TWI810946B - Method for identifying image, computer device and storage medium - Google Patents

Abstract

The present application relates to an image analysis technology, and provides a method for identifying image, a computer device and a storage medium. The method includes obtaining an image to be identified, an initial annotation image and an initial annotation result of the initial annotation image, comparing the initial annotation result with a preset annotation result; obtaining a target image and an annotation result of the target image by inputting the initial annotation image into a first pre-constructed semantic segmentation network in response that the initial annotation result is different from the preset annotation result; and obtaining a second semantic segmentation network by training the first semantic segmentation network based on the target image and the annotation result of the target image. Finally, an image annotation result of the image to be identified is obtained by inputting the image to be identified into the second semantic segmentation network. This application can improve the accuracy of image recognition.

Description

Image recognition method, computer equipment and storage medium

The present application relates to the field of image processing, in particular to an image recognition method, computer equipment and storage media.

In the current image recognition scheme, due to the large number of annotation dimensions of the training image, and the small number of training images for each annotation dimension, the generalization ability of the trained model will be poor, resulting in image recognition accuracy is low.

In view of the above, it is necessary to provide an image recognition method, computer equipment and storage medium, which can improve the accuracy of image recognition.

An image recognition method, the image recognition method comprising: Obtaining the image to be identified, and obtaining an initial labeling image and an initial labeling result of the initial labeling image; constructing a first semantic segmentation network; comparing the initial labeling result with a preset labeling result to obtain a comparison result; If the comparison result is that the initial labeling result is different from the preset labeling result, the initial labeling image is input into the first semantic segmentation network to obtain the corresponding initial labeling image. The target image of the target image and the target labeling result of the target image; the first semantic segmentation network is trained based on multiple target images and the target labeling results of each target image to obtain the second semantic segmentation A network; inputting the image to be recognized into the second semantic segmentation network to obtain an image annotation result of the image to be recognized.

According to an optional embodiment of the present application, before comparing the initial labeling result with the preset labeling result, the method further includes: acquiring a plurality of preset color labels, a plurality of preset values and a plurality of preset categories ; establishing a corresponding relationship between each preset value and each color label to obtain a plurality of target labels; establishing a corresponding relationship between each preset category and each target label to obtain the preset labeling result.

According to an optional embodiment of the present application, the initial labeling result includes a plurality of initial categories and initial labels corresponding to each initial category, and comparing the initial labeling result with a preset labeling result to obtain the comparison result includes: obtaining Each initial object in the initial annotation image, and determine the initial category corresponding to each initial object; if there is no corresponding preset category for any initial category, determine that the comparison result is the initial labeling result and the initial category The preset labeling results are different; or, if each initial category has a corresponding preset category and the initial label corresponding to each initial category is different from the target label of the corresponding preset category, then determine the comparison result The initial marking result is different from the preset marking result.

According to an optional embodiment of the present application, the first semantic segmentation network includes an autoencoder and a classifier, and the initial labeled image is input into the first semantic segmentation network to obtain the initial Labeling the target image corresponding to the image and the target labeling result of the target image includes: using the self-encoder to process the initial labeling image to generate a target feature map; based on the classifier to classify the target feature map Each pixel in is classified to obtain the labeling category corresponding to each pixel; based on the labeling category corresponding to each pixel and the target label corresponding to the labeling category, the target feature map is labeled to generate the The target image and the target labeling result.

According to an optional embodiment of the present application, the self-encoder includes a plurality of cascaded structures and decoders, and using the self-encoder to process the initial labeled image and generating the target feature map includes: based on any cascaded structure A plurality of hidden layers in the method extract features from the initial labeled image to obtain an initial feature map output by the last hidden layer; a pooling layer based on any cascaded structure performs a pooling operation on the initial feature map, Output the first feature map; output the current cascade structure The first feature map is input to the next cascade structure until the first feature map output by the last cascade structure is obtained as the second feature map; the pixel value of each pixel point in multiple first feature maps is obtained in The first pixel position in the corresponding initial feature map, and obtain the second pixel position of the pixel value of each pixel in the second feature map in the corresponding initial feature map; based on the decoder, the first A pixel position and the second pixel position perform a decoding operation on the second feature map to obtain the target feature map.

