当外接框包含稠密分布的背景目标时，本申请可以精确定位待分割目标边界和目标位置；当背景目标与待分割目标相互紧密连接时，即使本申请外接框中包含了大量的背景噪声，本申请依然可以清晰地标定待分割目标的完整准确边界。
需要注意的是，本申请可在软件和/或软件与硬件的组合体中被实施，例如，可采用专用集成电路(ASIC)、通用目的计算机或任何其他类似硬件设备来实现。在一个实施例中，本申请的软件程序可以通过处理器执行以实现上文所述步骤或功能。同样地，本申请的软件程序(包括相关的数据结构)可以被存储到计算机可读记录介质中，例如，RAM存储器，磁或光驱动器或软磁盘及类似设备。另外，本申请的一些步骤或功能可采用硬件来实现，例如，作为与处理器配合从而执行各个步骤或功能的电路。
另外，本申请的一部分可被应用为计算机程序产品，例如计算机程序指令，当其被计算机执行时，通过该计算机的操作，可以调用或提供根据本申请的方法和/或技术方案。而调用本申请的方法的程序指令，可能被存储在固定的或可移动的记录介质中，和/或通过广播或其他信号承载媒体中的数据流而被传输，和/或被存储在根据所述程序指令运行的计算机设备的工作存储器中。在此，根据本申请的一个实施例包括一个装置，该装置包括用于存储计算机程序指令的存储器和用于执行程序指令的处理器，其中，当该计算机程序指令被该处理器执行时，触发该装置运行基于前述根据本申请的多个实施例的方法和/或技术方案。
对于本领域技术人员而言，显然本申请不限于上述示范性实施例的细节，而且在不背离本申请的精神或基本特征的情况下，能够以其他的具体形式实现本申请。因此，无论从哪一点来看，均应将实施例看作是示范性的，而且是非限制性的，本申请的范围由所附权利要求而不是上述说明限定，因此旨在将落在权利要求的等同要件的含义和范围内的所有变化涵括在本申请内。不应将权利要求中的任何附图标记视为限制所涉及的权利要求。此外，显然“包括”一词不排除其他单元或步骤，单数不排除复数。装置权利要求中陈述的多个单元或装置也可以由一个单元或装置通过软件或者硬件来实现。第一，第二等词语用来表示名称，而并不表示任何特定的顺序。

Claims (11)

  1. 基于神经网络的稠密目标背景噪声抑制方法，其特征在于，所述方法包括：
    获取图像并确定多个稠密分布的待分割目标；
    分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框；
    基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理；
    基于上述待分割目标的层级排序处理，对待分割目标的旋转角度进行检测；
    基于上述对待分割目标的旋转角度的检测，对所述待分割目标进行外接框分割；
    基于待分割目标的外接框分割，输出待分割目标。
  2. 根据权利要求1所述的方法，其特征在于，上述的分别对上述多个稠密分布的待分割目标进行外接框分割，获得上述多个稠密分布的待分割目标所对应的外接框包括：通过外接矩形分割方法分别对上述多个稠密分布的待分割目标进行外接矩形框分割，获得所述多个稠密分布的待分割目标所对应的外接矩形框。
  3. 根据权利要求1或2所述的方法，其特征在于，基于上述多个待分割目标的外接框，对所述外接框进行编号，并为每个所述外接框建立一个相交框列表，并记录其中一个所述待分割目标的外接框与其他待分割目标的外接框的相交区域。
  4. 根据权利要求1或2所述的方法，其特征在于，上述的基于上述多个稠密分布的待分割目标所对应的外接框，进行所述待分割目标的层级排序处理，包括：
    基于每个上述外接框所建立一个相交框的列表，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，其中，
    P i＝area cover/area i
    其中，area i表示其中任意一个外接框所覆盖的面积，area cover表示该外接框被其他干扰的外接框所覆盖的面积；
    其中，定义外接框被覆盖的阈值为T，如果所覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理。
  5. 根据权利要求4所述的方法，其特征在于，
    基于上述的将符合该阈值要求的所述外接框并列在同一层级处理结果，将上述符合阈值要求的外接框从相交框列表中删除，获得新的外接框的相交框列表；
    基于上述获得的新的外接框的相交框列表结果，计算每个所述外接框被其他干扰的外接框所覆盖的面积比P i，并将该覆盖的面积比P i与阈值T比较，如果该覆盖的面积比P i小于该阈值T，那么将符合该阈值要求的所述外接框并列在同一层级处理；
    重复上述步骤，直至完成所有的待分割目标的层级排序处理。
  6. 根据权利要求1所述的方法，其特征在于，基于上述的待分割目标的层级排序处理结果，按照层级排序图从上到下依次处理每一层的目标。
  7. 根据权利要求6所述的方法，其特征在于，所述基于上述的待分割目标的层级排序处理结果，按照层级排序图从上到下依次处理每一层的目标， 包括：采用目标旋转角度检测方法，分别计算出同一层的每个待分割目标的旋转角度。
  8. 根据权利要求7所述的方法，其特征在于，基于上述的采用目标旋转角度检测方法，分别计算出同一层的每个待分割目标的旋转角度结果，再次进行外接矩形分割，获得同一层的每个待分割目标的精确位置。
  9. 根据权利要求8所述的方法，其特征在于，基于上述的获得同一层的每个待分割目标的精确位置，将上述已经完成分割的同一层的每个待分割目标去除后，对层级排序图中的下一层的所有待分割目标进行分割，直至所有层的待分割目标均被分割完成。
  10. 一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现权利要求1至9中任一项所述的方法。
  11. 一种计算机设备，包括存储器和处理器，所述存储器存储计算机程序，其特征在于，所述计算机程序被处理器执行时实现权利要求1至9中任一项所述的方法。
PCT/CN2020/128021 2019-11-11 2020-11-11 基于神经网络的目标背景噪声抑制方法及设备 WO2021093756A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911097155.0 2019-11-11
CN201911097155.0A CN110930420B (zh) 2019-11-11 2019-11-11 基于神经网络的稠密目标背景噪声抑制方法及设备

