CN108764325A - Image-recognizing method, device, computer equipment and storage medium - Google Patents

Abstract

The present invention relates to an image recognition method, device, computer equipment and storage medium. The method includes: obtaining a current image of an object to be recognized; obtaining a current pixel point from the current image, and determining according to the position of the current pixel point The corresponding background reference area; calculate the background similarity corresponding to the current pixel according to the background reference area; process the image feature corresponding to the current pixel according to the trained image target recognition model to obtain the current pixel The first probability corresponding to the point, the first probability is the probability that the current pixel belongs to the pixel of the target object; the background similarity and the first probability corresponding to each pixel in the current image are calculated, according to the The background similarity corresponding to each pixel and the first probability identify the image area where the target object is located in the current image. The image recognition accuracy of the above method is high.

Description

Image recognition method, device, computer equipment and storage medium

technical field

The invention relates to the field of image processing, in particular to an image recognition method, device, computer equipment and storage medium.

Background technique

With the development of science and technology, more and more information is included in the image. In order to obtain the content in the image, it is necessary to identify and process the image content. target.

At present, when it is necessary to identify the target object in the image, the area whose gray value is greater than a certain threshold is often regarded as the area where the target is located according to the relatively small gray value of the background of the image. However, the gray value of the background may also be large or the target The gray value of the image may also be very small, therefore, the method of identifying the target object in the image only based on the gray value is not accurate.

Contents of the invention

Based on this, it is necessary to provide an image recognition method, device, computer equipment, and storage medium for the above-mentioned problems. Since the background similarity can reflect whether a pixel is a background, the probability obtained by using the model positively reflects whether a pixel is a target. , combined with the background similarity corresponding to the pixel and the probability that the pixel belongs to the pixel of the target object to identify the image area where the target object is located, the image recognition accuracy is high.

An image recognition method, the method comprising: acquiring a current image of an object to be identified; acquiring a current pixel from the current image, determining a corresponding background reference area according to the position of the current pixel; Calculate the background similarity corresponding to the current pixel in the region; process the image feature corresponding to the current pixel according to the trained image target recognition model to obtain the first probability corresponding to the current pixel, the first The probability is the probability that the current pixel belongs to the pixel of the target object; the background similarity and the first probability corresponding to each pixel in the current image are calculated, and according to the background similarity and the first probability corresponding to each pixel The image region where the target object is located in the current image is obtained through probabilistic identification.

An image recognition device, the device comprising: a current image acquisition module, configured to acquire a current image of an object to be identified; a background area determination module, configured to acquire a current pixel from the current image, and according to the current pixel The position determines the corresponding background reference area; the similarity calculation module is used to calculate the background similarity corresponding to the current pixel point according to the background reference area; the first probability acquisition module is used to identify the model based on the trained image target The image feature corresponding to the current pixel is processed to obtain the first probability corresponding to the current pixel, and the first probability is the probability that the current pixel belongs to the pixel of the target object; the target area identification module uses Calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the image area where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel .

In one of the embodiments, the background area determination module includes: a first area acquisition unit, configured to acquire a first area and a second area on the current image according to the position of the current pixel point, wherein the The second area is a sub-area of the first area, and the current pixel is located inside the second area; the first area determination unit is configured to combine non-overlapping images between the first area and the second area area as the background reference area.

In one of the embodiments, the device further includes: a training area acquisition module, configured to acquire a training image, and acquire a training area corresponding to a target object in the training image; a training feature acquisition module, configured to acquire the training area The training image features corresponding to each pixel in the training module; the training module is used to perform model training according to the training image features, and obtain the feature mapping function that maps the training image features to the minimum feature space and the central value of the feature space .

In one of the embodiments, the device further includes: a second area acquisition unit, configured to acquire a third area and a fourth area on the current image according to the position of the current pixel point, wherein the fourth area The area is a sub-area of the third area, and the current pixel point is located inside the fourth area; the second area determination unit is configured to obtain a non-overlapping image area between the third area and the fourth area; The first statistical unit is configured to perform statistics on the grayscale values of the pixels corresponding to the non-overlapping image areas to obtain a first statistical result, and to perform statistics on the grayscale values of the pixels corresponding to the fourth area to obtain the first statistical result. Two statistical results: a contrast feature obtaining unit, configured to obtain the contrast feature according to the first statistical result and the second statistical result.

In one of the embodiments, the target area identification module includes: a second probability obtaining unit, configured to obtain a second probability corresponding to the current pixel point according to the background similarity, wherein the second probability is the The probability that the current pixel point belongs to the pixel point of the target object, the second probability is negatively correlated with the background similarity; the target probability obtaining unit is used to determine the The current target probability corresponding to the current pixel point, the current target probability is the probability that the current pixel point belongs to the target object pixel point; the target area identification unit is used to identify and obtain according to the target probability corresponding to each pixel point in the current image The image area where the target object is located in the current image.

In one of the embodiments, the target area identification unit is configured to: acquire the first pixel in the current image whose target probability is greater than the first threshold; acquire the distribution feature of the target probability corresponding to the first pixel; according to The distribution feature obtains a second threshold; and an area obtained by combining first pixels in the current image with a target probability greater than the second threshold is acquired as an image area where the target object is located in the current image.