According to an optional embodiment of the present application, classifying each pixel in the target feature map based on the classifier, and obtaining a label category corresponding to each pixel includes: Calculating a single score value for each pixel point in the target feature map based on the pixel value of each pixel point in the target feature map; calculating the single score value based on the single score value and the plurality of preset categories The category probability that the corresponding pixel belongs to each preset category; the preset category corresponding to the category probability with the largest value is determined as the labeled category corresponding to the pixel.

，i=1,2,...,k； 其中，S _i 表示每個像素點屬於第i個預設類別的類別概率，

表示所述目標特徵圖中的第j個像素點的單個評分值，z _j 表示所述目標特徵圖中的第j個像素點的像素值，

表示所述目標特徵圖中所有像素點的總評分值，i表示所述第i個預設類別，k表示所述多個預設類別的數量。 According to an optional embodiment of the present application, the formula for calculating the category probability is:

, i=1,2,...,k; Among them, S _i represents the category probability that each pixel belongs to the i-th preset category,

Represents the single scoring value of the jth pixel in the target feature map, z _j represents the pixel value of the jth pixel in the target feature map,

represents the total scoring value of all pixels in the target feature map, i represents the ith preset category, and k represents the number of the multiple preset categories.

According to an optional embodiment of the present application, the target feature map is marked based on the labeling category corresponding to each pixel point and the target label corresponding to the labeling category, and the target image and the target labeling result are generated. Including: determining the area formed by all the pixels corresponding to the same label category in the target feature map as the feature area; adjusting the pixel values of all the pixels in the feature area to the preset value corresponding to the same label category numerical value; according to the color label corresponding to the preset numerical value of the feature area, coloring is performed on each pixel in the feature area to obtain a destination area; according to multiple destination areas in the target feature map The area positions of the multiple destination areas are spliced together to obtain the Target image: determining the preset value, color label and labeling category corresponding to each destination area in the target image as the target labeling result.

The present application provides a computer device, which includes: a memory storing at least one instruction; and a processor executing the at least one instruction to implement the image recognition method.

The present application provides a computer-readable storage medium, at least one instruction is stored in the computer-readable storage medium, and the at least one instruction is executed by a processor in a computer device to implement the image recognition method.

It can be seen from the above technical solutions that since the initial labeled images in this application come from different data sets, the initial labeling results for the same target object between any two initial labeled images from different data sets are different (that is, any The same target object in the two initial annotation images is marked in different ways), and the initial annotation image is re-labeled through the first semantic segmentation network, because the first semantic segmentation network is based on the Therefore, the target image generated by the first semantic segmentation network can have the same target labeling result for the same target object (that is, the target image is unified. the labeling method of the same target object in the image), and then based on the multiple target images and the target labeling results of each target image, the first semantic segmentation network is trained to obtain the second semantic segmentation network, Due to the increase of the training data of the first semantic segmentation network and the unification of the labeling methods of the training images, the generated second semantic segmentation network has better generalization ability, thereby improving the image quality. Accuracy of labeling results.

1: Computer equipment

12: Storage

13: Processor

S10~S15: Steps

Fig. 1 is a flowchart of an image recognition method in a preferred embodiment of the present application.

Fig. 2 is a schematic diagram of an initially labeled image in the image recognition method of the present application.

Fig. 3 is a schematic structural diagram of a computer device implementing a preferred embodiment of the image recognition method of the present application.

In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

The image recognition method can be applied to one or more computer devices 1 . The computer device 1 is a device that can automatically perform parameter value calculation and/or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to: microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate array (Field-Programmable Gate Array, FPGA), digital signal processor (Digital Signal Processor, DSP), embedded devices, etc.

The computer device 1 can be any electronic product capable of man-machine interaction with the user, for example, a personal computer, a tablet computer, a smart phone, a personal digital assistant (Personal Digital Assistant, PDA), a game console, an interactive Internet TV (Internet Protocol Television, IPTV), smart wearable devices, etc.

The computer equipment 1 may also include network equipment and/or user equipment. Wherein, the network device includes, but is not limited to, a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of hosts or network servers based on Cloud Computing.

The network where the computer device 1 is located includes, but is not limited to: Internet, wide area network, metropolitan area network, local area network, virtual private network (Virtual Private Network, VPN) and so on.

