WO2021104125A1 - 禽蛋异常的识别方法、装置及系统、存储介质、电子装置 - Google Patents

禽蛋异常的识别方法、装置及系统、存储介质、电子装置 Download PDF

Info

Publication number
WO2021104125A1
WO2021104125A1 PCT/CN2020/129784 CN2020129784W WO2021104125A1 WO 2021104125 A1 WO2021104125 A1 WO 2021104125A1 CN 2020129784 W CN2020129784 W CN 2020129784W WO 2021104125 A1 WO2021104125 A1 WO 2021104125A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
network
target
egg
pixel
Prior art date
Application number
PCT/CN2020/129784
Other languages
English (en)
French (fr)
Inventor
苏睿
Original Assignee
京东数科海益信息科技有限公司
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 京东数科海益信息科技有限公司 filed Critical 京东数科海益信息科技有限公司
Publication of WO2021104125A1 publication Critical patent/WO2021104125A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches

Definitions

  • This application relates to the field of intelligent breeding, and in particular to a method, device and system, storage medium, and electronic device for identifying abnormalities of poultry eggs.
  • the egg breakage detection program in the related technology mainly uses optical illumination for detection, and the specific steps are as follows:
  • the lamp tube illuminates the eggs, and the detection device detects and identifies the eggs. If the eggs are intact, they will enter the conveyor chain through the screening device. If the detection device detects that the eggs are broken, the electronic control baffle in the screening device Just open, the eggs fall into the first-level temporary storage box, and the intact eggs enter the conveyor chain;
  • the embodiments of the present application provide a method, device, system, storage medium, and electronic device for recognizing abnormality of poultry eggs, so as to at least solve the technical problem of low accuracy of detecting abnormality of poultry eggs in related technologies.
  • a method for identifying abnormalities of poultry eggs including: obtaining a recognition request, wherein the recognition request is used to request recognition of the abnormal type of the target poultry egg; The first image collected by the egg; the target model is used to identify the abnormal type of the target egg in the first image, where the target model uses the spatial dimension characteristics of the first image when identifying the abnormal type of the target egg And the characteristics of the channel dimension.
  • a device for identifying abnormality of poultry eggs including: a first obtaining unit configured to obtain a recognition request, wherein the recognition request is used to request to recognize the abnormal type of the target poultry egg
  • the second acquisition unit is configured to acquire the first image obtained by collecting the target egg in response to the identification request;
  • the identification unit is configured to use the target model to identify the abnormal type of the target egg in the first image, wherein ,
  • the target model uses the spatial dimension of the first image and the channel dimension when identifying the abnormal type of the target egg.
  • a system for identifying abnormalities of poultry eggs including: an image acquisition device for acquiring a first image to be processed, wherein the first image is a collection of a target poultry egg The obtained image; the server is used to use the target model to identify the abnormal type of the target egg in the first image in the case of receiving the recognition request, wherein the target model uses the target model to identify the abnormal type of the target egg The features of the first image in the spatial dimension and the channel dimension, wherein the identification request is used to request to identify the abnormal type of the target egg.
  • a storage medium includes a stored program, and the above-mentioned method is executed when the program is running.
  • an electronic device including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor executes the above-mentioned method through the computer program.
  • the characteristics of the first image in the spatial dimension and the channel dimension are used to determine the abnormality type, and a two-step Attention-based egg breakage detection scheme is provided, and the spatial attention mechanism and the channel attention mechanism are added. Optimizing the segmentation results can solve the technical problem of low detection accuracy of egg abnormalities in related technologies, and then achieve the technical effect of improving the accuracy of detection of egg abnormalities.
  • Fig. 1 is a schematic diagram of the hardware environment of the method for recognizing an abnormality of a poultry egg according to an embodiment of the present application
  • FIG. 2 is a flowchart of an optional method for recognizing an abnormality of poultry eggs according to an embodiment of the present application
  • Fig. 3 is a schematic diagram of an optional network model according to an embodiment of the present application.
  • Fig. 4 is a schematic diagram of the network layer in an optional model according to an embodiment of the present application.
  • Fig. 5 is a schematic diagram of the network layer in an optional model according to an embodiment of the present application.
  • Fig. 6 is a schematic diagram of a network layer in an optional model according to an embodiment of the present application.
  • Fig. 7 is a schematic diagram of a network layer in an optional model according to an embodiment of the present application.
  • Fig. 8 is a schematic diagram of an optional device for identifying abnormalities of poultry eggs according to an embodiment of the present application.
  • Fig. 9 is a structural block diagram of a terminal according to an embodiment of the present application.
  • the above-mentioned method for recognizing abnormality of poultry eggs may be applied to the hardware environment constituted by the image acquisition device 101 and the server 103 as shown in FIG. 1.
  • the image acquisition device is used to acquire the first image of the target poultry egg to be processed
  • the server is used to identify the first image in the first image by using the target model upon receiving the acquisition recognition request.
  • the abnormal type of the target egg, the target model uses the spatial dimension feature and the channel dimension feature of the first image when identifying the abnormal type of the target egg
  • the server 103 communicates with the image through the network
  • the collection device 101 is connected and can be used to provide image recognition services for the image collection device 101.
  • the database 105 can be set on the server or independently of the server to provide data storage services for the server 103.
  • the above-mentioned networks include but are not limited to: wide area network, city Local area network or local area network.
  • the method for recognizing an abnormality of a poultry egg in the embodiment of the present application may be executed by the server 103 (specifically, the following steps S202 to S206 may be executed), or may be executed jointly by the server 103 and the image acquisition device 101.
  • Fig. 2 is a flowchart of an optional method for recognizing avian egg abnormalities according to an embodiment of the present application. As shown in Fig. 2, the method may include the following steps:
  • Step S202 Obtain an identification request, which is used to request identification of the abnormal type of the target egg.
  • the above abnormal type can be at least one of damage (or non-destructive testing), color abnormality, shape abnormality and other abnormalities.
  • damage or non-destructive testing
  • the following description will take the abnormal type as damage as an example; the above-mentioned poultry eggs can be eggs, duck eggs, goose eggs, turtles Eggs, such as eggs, will be described later by taking poultry eggs as eggs.
  • Non-destructive testing refers to the use of changes in heat, sound, light, electricity, and magnetism caused by abnormalities or defects in the internal structure of the material under the premise of not damaging or affecting the performance of the tested object, and not damaging the internal organization of the tested object. , Use physical or chemical methods as a means to inspect the structure, nature, state and type, nature, quantity, shape, location, size, distribution and changes of the internal and surface of the test piece with the help of modern technology and equipment. And the method of testing.
  • Step S204 in response to the identification request, obtain a first image obtained by collecting the target poultry egg.
  • step S206 the target model is used to identify the abnormal type of the target egg in the first image, and the target model uses the features in the space dimension and the channel dimension of the first image when identifying the abnormal type of the target egg, and then Determine the type of abnormality through semantic segmentation.
  • Semantic segmentation is a classification at the pixel level. Pixels belonging to the same category are all classified into one category. Therefore, semantic segmentation is to understand the image from the pixel level. For example, to perform semantic segmentation on broken eggs, it is necessary to clarify where the eggs are damaged. position.
  • semantic segmentation methods such as Texton Forest and random forest classifiers were used more frequently.
  • deep learning methods have improved a lot over traditional methods. The deep learning method has achieved great success in semantic segmentation.
  • the deep learning method to solve the semantic segmentation problem can be summarized into several ideas, such as patch classification, full convolution method, encoder-decoder architecture, hole convolution, conditional random field, etc.
  • the damage detection of poultry eggs basically relies on optical detection, which is low in efficiency and cannot be guaranteed.
  • the first image's characteristics and channels in the spatial dimension are used. Dimensional features are used to determine the type of abnormality, and a two-step Attention-based egg damage detection scheme is provided.
  • the spatial attention mechanism and channel attention mechanism are added to optimize the segmentation results, which can solve the problem of low accuracy in detecting egg abnormalities in related technologies.
  • the technical problem and then achieve the technical effect of improving the accuracy of detection of abnormal poultry eggs.
  • the technical solution of the present application will be further detailed below in conjunction with the steps shown in FIG. 2.
  • a timer can be set in the program to trigger the recognition request regularly (e.g. every 5 seconds, every 1 second), or it can be triggered when needed (e.g. the user actively triggers) the recognition request to Request to identify the abnormal type of the target egg.
  • step S204 in response to the identification request, the first image obtained by collecting the target poultry egg is obtained.
  • acquisition can be the real-time collection of the first image of the poultry egg by using the image acquisition device, or It is the first image of the poultry egg that is collected and saved by the image collection device in advance from the database.
  • step S206 when the target model is used for identification, the spatial dimension feature and the channel dimension feature of the first image are used to determine the abnormal type of the target egg in the first image.
  • the above-mentioned target model is a pre-trained semantic segmentation neural network model, such as a model such as Enet (the semantic segmentation neural network model is the Enet model as an example for illustration), and the network structure of the target model is shown in Figure 3.
  • using the target model to identify the abnormal type of the target egg in the first image includes steps S2062-step S2066:
  • Step S2062 Perform a compression operation on the first image through the first network in the target model to obtain a second image, where the compression operation includes eliminating visual redundant information in the first image.