A computer device includes a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor is made to execute the steps of the above image recognition method.

A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the processor is made to execute the steps of the above-mentioned image recognition method.

The above image recognition method, device, computer equipment and storage medium acquire the current image of the target to be recognized, obtain the current pixel from the current image, determine the corresponding background reference area according to the position of the current pixel, and calculate the current pixel according to the background reference area. The background similarity corresponding to the point, according to the trained image target recognition model, the image feature corresponding to the current pixel is processed to obtain the first probability corresponding to the current pixel, the first probability is the probability that the current pixel belongs to the pixel of the target object , calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the image area where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel. Since the background similarity can reflect whether the pixel is the background, and the probability obtained by using the model positively reflects whether the pixel is the target, combined with the background similarity corresponding to the pixel and the probability that the pixel belongs to the pixel of the target object, the location of the target object can be obtained. Image area, image recognition accuracy is high.

Description of drawings

Fig. 1 is the application environment diagram of the image recognition method provided in one embodiment;

Fig. 2 is the flowchart of image recognition method in an embodiment;

Figure 3A is a schematic diagram of a first area and a second area in one embodiment;

Fig. 3B is a schematic diagram of the first area and the second area in one embodiment;

Fig. 4 is a flow chart of calculating the background similarity corresponding to the current pixel according to the background reference area in one embodiment;

Fig. 5 is a schematic diagram of dividing a background reference area into a plurality of sub-areas in an embodiment;

Fig. 6 is a flow chart of processing the image feature corresponding to the current pixel point according to the trained image target recognition model to obtain the first probability corresponding to the current pixel point;

Fig. 7A is a flow chart of obtaining the image target recognition model in one embodiment;

FIG. 7B is a schematic diagram of an area corresponding to a target object in a training image in an embodiment;

Fig. 8 is a flow chart of obtaining the contrast feature corresponding to the current pixel point in one embodiment;

Fig. 9 is a flow chart of obtaining the gray gradient feature corresponding to the current pixel point in one embodiment;

Fig. 10 is a flow chart of identifying the image area where the target object is located in the current image according to the background similarity corresponding to each pixel point and the first probability in one embodiment;

Fig. 11 is a flow chart of identifying and obtaining the image area where the target object is located in the current image according to the target probability corresponding to each pixel in the current image in one embodiment;

Fig. 12 is a schematic diagram of target probabilities corresponding to each pixel of the current image in an embodiment;

Fig. 13 is a schematic diagram of image detection results in an embodiment;

Fig. 14 is a structural block diagram of an image recognition device in an embodiment;

Fig. 15 is a structural block diagram of a background area determination module in an embodiment;

Fig. 16 is a structural block diagram of a target area identification module in an embodiment;

Figure 17 is a block diagram of the internal structure of a computer device in one embodiment.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

FIG. 1 is an application environment diagram of an image recognition method provided in an embodiment. As shown in FIG. 1 , the application environment includes a camera 110 and a computer device 120 . After the camera device 110 captures the current image, it sends it to the computer device 120, and the computer device 120 obtains the current image of the target to be recognized, executes the image recognition method provided by the embodiment of the present invention, and recognizes the location of the target object in the current image. image area.

In one embodiment, after obtaining the image area where the target object is located in the current image, the computer 120 can display the current image and identify the image area where the target object is located, for example, add an arrow on the current image, and the arrow points to Region is the image region where the target object is located.

In one embodiment, after obtaining the image area where the target object is located in the current image, the pixel gray value of the image area where the target object is located can also be set to 1, and the pixels in the current image that are not in the image area where the target object is located The point gray value is set to 0 to obtain the binarized image corresponding to the current image.

In one embodiment, the current image is an infrared image, and the camera 110 may be an infrared camera.

In one embodiment, the computer device 120 can be an independent physical server or terminal, or a server cluster composed of multiple physical servers, or a cloud that provides basic cloud computing services such as cloud server, cloud database, cloud storage, and CDN. server.

It should be noted that the above-mentioned application scenario is only an example, and should not be regarded as a limitation to the present invention. In actual applications, other application scenarios may also exist. For example, computer device 120 may obtain the current image from local storage or obtain the current image from other devices. Alternatively, the computer device 120 and the imaging device 110 may be the same device.

As shown in FIG. 2 , in one embodiment, an image recognition method is proposed, and this embodiment is mainly illustrated by taking the method applied to the computer device 120 in the above-mentioned FIG. 1 as an example. Specifically, the following steps may be included:

Step S202, acquiring the current image of the target to be identified.

Specifically, the current image refers to an image that currently needs to identify a target object. The current image may be transmitted to the computer device by the camera device in real time or periodically, and the current image may also be an image locally stored by the computer device. The computer device may also acquire a current image of the target to be identified in response to the image recognition request. For example, when a user needs to identify a target object in an image, an image recognition request can be sent, and the image recognition request can carry an image or an image identifier. After receiving the image recognition request, the computer device obtains the current image according to the image recognition request. The target object refers to an object that needs to be recognized. For example, the target object can be a human face, a cat, or an airplane, and the specific target object can be set according to needs.