As shown in FIG. 1 , it is a flowchart of a preferred embodiment of an image recognition method of the present application. According to different requirements, the order of each step in the flowchart can be adjusted according to actual detection requirements, and some steps can be omitted. The execution body of the method is a computer device, such as the computer device 1 shown in FIG. 1 .

Step S10, acquiring an image to be recognized, and acquiring an initial labeled image and an initial labeling result of the initial labeled image.

In at least one embodiment of the present application, the initial labeled image refers to the initial Multiple images with known labeling results, the initial labeled image contains multiple initial objects, wherein the same initial object in different data sets corresponds to multiple initial categories, multiple initial colors and multiple initial numbers . For example: the first initial labeled image is obtained from data set A, and the second initial labeled image is obtained from data set B, the first initial labeled image and the second initial labeled image both contain the initial object car, The initial category of the initial object car in the first initial annotation image is a car and a bus, wherein the initial color of the car is blue, the initial number of the car is 3, and the initial color of the bus is Purple, the initial number of the bus is 5; the initial category of the initial object car in the second initial label image is a motorcycle and a truck, wherein the initial color of the motorcycle is blue, the motorcycle The initial number of the car is 1, the initial color of the truck is black, and the initial number of the truck is 8.

In at least one embodiment of the present application, the initial labeling result refers to the correspondence between the initial category of each initial object in the initial labeling image and each initial label, and each initial label refers to the The correspondence between the initial color of each initial object in the initial annotation image and the initial number of each initial object.

The initial categories of the initial objects may be: cars, trees, utility poles, pedestrians, double yellow lines, white lines and so on. The initial color may be: gray, red, orange, yellow, yellow-green, brown, light blue, and the like. For example, the initial objects in any initial labeled image are: people, cars, trees, road facilities, roads, lane lines, and the initial categories corresponding to each initial object are: pedestrians, cars, trees, utility poles, roads, Double yellow lines and white lines, the initial labeling results of any of the initial labeling images are shown in Table 1:

In one embodiment of the present application, the computer device obtains the initial tagged image and the initial tagged result of the initial tagged image from databases such as KITTI, Mapillary, CityScapes, and Daimler Urban.

As shown in Figure 2, it is the initial labeled image obtained by the application from the above-mentioned multiple databases. categories, but only the initial colors and initial numbers of some initial categories are listed, and the initial colors and initial numbers of each initial category in the partial initial categories are different.

In order to facilitate the presentation of the initial categories, initial colors and initial numbers corresponding to different initial objects (such as people, cars, roads, etc.), multiple dashed boxes are used in Figure 2 to illustrate the initial categories, initial colors and initial numbers corresponding to each initial object. The initial number and the text in the dotted line box are only exemplary descriptions of each initial object in the figure. In practical applications, the initial label image does not include the dotted line box. In at least one embodiment of the present application, the image to be recognized includes a plurality of objects to be recognized, and the image to be recognized refers to an image in which information such as the category and color of each object to be recognized is not marked image. In at least one embodiment of the present application, the computer device acquires images captured by a driving recorder or an in-vehicle camera to obtain the image to be recognized.

Step S11, constructing a first semantic segmentation network.

In at least one embodiment of the present application, the first semantic segmentation network refers to a network for re-labeling the initially labeled image.

In at least one embodiment of the present application, the first semantic segmentation network includes an autoencoder and a classifier. Wherein, the self-encoder includes multiple cascaded structures and decoders, each cascaded structure includes multiple hidden layers and pooling layers, and the decoder includes multiple cascaded structures, each cascaded structure includes anti-pooling layer and a plurality of computing layers, the computer device constructing the first semantic segmentation network includes: the computer device constructs a cascade structure of multiple hidden layers and pooling layers, and uses multiple cascade structures as encoders , the computer equipment constructs the anti-pooling layer and the series connection of multiple computing layers structure, using multiple serial structures as decoders, and constructing a classifier, further, the computer device generates a learner based on the encoder, the decoder and the classifier, further, the The computer device acquires training images, and trains the learner based on the training images, and calculates a loss value of the learner until the loss value is less than a preset value, and obtains the first semantic segmentation network road. Wherein, each hidden layer includes multiple convolution layers (Convolution Layer), batch standard layer (Batch Normalization Layer) and activation function layer (Activation Function Layer), and the pooling layer is a maximum pooling layer, and each operation layer includes A plurality of deconvolution layers, the batch standard layer and the activation function layer, the activation function layer may be a ReLu linear rectification function.