  • FIG. 4 An optional first network is shown in Figure 4, which can also be called the initial layer.
  • Enet solves the timeliness of the semantic segmentation model by reducing floating point operations.
  • the first image is compressed through the first network in the target model, and when the second image is obtained, the convolutional layer of the first network (as shown in Figure 4 has a size of 3*3 and stride Convolution kernel of 2) performs a convolution operation on the first image, and performs a pooling operation on the first image through the pooling layer MaxPooling of the first network; through the splicing layer Concat of the first network, the convolution operation is performed The result and the result of performing the pooling operation are spliced together to obtain the second image.
  • Step S2064 Perform semantic segmentation on the second image through the second network in the target model, and extract the features in the space dimension and the channel dimension carried in the second image to obtain the third image.
  • FIG. 5 An optional unit in the second network is shown in FIG. 5, and the second network may include multiple (such as 5) bottleneck structure units as shown in FIG. 5 (part of which is used for encoding and the other is used for decoding),
  • the third image is obtained through the encoding in the second network
  • the network performs feature encoding processing on the second image to obtain the fourth image; the fourth image is decoded through the decoding network in the second network to obtain the third image, where the third image carries the decoding network to the fourth image
  • the information obtained by semantic segmentation and the features in the space dimension and the channel dimension are extracted.
  • the five bottleneck structures include a larger encoder (such as the first three bottleneck structures) and a smaller decoder (such as the latter two bottleneck structures). Try to do as much as possible without significantly affecting the segmentation accuracy. It reduces the size of the network and the number of parameters.
  • batch normalized BN processing can be performed, PReLU represents the modified linear unit activation function with parameters; MaxPooling represents maximum pooling.
  • the constituent elements in the source data can be regarded as a series of ⁇ Key, Value> data pairs.
  • a certain element Query in the given target is calculated by calculating the similarity between the Query and each Key. Or correlation, the weight coefficient of each Key corresponding to Value is obtained, and then the Value is weighted and summed to obtain the final Attention value.
  • the Attention mechanism is to perform a weighted summation of the Value values of the elements in the Source, and Query and Key are used to calculate the weight coefficient of the corresponding Value.
  • performing decoding processing on the fourth image through the decoding network in the second network to obtain the third image includes determining the pixel value of each column of pixels in the third image in the spatial dimension according to the following formula: among them, E j represents the pixel value of a row of pixels in the j-th column of the third image in the spatial dimension, ⁇ is the coefficient of the spatial dimension, A j represents the pixel value of the pixel in the same row of the j-th column in the fourth image A, B, C, D is the image obtained by processing the fourth image A using different convolution kernels, B i represents the pixel value of the pixel in the i-th row in image B, C j represents the pixel value of the pixel in the j-th column of image C, and D i represents the image The pixel value of the pixel in the i-th row in D, N represents the maximum value of i, and s ji represents the pixel value of the pixel in the j-th column and the i-th row in
  • E ji ⁇ (s ji *D ji )+A ji
  • E ji represents the pixel value of the pixel in the jth column and row i in the third image in the spatial dimension (or called The feature value of the spatial dimension)
  • D ji represents the pixel value of the pixel in the jth column and the i row in the image D.
  • s ji is the element of the i-th row and j-th column of the Softmax result matrix after the B matrix is transposed and the C matrix is multiplied, (B, C, D are obtained from A through 1*1 convolution), this step It is to calculate each element of the similarity matrix of Query and Key in the spatial Attention matrix.
  • the D matrix is transformed into a (H*W)*(H*W) matrix through a reshape transformation operation (such as matrix change dimension) and multiplied by each element s ji of the above Attention matrix, which is equivalent to assigning a weight to each position ( Multiply by the ⁇ coefficient), and add the original matrix A to get the final spatial attention result.
  • a reshape transformation operation such as matrix change dimension
  • performing decoding processing on the fourth image through the decoding network in the second network to obtain the third image includes determining the pixel value of each column of pixels in the third image in the channel dimension according to the following formula: among them, E j represents the pixel value of the pixel in the jth column of the third image in the channel dimension, ⁇ is the coefficient of the channel dimension, A j represents the pixel value of the pixel in the jth column of the fourth image A, and Ax represents the pixel value of the jth column of the fourth image A.
  • E ji ⁇ (x ji *A ji )+A ji
  • E ji represents the pixel value of the pixel in the jth column and i row of the third image in the channel dimension (or called Eigenvalues of the channel dimensions)
  • x ji is the softmax result of the matrix product of A i and A j obtained after calculating the matrix A reshape (matrix variable dimension) (that is, the attention weight of each region), which is used to calculate different channels
  • matrix A reshape matrix variable dimension
  • Step S2066 Identify the abnormal type of the third image through the third network in the target model, and determine the abnormal type of the target poultry egg.
  • the above-mentioned third network can be a classification layer in a neural network model, or a fully connected layer, which can determine the specific abnormal type of the egg, such as whether it is damaged.
  • the technical solution of the present application will be described in detail below, taking poultry eggs as eggs as an example.
  • the egg detection scheme basically uses optical methods. In large-scale scenes such as chicken farms, the efficiency is low; optical detection is greatly affected by physical conditions, such as the intensity of light, the difference in media, etc., and the accuracy of detection No fundamental guarantee can be obtained; this application can implement an improved Enet network egg breakage detection algorithm based on two-step Attention for the broken egg detection scene, which effectively improves the detection efficiency and has high accuracy.
  • the improved Enet network egg breakage detection scheme based on the two-step Attention proposed in this application can solve the egg breakage detection problem and solve the problem of low efficiency and unguaranteed accuracy in related technologies.
  • This scheme is mainly based on Enet's semantic segmentation network, adding spatial attention mechanism and channel attention mechanism at the decoding end to optimize the segmentation results.
  • the main steps include:
  • Step 1 The first step is image acquisition.
  • the eggs to be inspected are collected by photographing equipment to perform image preprocessing, and then enter the network encoding stage.
  • Step 2 The encoding stage is mainly to obtain high-level semantic features through a series of convolution operations and down-sampling operations.
  • An optional coding part is shown in Table 1.
  • the structures of the initial part and bottleneck are shown in Figures 4 and 5:
  • the initial part of Enet first compresses the resolution of the image, which greatly improves the speed of the entire network.
  • the theory behind this is that there are a lot of visual information in space. Redundancy, so it can be compressed to a more effective feature instead.
  • the bottleneck part adopts a parallel structure. In the non-downsampling layer, there is no maximum pooling MaxPooling and padding on the left branch. The branch on the left side of the downsampling layer becomes MaxPooling and Padding. This parallel structure not only helps the characteristics Fusion and can accelerate the speed of forward calculation.
  • downsampling means downsampling
  • dilated means dilation processing
  • asymmetric means asymmetric convolution processing
  • Step 3 In the Enet decoding stage, you can replace Max Pooling in the left branch of Figure 5 with Max unpooling, perform two upsampling, and finally use deconvolution to perform the last upsampling to make the feature map the same size as the original image. Semantic segmentation is the main part, so the decoder is relatively simple just to fine-tune the result of the encoder. An optional decoding part is shown in Table 2.
  • Upsampling in Table 2 represents upsampling, fullconv represents a fully connected layer, and the three columns correspond to the name, type, and output size.
  • Step 4 This solution mainly improves the decoding part of Enet, that is, the original network of Enet has the same position except for the last deconvolution. This is done because there are many cases of damage detection in multiple locations, but the characteristics of the damage are similar, but it is difficult to effectively present the results of semantic segmentation, so two types of damage are added to the decoding part
  • the attention module models the semantic dependencies in the spatial dimension and channel dimension respectively.
  • the location attention module selectively aggregates the features of each location by weighting and summing the features of all locations. Regardless of distance, similar features are interrelated.
  • the channel attention module selectively emphasizes interdependent channel maps by integrating the relevant features of all channel maps. By summing the output of the two attention modules, the feature representation is further improved and the segmentation result is more accurate.
  • Step 4 The spatial attention module and the position attention mechanism module are shown in Figure 6.
  • the above operation is shown in the following formula:
  • Step 5 the channel Attention module is shown in Figure 7.
  • the transposition of B and C is multiplied by the C*C matrix, and then the softmax operation is performed on the C*C matrix, as shown in the following:
  • two-step Attention is used to decode the ENet network, which can still run in real time without increasing the amount of parameters, with high efficiency.
  • context information can be effectively aggregated. It is helpful to improve the accuracy of damage detection; solve the problem of pixel blocks in the edge detection of cracks in the E-NET network. This solution can effectively solve the problem of broken egg detection, with high efficiency and better accuracy than other solutions in related technologies.
  • the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is Better implementation.
  • the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, The optical disc) includes several instructions to enable a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the method described in each embodiment of the present application.
  • an abnormality recognition device for poultry eggs for implementing the above-mentioned method for recognizing abnormality of poultry eggs.
  • Fig. 8 is a schematic diagram of an optional device for identifying abnormalities of poultry eggs according to an embodiment of the present application. As shown in Fig. 8, the device may include:
  • the first obtaining unit 801 is configured to obtain an identification request, where the identification request is used to request to identify the abnormal type of the target poultry egg;
  • the second obtaining unit 803 is configured to obtain the first image obtained by collecting the target poultry egg in response to the identification request;
  • the recognition unit 805 is configured to use the target model to recognize the abnormal type of the target egg in the first image, where the target model uses the spatial dimensional features and channels of the first image when recognizing the abnormal type of the target egg Dimensional characteristics.
  • first obtaining unit 801 in this embodiment can be used to perform step S202 in the embodiment of the present application
  • second obtaining unit 803 in this embodiment can be used to perform step S204 in the embodiment of the present application
  • the identification unit 805 in this embodiment may be used to execute step S206 in the embodiment of the present application.
  • the characteristics of the first image in the spatial dimension and the channel dimension are used to determine the abnormality type, and a two-step attention-based poultry egg damage detection scheme is provided, and the spatial attention mechanism and the channel attention mechanism are added to optimize the segmentation results. , It can solve the technical problem of low accuracy in detecting abnormality of poultry eggs in related technologies, thereby achieving the technical effect of improving the accuracy of detecting abnormality of poultry eggs.
  • the recognition unit includes: a compression module configured to perform a compression operation on the first image through the first network in the target model to obtain a second image, wherein the compression operation includes eliminating visual redundant information in the first image
  • the encoding and decoding module is set to perform semantic segmentation on the second image through the second network in the target model, and extract the features in the space dimension and the channel dimension carried in the second image to obtain the third image; recognition;
  • the module is set to identify the abnormal type of the third image through the third network in the target model, and determine the abnormal type of the target egg.
  • the encoding and decoding module when the encoding and decoding module performs semantic segmentation on the second image through the second network in the target model, and extracts the features in the space dimension and the channel dimension carried in the second image, to obtain the third image,
  • the second image can be subjected to feature encoding processing through the encoding network in the second network to obtain the fourth image;
  • the fourth image can be decoded through the decoding network in the second network to obtain the third image, where in the third image It carries the information obtained by the semantic segmentation of the fourth image by the decoding network and the extracted features in the space dimension and the channel dimension.
  • the encoding and decoding module may also be configured to decode the fourth image through the decoding network in the second network to obtain the third image including determining the spatial dimension of each column of pixels in the third image according to the following formula Pixel values: among them, E j represents the pixel value of the pixel in the jth column of the third image in the spatial dimension, ⁇ is the coefficient of the spatial dimension, A j represents the pixel value of the pixel in the jth column of the fourth image A, and B, C, and D use different
  • the pixel value of the row pixel, N represents the maximum value of i, and s ji represents the pixel value of the pixel in the j-th column and the i-th row in the image
  • the encoding and decoding module may also be configured to decode the fourth image through the decoding network in the second network, to obtain the third image, including determining the channel dimension of each column of pixels in the third image according to the following formula Pixel values: among them, E j represents the pixel value of the j-th column pixel in the third image in the channel dimension, ⁇ is the coefficient of the channel dimension, A j represents the pixel value of the j-th column pixel in the fourth image A , and A i represents the pixel value of the fourth image A The pixel value of the pixel in the i-th row, C represents the maximum value of i, and x ji represents the pixel value of the pixel in the j-th column and the i-th row in the image x.
  • Pixel values including determining the channel dimension of each column of pixels in the third image according to the following formula Pixel values: among them, E j represents the pixel value of the j-th column pixel in the third image in the channel dimension, ⁇ is the
  • the compression module may also be configured to perform a convolution operation on the first image through the convolution layer of the first network, and perform a pooling operation on the first image through the pooling layer of the first network;
  • the stitching layer of the network stitches the result of performing the convolution operation and the result of performing the pooling operation to obtain the second image.
  • the above-mentioned modules can run in the hardware environment as shown in FIG. 1, and can be implemented by software or hardware, where the hardware environment includes a network environment.
  • a server or terminal for implementing the above-mentioned method for identifying abnormalities in poultry eggs.
  • FIG. 9 is a structural block diagram of a terminal according to an embodiment of the present application.
  • the terminal may include: one or more (only one is shown in FIG. 9) processors 901, a memory 903, and a transmission device 905, as shown in FIG. 9, the terminal may also include an input and output device 907.
  • the memory 903 can be used to store software programs and modules, such as the program instructions/modules corresponding to the method and device for recognizing abnormalities of poultry eggs in the embodiment of the present application.
  • the processor 901 runs the software programs and modules stored in the memory 903, In this way, various functional applications and data processing are executed, that is, the above-mentioned method for identifying abnormality of poultry eggs is realized.
  • the memory 903 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory.
  • the memory 903 may further include a memory remotely provided with respect to the processor 901, and these remote memories may be connected to the terminal through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
  • the aforementioned transmission device 905 is used for receiving or sending data via a network, and can also be used for data transmission between the processor and the memory.
  • the above-mentioned specific examples of networks may include wired networks and wireless networks.
  • the transmission device 905 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices and routers via a network cable so as to communicate with the Internet or a local area network.
  • the transmission device 905 is a radio frequency (RF) module, which is used to communicate with the Internet in a wireless manner.
  • RF radio frequency
  • the memory 903 is used to store application programs.
  • the processor 901 may call the application program stored in the memory 903 through the transmission device 905 to perform the following steps:
  • the identification request is used to request to identify the abnormal type of the target egg
  • the target model is used to identify the abnormal type of the target egg in the first image, wherein the target model uses the spatial dimension feature and the channel dimension feature of the first image when identifying the abnormal type of the target egg.
  • the processor 901 is further configured to perform the following steps:
  • the abnormal type of the third image is identified through the third network in the target model, and the abnormal type of the target egg is determined.
  • an "acquisition identification request is provided, wherein the identification request is used to request the identification of the abnormal type of the target egg; in response to the identification request, the first image obtained by collecting the target egg is obtained; using the target
  • the model recognizes the abnormal type of the target egg in the first image, where the target model uses the feature in the spatial dimension of the first image and the feature in the channel dimension when identifying the abnormal type of the target egg.
  • the characteristics of the first image in the spatial dimension and the channel dimension are used to determine the anomaly type.
  • a two-step Attention-based egg breakage detection solution is provided.
  • the spatial attention mechanism and the channel attention mechanism are added to optimize the segmentation results, which can solve related technologies.
  • the technical problem of low accuracy in detecting abnormalities of poultry eggs thereby achieving the technical effect of improving the accuracy of detecting abnormalities in poultry eggs.
  • the structure shown in Fig. 9 is only for illustration, and the terminal can be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, and a mobile Internet device (MID), Terminal equipment such as PAD.
  • FIG. 9 does not limit the structure of the above-mentioned electronic device.
  • the terminal may also include more or fewer components (such as a network interface, a display device, etc.) than those shown in FIG. 9, or have a configuration different from that shown in FIG.
  • the program can be stored in a computer-readable storage medium, and the storage medium can be Including: flash disk, read-only memory (Read-Only Memory, ROM), random access device (Random Access Memory, RAM), magnetic disk or optical disk, etc.
  • the embodiment of the present application also provides a storage medium.
  • the above-mentioned storage medium may be used to execute the program code of the method for identifying abnormality of poultry eggs.
  • the foregoing storage medium may be located on at least one of the multiple network devices in the network shown in the foregoing embodiment.
  • the storage medium is configured to store program code for executing the following steps:
  • the identification request is used to request to identify the abnormal type of the target egg
  • the target model is used to identify the abnormal type of the target egg in the first image, wherein the target model uses the spatial dimension feature and the channel dimension feature of the first image when identifying the abnormal type of the target egg.
  • the storage medium is also configured to store program code for executing the following steps:
  • the abnormal type of the third image is identified through the third network in the target model, and the abnormal type of the target egg is determined.
  • the foregoing storage medium may include, but is not limited to: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk Various media that can store program codes, such as discs or optical discs.
  • the integrated unit in the foregoing embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in the foregoing computer-readable storage medium.
  • the technical solution of the present application essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, It includes several instructions to make one or more computer devices (which may be personal computers, servers, or network devices, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the disclosed client can be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division.
  • there may be other division methods for example, multiple units or components may be combined or may be Integrate into another system, or some features can be ignored or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, units or modules, and may be in electrical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.