Publications (1)

Publication Number Publication Date
WO2021093756A1 true WO2021093756A1 (zh) 2021-05-20

Family

ID=69852676

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/128021 WO2021093756A1 (zh) 2019-11-11 2020-11-11 基于神经网络的目标背景噪声抑制方法及设备

Country Status (2)

Country Link
CN (1) CN110930420B (zh)
WO (1) WO2021093756A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110930420B (zh) * 2019-11-11 2022-09-30 中科智云科技有限公司 基于神经网络的稠密目标背景噪声抑制方法及设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8467607B1 (en) * 2011-11-21 2013-06-18 Google Inc. Segmentation-based feature pooling for object models
CN107451602A (zh) * 2017-07-06 2017-12-08 浙江工业大学 一种基于深度学习的果蔬检测方法
CN109948457A (zh) * 2019-02-26 2019-06-28 南京理工大学 基于卷积神经网络和cuda加速的实时目标识别算法
CN109961049A (zh) * 2019-03-27 2019-07-02 东南大学 一种复杂场景下香烟品牌识别方法
CN110298298A (zh) * 2019-06-26 2019-10-01 北京市商汤科技开发有限公司 目标检测及目标检测网络的训练方法、装置及设备
CN110930420A (zh) * 2019-11-11 2020-03-27 中科智云科技有限公司 基于神经网络的稠密目标背景噪声抑制方法及设备

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11423548B2 (en) * 2017-01-06 2022-08-23 Board Of Regents, The University Of Texas System Segmenting generic foreground objects in images and videos
EP3392832A1 (en) * 2017-04-21 2018-10-24 General Electric Company Automated organ risk segmentation machine learning methods and systems
CN108875577A (zh) * 2018-05-11 2018-11-23 深圳市易成自动驾驶技术有限公司 目标检测方法、装置及计算机可读存储介质
CN109255320B (zh) * 2018-09-03 2020-09-25 电子科技大学 一种改进的非极大值抑制方法
CN109800735A (zh) * 2019-01-31 2019-05-24 中国人民解放军国防科技大学 一种船目标精确检测与分割方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8467607B1 (en) * 2011-11-21 2013-06-18 Google Inc. Segmentation-based feature pooling for object models
CN107451602A (zh) * 2017-07-06 2017-12-08 浙江工业大学 一种基于深度学习的果蔬检测方法
CN109948457A (zh) * 2019-02-26 2019-06-28 南京理工大学 基于卷积神经网络和cuda加速的实时目标识别算法
CN109961049A (zh) * 2019-03-27 2019-07-02 东南大学 一种复杂场景下香烟品牌识别方法
CN110298298A (zh) * 2019-06-26 2019-10-01 北京市商汤科技开发有限公司 目标检测及目标检测网络的训练方法、装置及设备
CN110930420A (zh) * 2019-11-11 2020-03-27 中科智云科技有限公司 基于神经网络的稠密目标背景噪声抑制方法及设备

Also Published As

Publication number Publication date
CN110930420B (zh) 2022-09-30
CN110930420A (zh) 2020-03-27

Similar Documents

Publication Publication Date Title
JP6321681B2 (ja) ウェブサイトユーザを識別する方法および装置
CN109726658B (zh) 人群计数及定位方法、系统、电子终端及存储介质
CN111091123A (zh) 文本区域检测方法及设备
CN111598825B (zh) 数据处理方法、瑕疵的检测方法、计算设备及存储介质
CN104112284B (zh) 一种图片的相似度检测方法和设备
US11967088B2 (en) Method and apparatus for tracking target
CN111310759B (zh) 双模式协作的目标检测抑制优化方法及设备
WO2021082922A1 (zh) 一种检测屏幕显示断线的方法及设备
WO2021093756A1 (zh) 基于神经网络的目标背景噪声抑制方法及设备
CN109886956B (zh) 检测缺陷点聚集性的方法及装置
CN107993239B (zh) 一种计算单目图像的深度次序的方法和装置
WO2020135230A1 (zh) 一种内切圆确定方法及装置
CN109919095A (zh) 危化品仓库中码垛的监控方法及电子设备
CN109855634B (zh) 一种栅格地图的图像处理的方法及设备
CN111696152B (zh) 检测包裹堆垛的方法、装置、计算设备、系统及存储介质
CN110866928B (zh) 基于神经网络的目标边界分割及背景噪声抑制方法及设备
CN109961061A (zh) 一种边缘计算视频数据结构化方法及系统
CN114819742A (zh) 一种企业风险的评估方法及装置、电子设备、存储介质
JP2020165920A (ja) 飛跡検出装置、飛跡検出方法、及び飛跡検出プログラム
CN111310824A (zh) 多角度稠密目标检测抑制优化方法及设备
CN110458070A (zh) 基于机动车年检检验表图片识别检验次数的方法与系统
JP2001099625A (ja) パターン検査装置およびパターン検査方法
Kim et al. Deep Learning-Based Detection of Defects in Wafer Buffer Zone During Semiconductor Packaging Process
KR20190052785A (ko) 객체 검출 방법, 장치 및 컴퓨터 프로그램
TWI765475B (zh) 圖像瑕疵檢測方法、裝置、電腦裝置及介質

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: 20886528

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: 20886528

Country of ref document: EP

Kind code of ref document: A1