Wherein, the encoder refers to a network that extracts features from the initial labeled image, and the decoder refers to a network that restores the features extracted by the encoder, and the encoder and the decoder It is a symmetrical structure, and the decoding process is the reverse process of the encoding process. In other embodiments of the present application, the first semantic segmentation network may also be networks such as U-Net, DeepLab v1, DeepLab v2, and Mask R-CNN, which are not limited in this application.

Step S12, comparing the initial labeling result with the preset labeling result to obtain a comparison result.

In at least one embodiment of the present application, the preset labeling result refers to a correspondence between each preset category, each preset color label, and each preset value.

In at least one embodiment of the present application, the comparison result includes: the initial tagging result is the same as the preset tagging result, or the initial tagging result is different from the preset tagging result.

In at least one embodiment of the present application, before comparing the initial tagging result with the preset tagging result, the method further includes: the computer device acquires a plurality of preset color tags and a plurality of preset values and a plurality of preset categories, further, the computer equipment establishes the corresponding relationship between each preset value and each color label, and obtains multiple target labels, and further, the computer equipment establishes the correspondence between each preset category and each color label The corresponding relationship of each target label is used to obtain the preset labeling result. Wherein, the plurality of color tags include: red, green, blue, gray, etc., the plurality of preset values Can be: 1, 2, 3, 4, etc.

The plurality of preset categories may be: lanes, lane markings, crossings, zebra crossings, curbs, trees, sidewalks, cars, bicycles, motorcycles, pedestrians, baby carriages, large vehicles, trucks, traffic lights, traffic signals, road signs , buildings, street lights, utility poles, static and dynamic obstacles, etc. It can be understood that the preset category includes the category of each target object appearing on the road as much as possible. For example, the target objects are: people, cars, trees, road facilities, roads, lane lines, obstacles, and the initial categories corresponding to each target object are: pedestrians, cars, trucks, trucks, trees, utility poles, roads, lane lines , dynamic obstacles and static obstacles.

For example, the preset labeling results can be shown in Table 2:

In at least one embodiment of the present application, the initial labeling result includes a plurality of initial categories and initial labels corresponding to each initial category, and comparing the initial labeling result with the preset labeling result, the comparison result includes : The computer device obtains each initial object in the initial labeled image, and determines The initial category corresponding to each initial object, if any initial category does not have a corresponding preset category, the computer device determines that the comparison result is that the initial labeling result is different from the preset labeling result, or, if There is a corresponding preset category for each initial category and the initial label corresponding to each initial category is different from the target label of the corresponding preset category, and the computer device determines that the comparison result is that the initial labeling result and the The preset labeling results are different. In this embodiment, when there is no corresponding preset category for any initial category, it includes the following situation: the initial object corresponding to any initial category is the same as the corresponding target object in the plurality of preset categories , or the initial object corresponding to any initial category is different from the target objects corresponding to the multiple preset categories.

Further, when the initial object corresponding to any initial category is the same as the target object corresponding to the plurality of preset categories, there are many situations, for example, the naming of the initial category and the preset category Inconsistent, so that the initial category does not have a corresponding preset category, or, the classification of the preset category is too low, so that the initial category does not have a corresponding preset category, or, the preset category is too high, So that the initial category does not have a corresponding preset category. The difference between the initial label corresponding to each initial category and the target label of the corresponding preset category means that the correspondence between the initial color label and the initial value is inconsistent with the correspondence between the preset value and the color label.

In this embodiment, if there is a corresponding preset category for each initial category, and the initial label corresponding to each initial category is the same as the target label of the corresponding preset category, the computer device determines that the comparison result is The initial tagging result is the same as the preset tagging result. Through the above implementation, it can be judged whether the initial labeling result of any initial labeling image is the same as the preset labeling result, and when the initial labeling result of the initial labeling image is the same as the preset labeling result, there is no need to Relabeling the initially labeled image can improve labeling efficiency.

Following the above example, comparing Table 1 with Table 2, it can be seen that the road, double yellow line and white line in the initial labeling results of Table 1 do not have corresponding preset categories in Table 2. After comparison, the road is in the The preset category of the preset labeling result is road, and the preset category of double yellow lines and white lines in the preset labeling result is lane line, so the initial labeling results in Table 1 are different from the preset labeling results in Table 2.