Landscapes

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

Abstract

本申请公开了一种禽蛋异常的识别方法、装置及系统、存储介质、电子装置。其中，该方法包括：获取识别请求，识别请求用于请求识别目标禽蛋的异常类型；响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；利用目标模型识别出第一图像中目标禽蛋的异常类型，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。本申请解决了相关技术中对禽蛋异常的检测准确率较低的技术问题。

Description

禽蛋异常的识别方法、装置及系统、存储介质、电子装置
本申请要求于提交中国专利局，优先权号为201911175385.4、发明名称“禽蛋异常的识别方法、装置及系统、存储介质、电子装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及智能养殖领域，具体而言，涉及一种禽蛋异常的识别方法、装置及系统、存储介质、电子装置。
背景技术
相关技术中的鸡蛋破损检测方案主要是利用光学照射进行检测，具体步骤如下：
1)将鸡蛋放置在传送带上，鸡蛋在传送带、辅助棍以及输送链作用下逐个进入输送链；
2)鸡蛋在输送链传输过程中灯管发光照射鸡蛋，检测装置对鸡蛋进行检测鉴别，若鸡蛋完好则经过筛选装置进入输送链，如检测装置检测到鸡蛋破损则筛选装置内的电控挡板就打开，鸡蛋落入一级暂存箱，完好的鸡蛋进入输送链；
3)经输送链落入重力感应器，重力感应器对鸡蛋进行重力检测分类，若鸡蛋重量达标则鸡蛋从进蛋口落入三级暂存箱，不达标则电控挡板打开，鸡蛋落入二级暂存箱。
相关技术的方案基本上是采用光学的方法，在养鸡场等大型场景下，效率低下；光学检测受物理条件影响巨大，比如光的强弱，介质的不同等等，检测的准确率不能得到根本的保证。类似地，在鸭蛋、鹅蛋等禽蛋中也存在类似地问题。
针对上述的问题，目前尚未提出有效的解决方案。
发明内容
本申请实施例提供了一种禽蛋异常的识别方法、装置及系统、存储介质、电子装置，以至少解决相关技术中对禽蛋异常的检测准确率较低的技术问题。
根据本申请实施例的一个方面，提供了一种禽蛋异常的识别方法，包括：获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。
根据本申请实施例的另一方面，还提供了一种禽蛋异常的识别装置，包括：第一获取单元，被设置为获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；第二获取单元，被设置为响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；识别单元，被设置为利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。
根据本申请实施例的另一方面，还提供了一种禽蛋异常的识别系统，包括：图像采集设备，用于采集待处理的第一图像，其中，第一图像是对目标禽蛋进行采集得到的图像；服务器，用于在接收到获取识别请求的情况下，利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征，其中，识别请求用于请求识别目标禽蛋的异常类型。
根据本申请实施例的另一方面，还提供了一种存储介质，该存储介质包括存储的程序，程序运行时执行上述的方法。
根据本申请实施例的另一方面，还提供了一种电子装置，包括存 储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，处理器通过计算机程序执行上述的方法。
在本申请实施例中，使用第一图像在空间维度上的特征和通道维度上的特征来确定异常类型，提供了基于两步Attention的禽蛋破损检测方案，加入空间注意机制和通道注意机制来优化分割结果，可以解决相关技术中对禽蛋异常的检测准确率较低的技术问题，进而达到提高禽蛋异常的检测准确率的技术效果。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的不当限定。在附图中：
图1是根据本申请实施例的禽蛋异常的识别方法的硬件环境的示意图；
图2是根据本申请实施例的一种可选的禽蛋异常的识别方法的流程图；
图3是根据本申请实施例的一种可选的网络模型的示意图；
图4是根据本申请实施例的一种可选的模型中网络层的示意图；
图5是根据本申请实施例的一种可选的模型中网络层的示意图；
图6是根据本申请实施例的一种可选的模型中网络层的示意图；
图7是根据本申请实施例的一种可选的模型中网络层的示意图；
图8是根据本申请实施例的一种可选的禽蛋异常的识别装置的示意图；以及，
图9是根据本申请实施例的一种终端的结构框图。
具体实施方式
为了使本技术领域的人员更好地理解本申请方案，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，所描述的实施例仅仅是本申请一部分的实施例，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，都应当属于本申请保护的范围。
需要说明的是，本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
根据本申请实施例的一方面，提供了一种禽蛋异常的识别方法的方法实施例。
可选地，在本实施例中，上述禽蛋异常的识别方法可以应用于如图1所示的由图像采集设备101和服务器103所构成的硬件环境中。如图1所示，图像采集设备用于采集待处理的所述目标禽蛋的第一图像，服务器用于在接收到获取识别请求的情况下，利用目标模型识别出所述第一图像中所述目标禽蛋的异常类型，所述目标模型在识别出所述目标禽蛋的异常类型时使用了所述第一图像在空间维度上的特征和通道维度上的特征，服务器103通过网络与图像采集设备101进行连接，可用于为图像采集设备101提供图像识别服务，可在服务器上或独立于服务器设置数据库105，用于为服务器103提供数据存储服务，上述网络包括但不限于：广域网、城域网或局域网。
本申请实施例的禽蛋异常的识别方法可以由服务器103来执行(具体可执行下述步骤S202至步骤S206)，也可以由服务器103和图像采集设备101共同执行。图2是根据本申请实施例的一种可选的禽蛋异常的识别方法的流程图，如图2所示，该方法可以包括以下步骤：
步骤S202，获取识别请求，识别请求用于请求识别目标禽蛋的异常类型。
上述异常类型可为破损(或称无损检测)、颜色异常、形状异常等异常中的至少之一，后续以异常类型为破损为例进行描述；上述禽蛋可以为鸡蛋、鸭蛋、鹅蛋、龟蛋等蛋类，后续以禽蛋为鸡蛋为例进行描述。
无损检测是指在不损害或不影响被检测对象使用性能,不伤害被检测对象内部组织的前提下，利用材料内部结构异常或缺陷存在引起的热、声、光、电、磁等反应的变化，以物理或化学方法为手段，借助现代化的技术和设备器材，对试件内部及表面的结构、性质、状态及缺陷的类型、性质、数量、形状、位置、尺寸、分布及其变化进行检查和测试的方法。
步骤S204，响应于识别请求，获取对目标禽蛋进行采集得到的第一图像。
步骤S206，利用目标模型识别出第一图像中目标禽蛋的异常类型，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征，然后通过语义分割确定异常类型。
语义分割是在像素级别上的分类，属于同一类的像素都要被归为一类，因此语义分割是从像素级别来理解图像的，比如说对破损鸡蛋进行语义分割，需要明确鸡蛋破损处的位置。在深度学习方法流行之前，纹理基元森林Texton Forest和基于随机森林分类器等语义分割方法是用得比较多的方法，不过在深度卷积网络流行之后，深度学习方 法比传统方法提升了很多，深度学习方法在语义分割上得到了巨大成功，深度学习方法解决语义分割问题可以概括为几种思路，如Patch classification、全卷积方法、encoder-decoder架构、空洞卷积、条件随机场等。
相关技术中对禽蛋进行破损检测，基本上依靠光学检测，效率低且准确率得不到保证，在本申请的技术方案中，通过上述步骤，使用第一图像在空间维度上的特征和通道维度上的特征来确定异常类型，提供了基于两步Attention的禽蛋破损检测方案，加入空间注意机制和通道注意机制来优化分割结果，可以解决相关技术中对禽蛋异常的检测准确率较低的技术问题，进而达到提高禽蛋异常的检测准确率的技术效果。下面结合图2所示的步骤进一步详述本申请的技术方案。
在步骤S202提供的技术方案中，可以在程序中设置定时器，定时触发(如每5秒、每1秒触发一次)识别请求，也可以在需要时触发(如用户主动触发)识别请求，以请求识别目标禽蛋的异常类型。
在步骤S204提供的技术方案中，响应于识别请求，获取对目标禽蛋进行采集得到的第一图像，此处的“获取”可以是利用图像采集设备实时采集禽蛋的第一图像，也可以是从数据库获取图像采集设备预先采集并保存的禽蛋的第一图像。
在步骤S206提供的技术方案中，利用目标模型识别时，使用了第一图像在空间维度上的特征和通道维度上的特征，从而确定第一图像中目标禽蛋的异常类型。
上述目标模型是预先训练好的语义分割神经网络模型，如Enet等模型(后续以语义分割神经网络模型为Enet模型为例进行说明)，目标模型的网络结构如图3所示。
可选地，利用目标模型识别出第一图像中目标禽蛋的异常类型包括步骤S2062-步骤S2066：
步骤S2062，通过目标模型中的第一网络对第一图像执行压缩操作，得到第二图像，其中，压缩操作包括消除第一图像中的视觉冗余信息。
一种可选的第一网络如图4所示，也可称为初始层initial层，为了满足实际应用中对方案的时效性极高的要求，Enet通过减少浮点运算解决了语义分割模型时效性差的问题，通过目标模型中的第一网络对第一图像执行压缩操作，得到第二图像时，通过第一网络的卷积层(如图4所示的大小为3*3、步长stride为2的卷积核)对第一图像执行卷积操作，并通过第一网络的池化层MaxPooling对第一图像执行池化操作；通过第一网络的拼接层Concat将执行卷积操作得到的结果和执行池化操作的结果拼接起来，得到第二图像。上述操作可消除第一图像中的视觉冗余信息(描述信源的数据是信息和数据冗余之和，即，数据＝信息+数据冗余，视觉冗余是图像数据中普遍存在的一种数据冗余)，过滤图像中的视觉冗余信息后，更加有利于后续的异常识别。
步骤S2064，通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像。