Step S13, if the comparison result is the initial labeling result and the preset labeling If the results are not the same, then input the initial tagged image into the first semantic segmentation network to obtain the target image corresponding to the initial tagged image and the target tagging result of the target image.

In at least one embodiment of the present application, the target image refers to an image generated after re-labeling an initial labeled image according to the first semantic segmentation network.

In at least one embodiment of the present application, the computer device inputs the initial tagged image into the first semantic segmentation network to obtain the target image corresponding to the initial tagged image and the target The target labeling result of the image includes: the computer device uses the self-encoder to process the initial label image to generate a target feature map, and further, the computer device classifies the target feature map based on the classifier Each pixel is classified to obtain the labeling category corresponding to each pixel. Further, the computer device classifies the target feature based on the labeling category corresponding to each pixel and the target label corresponding to the labeling category. Annotate the image to generate the target image and the target labeling result.

Through the above-mentioned implementation manner, based on the self-encoder, the initial labeled image is compressed and decompressed to obtain the target feature map. Since the image noise of the initial labeled image is filtered during the compression process, The generated target feature map can be made clearer, and multiple initial labeled images corresponding to initial labeling results different from the preset labeling result are relabeled through the preset labeling result to obtain multiple target images image, it can ensure that the labeling results of the multiple target images for the same object are unified.

Specifically, the computer device uses the self-encoder to process the initial tagged image, and generating the target feature map includes: the computer device performs the initial tagged image based on multiple hidden layers in any cascaded structure The feature extraction is to obtain the initial feature map output by the last hidden layer, further, the computer device performs pooling operation on the initial feature map based on the pooling layer of any cascaded structure, and outputs the first feature map, and further Further, the computer device inputs the first feature map output by the current cascade structure into the next cascade structure until the first feature map output by the last cascade structure is obtained as the second feature map, and the computer device Obtain the first pixel position of the pixel value of each pixel in the first feature map in the corresponding initial feature map, and obtain each pixel value in the second feature map The pixel value of the pixel point is in the second pixel position in the corresponding initial feature map, further, the computer device calculates the second feature based on the decoder, the first pixel position and the second pixel position The map is decoded to obtain the target feature map.

Specifically, the computer device performs a decoding operation on the second feature map based on the decoder, the first pixel position, and the second pixel position, and obtaining the target feature map includes: the computer device according to The size of the initial feature map output by each cascaded structure constructs an all-zero feature map of a corresponding size, and fills each pixel value in the second feature map into the corresponding all-zero feature map according to the second pixel position In the method, the third feature map output by the unpooling layer in the first series structure is obtained, and further, the computer device performs calculation on the third feature map based on the multiple operation layers in the first series structure Deconvolution operation to obtain the fourth feature map output by the first series structure, further, the computer device inputs the fourth feature map into the next series structure, and based on the fourth feature map Figure, the corresponding first pixel position and multiple operation layers in the next series structure generate the fifth feature map of the next series structure, and the computer device obtains the fifth feature map output by the last series structure as The target feature map.

In this embodiment, the second feature map is decoded based on the first pixel position and the second pixel position to obtain the target feature map. Since more pixel position information is retained, the The features contained in the target feature map are more complete.

Specifically, the computer device classifies each pixel in the target feature map based on the classifier, and obtaining the label category corresponding to each pixel includes: the computer device classifies each pixel in the target feature map based on Calculate the single score value of each pixel point in the target feature map based on the pixel value of each pixel point, and further, the computer device calculates the corresponding score value of the single score value based on the single score value and the plurality of preset categories The pixel points belong to the category probability of each preset category, and further, the computer device determines the preset category corresponding to the category probability with the largest value as the label category corresponding to the pixel point.

，i=1,2,...,k； 其中，S _i 表示每個像素點屬於第i個預設類別的類別概率，

表示所述目標特徵圖中的第j個像素點的單個評分值，z _j 表示所述目標特徵圖中的第j個像素點的像素值，

表示所述目標特徵圖中所有像素點的總評分值，i表示所述第i個預設類別，k表示所述多個預設類別的數量。 In this embodiment, the preset category corresponding to the category probability with the largest value is determined as the labeled category corresponding to the pixel, which can improve the classification accuracy of each pixel. Specifically, the formula for calculating the class probability is:

, i=1,2,...,k; Among them, S _i represents the category probability that each pixel belongs to the i-th preset category,

Represents the single scoring value of the jth pixel in the target feature map, z _j represents the pixel value of the jth pixel in the target feature map,

represents the total scoring value of all pixels in the target feature map, i represents the ith preset category, and k represents the number of the multiple preset categories.

Specifically, the computer device uses the pixel value of each pixel as the logarithm of an exp exponential function to obtain a single scoring value for each pixel, and the exp exponential function refers to an exponential function with e as the base.

Specifically, the computer device labels the target feature map based on the labeling category corresponding to each pixel point and the target label corresponding to the labeling category, and generating the target image and the target labeling result includes: the The computer device determines the area formed by all the pixels corresponding to the same label category in the target feature map as the feature area, and further, the computer device adjusts the pixel values of all the pixels in the feature area to the same Labeling the preset value corresponding to the category, further, the computer device performs coloring processing on each pixel in the feature area according to the color label corresponding to the preset value of the feature area to obtain the destination area , further, the computer device stitches the multiple destination areas according to the area positions of the target feature map to obtain the target image, further, the computer device The preset value, color label and labeling category corresponding to each destination area in the target image are determined as the target labeling result.

In this embodiment, according to the preset labeling result, the pixel value of each pixel in the feature region and the value of each pixel can be quickly adjusted, so that each region in the target image is more prominent .

Following the above example, the target labeling results of any one of the initial labeling images are shown in Table 3:

Step S14: Train the first semantic segmentation network based on multiple target images and the target labeling results of each target image to obtain a second semantic segmentation network.

In at least one embodiment of the present application, the second semantic segmentation network refers to using the multiple target images and the target labeling results of each target image to train the first semantic segmentation network to generate network. In at least one embodiment of the present application, the process of generating the second semantic segmentation network is basically the same as the process of generating the first semantic segmentation network, which will not be described in detail in this application.

Through the above implementation, the first semantic segmentation network is trained using the plurality of target images and the target labeling results corresponding to each target image to obtain the second semantic segmentation image. By adding the training image The number of images makes the second semantic segmentation network have higher recognition accuracy. Since the labeling results of the multiple target images for the same object are uniform, the second semantic segmentation network It can automatically and uniformly mark multiple images to be recognized.

Step S15, inputting the image to be recognized into the second semantic segmentation network to obtain an image annotation result of the image to be recognized.

In at least one embodiment of the present application, the image annotation result includes the category of each object to be recognized, the number of each object to be recognized and the color label of each object to be recognized. In at least one embodiment of the present application, the generation process of the image annotation result is related to the generation of the target annotation result The process is basically the same, and this application does not limit it.

In this embodiment, since the image to be recognized is an image captured by a driving recorder or an in-vehicle camera, each object to be recognized in the image to be recognized can be analyzed by using the second semantic segmentation network Accurate identification is performed to obtain the image annotation results, and the driver can obtain accurate road condition information according to the plurality of image annotation results during driving, thereby improving driving safety.

It can be seen from the above technical solutions that since the initial labeled images in this application come from different data sets, the initial labeling results for the same target object between any two initial labeled images from different data sets are different (that is, any The same target object in the two initial annotation images is marked in different ways), and the initial annotation image is re-labeled through the first semantic segmentation network, because the first semantic segmentation network is based on the Therefore, the target image generated by the first semantic segmentation network can have the same target labeling result for the same target object (that is, the target image is unified. the labeling method of the same target object in the image), and then based on the multiple target images and the target labeling results of each target image, the first semantic segmentation network is trained to obtain the second semantic segmentation network, Due to the increase of the training data of the first semantic segmentation network and the unification of the labeling methods of the training images, the generated second semantic segmentation network has better generalization ability, thereby improving the image quality. Accuracy of labeling results.

As shown in FIG. 3 , it is a schematic structural diagram of a computer device implementing a preferred embodiment of the image recognition method of the present application.

In one embodiment of the present application, the computer device 1 includes, but is not limited to, a storage 12, a processor 13, and a computer program stored in the storage 12 and operable on the processor 13 , such as an image recognition program.