一种可选的第二网络中的单元如图5所示，第二网络可以包括多个(如5个)如图5所示的瓶颈结构单元(其中一部分用来编码一部分用来解码)，在通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像时，通过第二网络中的编码网络对第二图像进行特征编码处理，得到第四图像；通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像，其中，第三图像中携带有解码网络对第四图像进行语义分割得到的信息和提取出的空间维度上的特征和通道维度上的特征。如在Enet中，五个瓶颈结构中包括较大的编码器(如前三个瓶颈结构)和较小的解码器(如后两个瓶颈结构)，在不明显影响 分割精度的情况下尽可能地缩小了网络的体积，减少了参数数量。
继续以Enet为例，瓶颈结构如图5所示，瓶颈模块可包括：瓶颈模块可包括：1*1的投影层，用于降低特征的维度；主卷积层(conv)，用于进行特征卷积；1*1的扩展层；在所有卷积层之间放置批量标准化Regularizer和PReLU。如果瓶颈模块是下采样，则可将最大池化层(MaxPooling)添加到主分支，相反如果是上采样，则将填充层Padding添加到主分支。第一个1*1的投影可被替换为2*2卷积，步长stride＝2，对于Regularizer(正则化器)，可使用Spatial Dropout。在图5中，可以进行批量归一化BN处理，PReLU表示带参数的修正线性单元激活函数；MaxPooling表示最大池化。
在Attention机制中，可将源数据中的构成元素看成是由一系列的<Key,Value>数据对构成，此时给定目标中的某个元素Query，通过计算Query和各个Key的相似性或者相关性，得到每个Key对应Value的权重系数，然后对Value进行加权求和，即得到了最终的Attention数值。所以本质上Attention机制是对Source中元素的Value值进行加权求和，而Query和Key用来计算对应Value的权重系数。
可选地，通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像包括按照如下公式确定第三图像中每列像素在空间维度上的像素值：
Figure PCTCN2020129784-appb-000001
其中，
Figure PCTCN2020129784-appb-000002
E j表示第三图像中第j列某行像素在空间维度上的像素值，α是空间维度的系数，A j表示第四图像A中第j列相同行像素的像素值，B、C、D是利用不同的卷积核对第四图像A进行处理得到的图像，B i表示图像B 中第i行像素的像素值，C j表示图像C中第j列像素的像素值，D i表示图像D中第i行像素的像素值，N表示i的最大取值，s ji表示根据图像B和图像C确定的图像s中第j列第i行的像素的像素值。
对于单个像素，还可以采用如下公式来描述：E ji＝α(s ji*D ji)+A ji，E ji表示第三图像中第j列i行像素在空间维度上的像素值(或称空间维度的特征值)，D ji表示图像D中第j列i行像素的像素值。
s ji是B矩阵转置后和C矩阵相乘的Softmax结果矩阵的第i行和第j列的元素，(B、C、D是来自于A通过1*1卷积得到的)，这步是为了计算空间Attention矩阵中的Query和Key的相似度矩阵的每个元素。
D矩阵通过reshape变换操作(如矩阵变维)变成(H*W)*(H*W)矩阵并和上面的Attention矩阵的每个元素s ji相乘，相当于每个位置赋予了权重(再乘上α系数)，再加上原始的矩阵A，就得到最终空间attention的结果了。
可选地，通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像包括按照如下公式确定第三图像中每列像素在通道维度上的像素值：
Figure PCTCN2020129784-appb-000003
其中，
Figure PCTCN2020129784-appb-000004
E j 表示第三图像中第j列像素在通道维度上的像素值，β是通道维度的系数，A j表示第四图像A中第j列像素的像素值，Ax表示第四图像A的第i行像素的像素值，C表示i的最大取值，x ji表示图像x中第j列第i行的像素的像素值，exp为指数函数。
对于单个像素，还可以采用如下公式来描述：E ji＝β(x ji*A ji)+A ji，E ji表示第三图像中第j列i行像素在通道维度上的像素值(或称通道维度的特征值)，x ji是计算矩阵A reshape(矩阵变维)后得到的A i和A j的矩阵乘积的softmax结果(就是每个区域的注意力权值)，用于计算不同通道间的相似程度的矩阵的每个元素，再乘以β系数，再加上原始矩阵A，就得到了最终的通道Attention的结果了。
步骤S2066，通过目标模型中的第三网络对第三图像进行异常类型的识别，确定目标禽蛋的异常类型。
上述的第三网络可以为神经网络模型中的分类层，或称全连接层，可以确定出禽蛋的具体异常类型，如是否破损。
作为一种可选的实施例，下面以禽蛋为鸡蛋为例详述本申请的技术方案。相关技术中鸡蛋检测方案基本上是采用光学的方法，在养鸡场等大型场景下，效率低下；光学检测受物理条件影响巨大，比如光的强弱，介质的不同等等，检测的准确率不能得到根本的保证；本申 请可针对破损鸡蛋检测场景，实现了一种基于两步Attention的改进Enet网络鸡蛋破损检测算法，有效的提升检测效率，而且准确率高。
本申请提出的基于两步Attention的改进Enet网络鸡蛋破损检测方案，可解决鸡蛋破损检测问题，解决相关技术中效率低且准确率得不到保证的问题。本方案主要是基于Enet的语义分割网络，在解码端加入空间注意机制和通道注意机制来优化分割结果，主要步骤包括：
步骤1，首先是图像采集，将待检测的鸡蛋通过照相设备采集图片数据，进行图像预处理，之后进入网络编码阶段。
步骤2，编码阶段主要是通过一系列卷积操作和下采样操作，得到高层语义特征。
一种可选的编码部分如表1所示。其中initial部分和bottleneck的结构如图4和图5所示：Enet的initial部分先把图像的分辨率压缩了，大大提高了整个网络的运行速度，这背后的理论是视觉信息在空间上有很多冗余，所以可以被压缩到一个更有效的特征来代替。bottleneck部分采用的是并联结构，在非下采样层没有左边分支的最大池化MaxPooling和填充Padding，在下采样层左侧的分支就变成了MaxPooling和Padding，这样的并行结构不仅有助于特征的融合而且能够加速前向运算的速度。
表1
Figure PCTCN2020129784-appb-000005
表1中downsampling表示下采样，dilated表示膨胀处理，asymmetric表示非对称卷积处理。
步骤3，在Enet解码阶段可通过对图5中左分支的MaxPooling替换为Max unpooling，进行两次上采样，最后在用反卷积进行最后一次上采样使得特征图和原图大小相同，编码器在语义分割中是占主要部分的，所以解码器相对简单只是为了微调编码器的结果，一种可选的解码部分如表2所示。
表2
Figure PCTCN2020129784-appb-000006
表2中的upsampling表示上采样，fullconv表示全连接层，三列分别对应名称、类型以及输出大小。
步骤4，本方案主要改进了Enet的解码部分，也是就是Enet的原始网络除了最后一次反卷积以外，其他的位置不变。这样做是因为做破损检测的时候有很多情况是多位置破损的，但是破损的特征是相近的，但是在语义分割的结果上很难有效的呈现出来，所以在解码部分 附加了两种类型的注意力模块，分别在空间维度和通道维度上对语义依赖关系进行建模。位置注意模块通过对所有位置的特征进行加权和，有选择地聚合每个位置的特征。不管距离远近，相似的特征都是相互关联的。同时，通道注意模块通过整合所有通道图的相关特征，选择性地强调相互依赖的通道图。通过对两个注意模块的输出进行求和，进一步改进了特征表示，使分割结果更加精确。
步骤4，空间Attention模块，位置注意力机制模块如图6所示，B、C、D由A卷积得到，维度均为C*H*W。将B、C、D都降维reshape到C*N(N＝H*W)；然后将C的转置于B相乘到N*N的矩阵，对于矩阵的每一个点进行softmax处理，例如对于i行j列的点(i，j)，softmax考虑的是i行的每一个点，然后将D与softmax后的结果相乘并reshape到C*H*W，与A进行elementwise(即进行特征融合)。以上操作如下式所示：
Figure PCTCN2020129784-appb-000007
步骤5，通道Attention模块如图7所示，在通道注意力机制中，B与C的转置相乘C*C的矩阵，然后针对C*C的矩阵进行softmax操作，具体如下是所示：
Figure PCTCN2020129784-appb-000008
在本申请的技术方案中，利用两步Attention对ENet网络进行解码，没有增加多少参数量仍可以实时运行，效率高，利用在E-NET后加入两步Attention，可以有效的聚合上下文信息，有利于提高破损检测的精度；解决E-NET网络对裂缝边缘检测出现的像素块问题。本方案可以有效地解决破损鸡蛋检测问题，效率高，准确率优于相关技术中的其他方案。
需要说明的是，对于前述的各方法实施例，为了简单描述，故将其都表述为一系列的动作组合，但是本领域技术人员应该知悉，本申请并不受所描述的动作顺序的限制，因为依据本申请，某些步骤可以采用其他顺序或者同时进行。其次，本领域技术人员也应该知悉，说明书中所描述的实施例均属于优选实施例，所涉及的动作和模块并不一定是本申请所必须的。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到根据上述实施例的方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。
根据本申请实施例的另一个方面，还提供了一种用于实施上述禽蛋异常的识别方法的禽蛋异常的识别装置。图8是根据本申请实施例的一种可选的禽蛋异常的识别装置的示意图，如图8所示，该装置可以包括：
第一获取单元801，被设置为获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；
第二获取单元803，被设置为响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；
识别单元805，被设置为利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。