Those skilled in the art can understand that the schematic diagram is only an example of the computer device 1 and does not constitute a limitation to the computer device 1. It may include more or less components than those shown in the illustration, or combine certain components, or have different Components, for example, the computer device 1 may also include input and output devices, network access devices, bus bars, and the like. The processor 13 can be a central processing unit (Central Processing Unit, CPU), and can also be other general processors, digital signal processors (Digital Signal Processor, DSP), Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., and the processor 13 is the computing core and control center of the computer device 1, and utilizes various interfaces and lines to connect the entire computer device 1, and obtain the operating system of the computer device 1 and various installed applications, program codes, etc.

The processor 13 acquires the operating system of the computer device 1 and various installed applications. The processor 13 acquires the application program to implement the steps in the above embodiments of the image recognition method, for example, the steps shown in FIG. 1 .

Exemplarily, the computer program can be divided into one or more modules/units, and the one or more modules/units are stored in the storage 12 and acquired by the processor 13, to complete this application. The one or more modules/units may be a series of computer program instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the acquisition process of the computer program in the computer device 1 .

The storage 12 can be used to store the computer programs and/or modules, and the processor 13 executes or obtains the computer programs and/or modules stored in the storage 12, and calls the computer programs and/or modules stored in the storage 12 to realize various functions of the computer device 1. The storage device 12 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playback function, an image playback function, etc.); The area can store data created according to the use of the computer equipment, etc. In addition, the storage 12 may include non-volatile storage, such as hard disk, memory, plug-in hard disk, smart memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card, flash memory A memory card (Flash Card), at least one magnetic disk storage device, a flash memory device, or other non-volatile solid state storage devices.

The storage 12 may be an external storage and/or an internal storage of the computer device 1 . Further, the storage 12 can be a storage in physical form, such as a storage bar, a TF card (Trans-flash Card) and so on.

If the integrated modules/units of the computer device 1 are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments in the present application can also be completed by instructing related hardware through computer programs, and the computer programs can be stored in a computer-readable storage medium, so When the computer program is acquired by the processor, it can realize the steps of the above-mentioned various method embodiments.

Wherein, the computer program includes computer program code, and the computer program code may be in the form of original code, object code, obtainable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a flash drive, a removable hard disk, a magnetic disk, an optical disk, a computer storage, a read-only memory (ROM, Read- Only Memory).

Referring to FIG. 1, the memory 12 in the computer device 1 stores a plurality of instructions to implement an image recognition method, and the processor 13 can acquire the plurality of instructions to realize: acquire an image to be identified, and Obtaining an initial labeling image and an initial labeling result of the initial labeling image; constructing a first semantic segmentation network; comparing the initial labeling result with a preset labeling result to obtain a comparison result; if the comparison result is the If the initial labeling result is different from the preset labeling result, the initial labeling image is input into the first semantic segmentation network to obtain the target image corresponding to the initial labeling image and the The target labeling result of the target image; the first semantic segmentation network is trained based on the target labeling results of multiple target images and each target image to obtain the second semantic segmentation network; The recognized image is input into the second semantic segmentation network to obtain an image annotation result of the image to be recognized. Specifically, for the specific implementation method of the above instruction by the processor 13, reference may be made to the description of the relevant steps in the embodiment corresponding to FIG. 2 , which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, they may be located in one place, or may also be distributed to multiple networks on the unit. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional module in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented not only in the form of hardware, but also in the form of hardware plus software function modules.

Therefore, no matter from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the application is defined by the appended claims rather than the above description, so it is intended to All changes within the meaning and range of equivalents of the elements are embraced in this application. Any attached reference mark in a claim shall not be deemed to limit the claim to which it relates.

In addition, it is obvious that the word "comprising" does not exclude other elements or steps, and the singular does not exclude the plural. A plurality of units or devices stated in this application may also be realized by one unit or device through software or hardware. The terms first, second, etc. are used to denote names and do not imply any particular order.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application without limitation. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be Make modifications or equivalent replacements without departing from the spirit and scope of the technical solutions of the present application.