需要说明的是，该实施例中的第一获取单元801可以用于执行本 申请实施例中的步骤S202，该实施例中的第二获取单元803可以用于执行本申请实施例中的步骤S204，该实施例中的识别单元805可以用于执行本申请实施例中的步骤S206。
此处需要说明的是，上述模块与对应的步骤所实现的示例和应用场景相同，但不限于上述实施例所公开的内容。需要说明的是，上述模块作为装置的一部分可以运行在如图1所示的硬件环境中，可以通过软件实现，也可以通过硬件实现。
通过上述模块，使用第一图像在空间维度上的特征和通道维度上的特征来确定异常类型，提供了基于两步Attention的禽蛋破损检测方案，加入空间注意机制和通道注意机制来优化分割结果，可以解决相关技术中对禽蛋异常的检测准确率较低的技术问题，进而达到提高禽蛋异常的检测准确率的技术效果。
可选地，识别单元包括：压缩模块，被设置为通过目标模型中的第一网络对第一图像执行压缩操作，得到第二图像，其中，压缩操作包括消除第一图像中的视觉冗余信息；编解码模块，被设置为通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像；识别模块，被设置为通过目标模型中的第三网络对第三图像进行异常类型的识别，确定目标禽蛋的异常类型。
可选地，编解码模块在通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像时，可通过第二网络中的编码网络对第二图像进行特征编码处理，得到第四图像；通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像，其中，第三图像中携带有解码网络对第四图像进行语义分割得到的信息和提取出的空间维度上的特征和通道维度上的特征。
可选地，编解码模块还可被设置为：通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像包括按照如下公式确定第三图像中每列像素在空间维度上的像素值：
Figure PCTCN2020129784-appb-000009
其中，
Figure PCTCN2020129784-appb-000010
E j表示第三图像中第j列像素在空间维度上的像素值，α是空间维度的系数，A j表示第四图像A中第j列像素的像素值，B、C、D是利用不同的卷积核对第四图像A进行处理得到的图像，B i表示图像B中第i行像素的像素值，C j表示图像C中第j列像素的像素值，D i表示图像D中第i行像素的像素值，N表示i的最大取值，s ji表示根据图像B和图像C确定的图像s中第j列第i行的像素的像素值。
可选地，编解码模块还可被设置为：通过第二网络中的解码网络对第四图像进行解码处理，得到第三图像包括按照如下公式确定第三图像中每列像素在通道维度上的像素值：
Figure PCTCN2020129784-appb-000011
其中，
Figure PCTCN2020129784-appb-000012
E j表示第三图像中第j列像素在通道维度上的像素值，β是通道维度的系数，A j表示第四图像A中第j列像素的像素值，A i表示第四图像A的第i行像素的像素值，C表示i的最大取值， x ji表示图像x中第j列第i行的像素的像素值。
可选地，压缩模块还可被设置为：通过第一网络的卷积层对第一图像执行卷积操作，并通过第一网络的池化层对第一图像执行池化操作；通过第一网络的拼接层将执行卷积操作得到的结果和执行池化操作的结果拼接起来，得到第二图像。
此处需要说明的是，上述模块与对应的步骤所实现的示例和应用场景相同，但不限于上述实施例所公开的内容。需要说明的是，上述模块作为装置的一部分可以运行在如图1所示的硬件环境中，可以通过软件实现，也可以通过硬件实现，其中，硬件环境包括网络环境。
根据本申请实施例的另一个方面，还提供了一种用于实施上述禽蛋异常的识别方法的服务器或终端。
图9是根据本申请实施例的一种终端的结构框图，如图9所示，该终端可以包括：一个或多个(图9中仅示出一个)处理器901、存储器903、以及传输装置905，如图9所示，该终端还可以包括输入输出设备907。
其中，存储器903可用于存储软件程序以及模块，如本申请实施例中的禽蛋异常的识别方法和装置对应的程序指令/模块，处理器901通过运行存储在存储器903内的软件程序以及模块，从而执行各种功能应用以及数据处理，即实现上述的禽蛋异常的识别方法。存储器903可包括高速随机存储器，还可以包括非易失性存储器，如一个或者多个磁性存储装置、闪存、或者其他非易失性固态存储器。在一些实例中，存储器903可进一步包括相对于处理器901远程设置的存储器，这些远程存储器可以通过网络连接至终端。上述网络的实例包括但不限于互联网、企业内部网、局域网、移动通信网及其组合。
上述的传输装置905用于经由一个网络接收或者发送数据，还可 以用于处理器与存储器之间的数据传输。上述的网络具体实例可包括有线网络及无线网络。在一个实例中，传输装置905包括一个网络适配器(Network Interface Controller，NIC)，其可通过网线与其他网络设备与路由器相连从而可与互联网或局域网进行通讯。在一个实例中，传输装置905为射频(Radio Frequency，RF)模块，其用于通过无线方式与互联网进行通讯。
其中，可选地，存储器903用于存储应用程序。
处理器901可以通过传输装置905调用存储器903存储的应用程序，以执行下述步骤：
获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；
响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；
利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。
处理器901还用于执行下述步骤：
通过目标模型中的第一网络对第一图像执行压缩操作，得到第二图像，其中，压缩操作包括消除第一图像中的视觉冗余信息；
通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像；
通过目标模型中的第三网络对第三图像进行异常类型的识别，确定目标禽蛋的异常类型。
采用本申请实施例，提供了一种“获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；响应于识别请求，获取对目标 禽蛋进行采集得到的第一图像；利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征”的方案，使用第一图像在空间维度上的特征和通道维度上的特征来确定异常类型，提供了基于两步Attention的禽蛋破损检测方案，加入空间注意机制和通道注意机制来优化分割结果，可以解决相关技术中对禽蛋异常的检测准确率较低的技术问题，进而达到提高禽蛋异常的检测准确率的技术效果。
可选地，本实施例中的具体示例可以参考上述实施例中所描述的示例，本实施例在此不再赘述。
本领域普通技术人员可以理解，图9所示的结构仅为示意，终端可以是智能手机(如Android手机、iOS手机等)、平板电脑、掌上电脑以及移动互联网设备(Mobile Internet Devices，MID)、PAD等终端设备。图9其并不对上述电子装置的结构造成限定。例如，终端还可包括比图9中所示更多或者更少的组件(如网络接口、显示装置等)，或者具有与图9所示不同的配置。
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令终端设备相关的硬件来完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：闪存盘、只读存储器(Read-Only Memory，ROM)、随机存取器(Random Access Memory，RAM)、磁盘或光盘等。
本申请的实施例还提供了一种存储介质。可选地，在本实施例中，上述存储介质可以用于执行禽蛋异常的识别方法的程序代码。
可选地，在本实施例中，上述存储介质可以位于上述实施例所示的网络中的多个网络设备中的至少一个网络设备上。
可选地，在本实施例中，存储介质被设置为存储用于执行以下步 骤的程序代码：
获取识别请求，其中，识别请求用于请求识别目标禽蛋的异常类型；
响应于识别请求，获取对目标禽蛋进行采集得到的第一图像；
利用目标模型识别出第一图像中目标禽蛋的异常类型，其中，目标模型在识别出目标禽蛋的异常类型时使用了第一图像在空间维度上的特征和通道维度上的特征。
可选地，存储介质还被设置为存储用于执行以下步骤的程序代码：
通过目标模型中的第一网络对第一图像执行压缩操作，得到第二图像，其中，压缩操作包括消除第一图像中的视觉冗余信息；
通过目标模型中的第二网络对第二图像进行语义分割、并提取第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像；
通过目标模型中的第三网络对第三图像进行异常类型的识别，确定目标禽蛋的异常类型。
可选地，本实施例中的具体示例可以参考上述实施例中所描述的示例，本实施例在此不再赘述。
可选地，在本实施例中，上述存储介质可以包括但不限于：U盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
上述实施例中的集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在上述计算机可读取的存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术 做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在存储介质中，包括若干指令用以使得一台或多台计算机设备(可为个人计算机、服务器或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。
在本申请的上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述的部分，可以参见其他实施例的相关描述。
在本申请所提供的几个实施例中，应该理解到，所揭露的客户端，可通过其它的方式实现。其中，以上所描述的装置实施例仅仅是示意性的，例如所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，单元或模块的间接耦合或通信连接，可以是电性或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
以上所述仅是本申请的优选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本申请的保护范围。