S10~S15: Steps

Claims (9)

An image recognition method applied to computer equipment, wherein the image recognition method includes: obtaining an image to be recognized, and obtaining an initial tagged image and an initial tagged result of the initial tagged image; constructing a first semantic segmentation Network; obtain preset multiple color labels, multiple preset values and multiple preset categories, establish the corresponding relationship between each preset value and each color label, obtain multiple target labels, and establish each preset The corresponding relationship between the category and each target label to obtain the preset labeling result; compare the initial labeling result with the preset labeling result to obtain the comparison result; if the comparison result is the initial labeling result and the If the preset labeling results are different, the initial labeling image is input into the first semantic segmentation network to obtain the target image corresponding to the initial labeling image and the target labeling result of the target image ; Based on a plurality of target images and the target labeling results of each target image, the first semantic segmentation network is trained to obtain a second semantic segmentation network; the image to be recognized is input to the In the second semantic segmentation network, an image annotation result of the image to be recognized is obtained. The image recognition method according to claim 1, wherein the initial labeling result includes a plurality of initial categories and an initial label corresponding to each initial category, and comparing the initial labeling result with a preset labeling result, Obtaining the comparison result includes: obtaining each initial object in the initial marked image, and determining the initial category corresponding to each initial object; if any initial category does not have a corresponding preset category, determining that the comparison result is the The initial labeling result is different from the preset labeling result; or if each initial category has a corresponding preset category and the initial label corresponding to each initial category is different from the target label of the corresponding preset category, then determining that the comparison result is the initial labeling result The result is different from the preset labeling result. The image recognition method according to claim 1, wherein the first semantic segmentation network includes an autoencoder and a classifier, and the initial labeled image is input into the first semantic segmentation network Obtaining the target image corresponding to the initial labeled image and the target labeling result of the target image includes: using the self-encoder to process the initial labeled image to generate a target feature map; based on the classifier Classifying each pixel in the target feature map to obtain a label category corresponding to each pixel; based on the label category corresponding to each pixel point and the target label corresponding to the label category Marking is performed to generate the target image and the target marking result. The image recognition method according to claim 3, wherein the self-encoder includes a plurality of cascaded structures and decoders, and the process of using the self-encoder to process the initial labeled image to generate a target feature map includes : Based on multiple hidden layers in any cascade structure, feature extraction is performed on the initial labeled image to obtain an initial feature map output by the last hidden layer; based on the pooling layer of any cascade structure, the initial feature The graph is pooled and the first feature map is output; the first feature map output by the current cascade structure is input to the next cascade structure until the first feature map output by the last cascade structure is obtained as the second feature map ; Obtain the first pixel position of the pixel value of each pixel in the first feature map in the corresponding initial feature map, and obtain the pixel value of each pixel in the second feature map in the corresponding A second pixel position in the initial feature map; performing a decoding operation on the second feature map based on the decoder, the first pixel position, and the second pixel position to obtain the target feature map. The image recognition method according to claim 3 or 4, wherein the classifier is used to classify each pixel in the target feature map to obtain the label corresponding to each pixel The classification includes: calculating a single score value of each pixel point in the target feature map based on the pixel value of each pixel point in the target feature map; based on the single score value and the plurality of preset categories, calculating the The category probability that the pixel point corresponding to the single scoring value belongs to each preset category; the preset category corresponding to the category probability with the largest value is determined as the labeled category corresponding to the pixel point. The image recognition method as described in claim item 5, wherein the calculation formula of the category probability is:

, i=1,2,...,k; among them, S _i represents the category probability that each pixel belongs to the i-th preset category,

Represents the single scoring value of the jth pixel in the target feature map, z _j represents the pixel value of the jth pixel in the target feature map,

represents the total scoring value of all pixels in the target feature map, i represents the ith preset category, and k represents the number of the multiple preset categories. The image recognition method according to claim 4, wherein the target feature map is marked based on the label category corresponding to each pixel point and the target label corresponding to the label category to generate the target image And the target labeling result includes: determining the area formed by all pixels corresponding to the same label category in the target feature map as a feature area; adjusting the pixel values of all pixels in the feature area to the same label The preset value corresponding to the category; coloring each pixel in the feature area according to the color label corresponding to the preset value of the feature area to obtain a destination area; according to a plurality of the destination areas Splicing the plurality of destination areas at the area positions of the target feature map to obtain the target image; Don't be sure to label the result for said goal. A computer device, wherein the computer device includes: a memory storing at least one instruction; and a processor obtaining the instruction stored in the memory to realize the image as described in any one of claims 1 to 7 recognition methods. A computer-readable storage medium, wherein: at least one instruction is stored in the computer-readable storage medium, and the at least one instruction is executed by a processor in a computer device to implement any one of claims 1 to 7. image recognition method.