Claims (11)

  1. 一种禽蛋异常的识别方法，包括：
    获取识别请求，其中，所述识别请求用于请求识别目标禽蛋的异常类型；
    响应于所述识别请求，获取对所述目标禽蛋进行采集得到的第一图像；
    利用目标模型识别出所述第一图像中所述目标禽蛋的异常类型，其中，所述目标模型在识别出所述目标禽蛋的异常类型时使用了所述第一图像在空间维度上的特征和通道维度上的特征。
  2. 根据权利要求1所述的方法，其中，利用目标模型识别出所述第一图像中所述目标禽蛋的异常类型包括：
    通过所述目标模型中的第一网络对所述第一图像执行压缩操作，得到第二图像，其中，所述压缩操作包括消除所述第一图像中的视觉冗余信息；
    通过所述目标模型中的第二网络对所述第二图像进行语义分割、并提取所述第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像；
    通过所述目标模型中的第三网络对所述第三图像进行异常类型的识别，确定所述目标禽蛋的异常类型。
  3. 根据权利要求2所述的方法，其中，通过所述目标模型中的第二网络对所述第二图像进行语义分割、并提取所述第二图像中携带的空间维度上的特征和通道维度上的特征，得到所述第三图像包括：
    通过所述第二网络中的编码网络对所述第二图像进行特征编码处理，得到第四图像；
    通过所述第二网络中的解码网络对所述第四图像进行解码处理，得到所述第三图像，其中，所述第三图像中携带有所述解码网络对所述第四图像进行语义分割得到的信息和提取出的空间维度上的特征、通道维度上的特征。
  4. 根据权利要求3所述的方法，其中，通过所述第二网络中的解码网络对所述第四图像进行解码处理，得到所述第三图像包括按照如下公式确定所述第三图像中每列像素在空间维度上的像素值：
    Figure PCTCN2020129784-appb-100001
    其中，
    Figure PCTCN2020129784-appb-100002
    E j表示所述第三图像中第j列像素在空间维度上的像素值，α是空间维度的系数，A j表示所述第四图像A中第j列像素的像素值，B、C、D是利用不同的卷积核对所述第四图像A进行处理得到的图像，B i表示图像B中第i行像素的像素值，C j表示图像C中第j列像素的像素值，D i表示图像D中第i行像素的像素值，N表示i的最大取值，s ji表示根据图像B和图像C确定的图像s中第j列第i行的像素的像素值。
  5. 根据权利要求3所述的方法，其中，通过所述第二网络中的解码网络对所述第四图像进行解码处理，得到所述第三图像包括按照如下公式确定所述第三图像中每列像素在通道维度上的像素值：
    Figure PCTCN2020129784-appb-100003
    其中，
    Figure PCTCN2020129784-appb-100004
    E j表示所述第三图像中第j列像素在通道维度上的像素值，β是通道维度的系数， A j表示所述第四图像A中第j列像素的像素值，A i表示所述第四图像A的第i行像素的像素值，C表示i的最大取值，x ji表示图像x中第j列第i行的像素的像素值。
  6. 根据权利要求2所述的方法，其中，通过所述目标模型中的第一网络对所述第一图像执行压缩操作，得到第二图像包括：
    通过所述第一网络的卷积层对所述第一图像执行卷积操作，并通过所述第一网络的池化层对所述第一图像执行池化操作；
    通过所述第一网络的拼接层将执行所述卷积操作得到的结果和执行所述池化操作的结果拼接起来，得到所述第二图像。
  7. 一种禽蛋异常的识别装置，包括：
    第一获取单元，被设置为获取识别请求，其中，所述识别请求用于请求识别目标禽蛋的异常类型；
    第二获取单元，被设置为响应于所述识别请求，获取对所述目标禽蛋进行采集得到的第一图像；
    识别单元，被设置为利用目标模型识别出所述第一图像中所述目标禽蛋的异常类型，其中，所述目标模型在识别出所述目标禽蛋的异常类型时使用了所述第一图像在空间维度上的特征和通道维度上的特征。
  8. 根据权利要求7所述的装置，其中，所述识别单元包括：
    压缩模块，被设置为通过所述目标模型中的第一网络对所述第一图像执行压缩操作，得到第二图像，其中，所述压缩操作包括消除所述第一图像中的视觉冗余信息；
    编解码模块，被设置为通过所述目标模型中的第二网络对所述第二图像进行语义分割、并提取所述第二图像中携带的空间维度上的特征和通道维度上的特征，得到第三图像；
    识别模块，被设置为通过所述目标模型中的第三网络对所述第三图像进行异常类型的识别，确定所述目标禽蛋的异常类型。
  9. 一种禽蛋异常的识别系统，包括：
    图像采集设备，被设置为采集待处理的第一图像，其中，所述第一图像是对目标禽蛋进行采集得到的图像；
    服务器，被设置为在接收到获取识别请求的情况下，利用目标模型识别出所述第一图像中所述目标禽蛋的异常类型，其中，所述目标模型在识别出所述目标禽蛋的异常类型时使用了所述第一图像在空间维度上的特征和通道维度上的特征，所述识别请求被设置为请求识别目标禽蛋的异常类型。
  10. 一种存储介质，所述存储介质包括存储的程序，其中，所述程序运行时执行上述权利要求1至6任一项中所述的方法。
  11. 一种电子装置，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器通过所述计算机程序执行上述权利要求1至6任一项中所述的方法。。
PCT/CN2020/129784 2019-11-26 2020-11-18 禽蛋异常的识别方法、装置及系统、存储介质、电子装置 WO2021104125A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911175385.4 2019-11-26
CN201911175385.4A CN111160096A (zh) 2019-11-26 2019-11-26 禽蛋异常的识别方法、装置及系统、存储介质、电子装置

Publications (1)

Publication Number Publication Date
WO2021104125A1 true WO2021104125A1 (zh) 2021-06-03

Family

ID=70556148

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/129784 WO2021104125A1 (zh) 2019-11-26 2020-11-18 禽蛋异常的识别方法、装置及系统、存储介质、电子装置

Country Status (2)

Country Link
CN (1) CN111160096A (zh)
WO (1) WO2021104125A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113688645A (zh) * 2021-08-11 2021-11-23 广州爱格尔智能科技有限公司 一种识别方法、系统及设备
CN113705346A (zh) * 2021-07-22 2021-11-26 中国人民解放军陆军工程大学 一种遥感影像地物分类泛化性提升方法及装置
CN114898148A (zh) * 2022-05-10 2022-08-12 中国农业大学 一种基于深度学习的鸡蛋污损检测方法及系统
CN117390089A (zh) * 2023-11-14 2024-01-12 河北玖兴农牧发展有限公司 一种孵前种蛋受精信息统计方法及系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111160096A (zh) * 2019-11-26 2020-05-15 北京海益同展信息科技有限公司 禽蛋异常的识别方法、装置及系统、存储介质、电子装置
CN112465722A (zh) * 2020-12-04 2021-03-09 武汉大学 一种异常相位图像的修复方法

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120008868A1 (en) * 2010-07-08 2012-01-12 Compusensor Technology Corp. Video Image Event Attention and Analysis System and Method
CN110070073A (zh) * 2019-05-07 2019-07-30 国家广播电视总局广播电视科学研究院 基于注意力机制的全局特征和局部特征的行人再识别方法
CN110110689A (zh) * 2019-05-15 2019-08-09 东北大学 一种行人重识别方法
CN110309800A (zh) * 2019-07-05 2019-10-08 中国科学技术大学 一种林火烟雾探测方法及装置
CN110309880A (zh) * 2019-07-01 2019-10-08 天津工业大学 一种基于注意力机制cnn的5天和9天孵化鸡蛋胚胎图像分类方法
CN111160096A (zh) * 2019-11-26 2020-05-15 北京海益同展信息科技有限公司 禽蛋异常的识别方法、装置及系统、存储介质、电子装置
CN111178183A (zh) * 2019-12-16 2020-05-19 深圳市华尊科技股份有限公司 人脸检测方法及相关装置

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107730503B (zh) * 2017-09-12 2020-05-26 北京航空航天大学 三维特征嵌入的图像对象部件级语义分割方法与装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120008868A1 (en) * 2010-07-08 2012-01-12 Compusensor Technology Corp. Video Image Event Attention and Analysis System and Method
CN110070073A (zh) * 2019-05-07 2019-07-30 国家广播电视总局广播电视科学研究院 基于注意力机制的全局特征和局部特征的行人再识别方法
CN110110689A (zh) * 2019-05-15 2019-08-09 东北大学 一种行人重识别方法
CN110309880A (zh) * 2019-07-01 2019-10-08 天津工业大学 一种基于注意力机制cnn的5天和9天孵化鸡蛋胚胎图像分类方法
CN110309800A (zh) * 2019-07-05 2019-10-08 中国科学技术大学 一种林火烟雾探测方法及装置
CN111160096A (zh) * 2019-11-26 2020-05-15 北京海益同展信息科技有限公司 禽蛋异常的识别方法、装置及系统、存储介质、电子装置
CN111178183A (zh) * 2019-12-16 2020-05-19 深圳市华尊科技股份有限公司 人脸检测方法及相关装置

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113705346A (zh) * 2021-07-22 2021-11-26 中国人民解放军陆军工程大学 一种遥感影像地物分类泛化性提升方法及装置
CN113705346B (zh) * 2021-07-22 2023-09-19 中国人民解放军陆军工程大学 一种遥感影像地物分类泛化性提升方法及装置
CN113688645A (zh) * 2021-08-11 2021-11-23 广州爱格尔智能科技有限公司 一种识别方法、系统及设备
CN114898148A (zh) * 2022-05-10 2022-08-12 中国农业大学 一种基于深度学习的鸡蛋污损检测方法及系统
CN114898148B (zh) * 2022-05-10 2024-04-26 中国农业大学 一种基于深度学习的鸡蛋污损检测方法及系统
CN117390089A (zh) * 2023-11-14 2024-01-12 河北玖兴农牧发展有限公司 一种孵前种蛋受精信息统计方法及系统
CN117390089B (zh) * 2023-11-14 2024-03-19 河北玖兴农牧发展有限公司 一种孵前种蛋受精信息统计方法及系统

Also Published As

Publication number Publication date
CN111160096A (zh) 2020-05-15

Similar Documents

Publication Publication Date Title
WO2021104125A1 (zh) 禽蛋异常的识别方法、装置及系统、存储介质、电子装置
CN108960141B (zh) 基于增强型深度卷积神经网络的行人再识别方法
US20190156122A1 (en) Intelligent digital image scene detection
CN106897673B (zh) 一种基于retinex算法和卷积神经网络的行人再识别方法
WO2019033715A1 (zh) 人脸图像数据采集方法、装置、终端设备及存储介质
CN111178197A (zh) 基于Mask R-CNN和Soft-NMS融合的群养粘连猪实例分割方法
US20230334890A1 (en) Pedestrian re-identification method and device
CN110689021A (zh) 一种基于深度学习的低可见度环境下实时目标检测方法
Zhao et al. SEV‐Net: Residual network embedded with attention mechanism for plant disease severity detection
EP3698268A1 (en) Methods and systems for face recognition
WO2021104124A1 (zh) 圈养栏信息的确定方法、装置及系统、存储介质
CN115331087A (zh) 融合区域语义与像素特征的遥感影像变化检测方法及系统
CN113034506B (zh) 遥感图像语义分割方法、装置、计算机设备和存储介质
CN110046568B (zh) 一种基于时间感知结构的视频动作识别方法
Hu et al. LE–MSFE–DDNet: a defect detection network based on low-light enhancement and multi-scale feature extraction
CN114333062B (zh) 基于异构双网络和特征一致性的行人重识别模型训练方法
Xie et al. Recognition of big mammal species in airborne thermal imaging based on YOLO V5 algorithm
CN116206334A (zh) 一种野生动物识别方法和装置
Cui et al. Real-time detection of wood defects based on SPP-improved YOLO algorithm
CN116778346B (zh) 一种基于改进自注意力机制的管线识别方法及系统
CN113449676A (zh) 一种基于双路互促进解纠缠学习的行人重识别方法
CN116503399A (zh) 基于yolo-afps的绝缘子污闪检测方法
CN111079930A (zh) 数据集质量参数的确定方法、装置及电子设备
CN116151323A (zh) 模型生成方法、装置、电子设备及存储介质
CN115641445A (zh) 一种非对称内卷积与Transformer相融合的遥感图像阴影检测方法

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: 20891867

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: 20891867

Country of ref document: EP

Kind code of ref document: A1