Step S204, obtaining the current pixel point from the current image, and determining the corresponding background reference area according to the position of the current pixel point.

Specifically, a pixel is the smallest image unit in an image represented by a sequence of numbers. The current image includes a plurality of pixels (pixels), and the image recognition of the current image is performed in units of pixels. The current pixel point refers to the currently acquired pixel point. For each pixel of the current image, steps S204-S208 may be performed sequentially as the current pixel, or multiple pixels may be used as the current pixel at the same time, and steps S204-S208 may be respectively performed for each current pixel. The background reference area is an image area corresponding to the background, and the background reference area is determined according to the position of the current pixel. For example, an area composed of pixels adjacent to the current pixel can be used as a background reference area. The area within the preset range of the current pixel can also be used as the background reference area.

In one embodiment, determining the corresponding background reference area according to the position of the current pixel includes: acquiring the first area and the second area on the current image according to the position of the current pixel, and dividing the area between the first area and the second area Non-overlapping image regions serve as background reference regions. Wherein, the second area is a sub-area of the first area, and the current pixel is located inside the second area.

Specifically, the sizes of the first area and the second area can be set as required. The first area and the second area may be partial areas in the current image. For example, the first area may include 7*7 pixels, and the second area may include 3*3 pixels. The fact that the second area is a sub-area of the first area means that the second area belongs to the first area. It can be understood that since the current pixel point is located in the second area, and the second area is a sub-area of the first area, the current pixel point is also located in the first area. After the first area and the second area are obtained, the area outside the second area in the first area, that is, the non-overlapping area between the first area and the second area is used as the background reference area. In the embodiment of the present invention, since the pixels around the current pixel may also be the pixels corresponding to the target object, the non-overlapping image area between the first area and the second area is used as the background reference area, which improves the selection of the background. The accuracy of the reference area.

As shown in Figure 3A, assume that a square in Figure 3A represents a pixel point, P _ij represents the pixel point in row i and column j, where P ₄₄ is the current pixel point, P ₃₂ , P ₃₃ , P ₃₄ , P ₄₂ , P ₄₃ , P ₄₄ , P ₅₂ , P ₅₃ and P ₅₄ are the second area, that is, the area corresponding to the oblique line in Figure 3A, the first area is the area composed of all the pixels in Figure 3, and the non-overlapping area That is, the background reference area is the area in Figure 3A except the area corresponding to the oblique line.

In one embodiment, at least one of the first area and the second area takes the current pixel as a symmetrical center. For example, when the shape of the first region is a rectangle, the center of symmetry is the intersection point of the diagonals. As shown in FIG. 3B , P ₄₄ is the current pixel point, the second area is the area corresponding to the oblique line, and the background reference area is an area other than the area corresponding to the oblique line in FIG. 3B .

Step S206, calculating the background similarity corresponding to the current pixel point according to the background reference area.

Specifically, the background similarity is used to indicate the degree of similarity between the current pixel and the background. The greater the similarity, the greater the possibility that the current pixel is the background. After the background reference area is obtained, each pixel can be used as a target pixel to calculate the similarity between each pixel in the background reference area and the current pixel. It is also possible to select some pixels from the background reference area as the target pixels, and calculate the similarity between the target pixels and the current pixel. After obtaining the similarity between the target pixel and the current pixel, the background similarity is obtained according to the calculated similarity. For example, the average value, the maximum similarity, the minimum similarity, or the median value of the calculated similarity may be used as the background similarity. The calculation method of the similarity can be set as required. For example, the pixel features corresponding to the pixel points can be obtained, and the similarity between the pixel features can be calculated. The pixel features can be one or more of color features, texture features, and grayscale features.

In one embodiment, the similarity can be calculated according to the gray value of the pixel. For example, the gray value of the pixel can be obtained, and the gray value of the pixel can be reversed to obtain the complementary current gray value, and then the gray value corresponding to the pixel and the complementary gray value can be combined into a gray value vector. Then calculate the similarity between vectors to obtain the similarity between pixels. The similarity between vectors can be calculated using the cosine similarity algorithm and the Euclidean distance similarity algorithm.

Step S208: Process the image features corresponding to the current pixel according to the trained image object recognition model to obtain a first probability corresponding to the current pixel, where the first probability is the probability that the current pixel belongs to the pixel of the target object.

Specifically, the image features are used to represent the properties corresponding to the image, such as one or more of the contrast features, color features, or grayscale features corresponding to the pixels, which can be selected according to needs. The pixel point of the target object is a pixel point corresponding to the target object. Before processing the image features according to the trained image target recognition model, it is necessary to carry out model training through the training data to determine the model parameters corresponding to the image target recognition model, and establish a mapping from image features to the probability that the pixel belongs to the pixel of the target object. The method of model training can be a supervised training method or an unsupervised training method. For a supervised training method, it is known whether the pixels in the training data are the pixels of the target object, and the supervised training model can be, for example, a support vector machine or a deep neural learning model. For the unsupervised training method, it may be unknown whether the pixels in the training data are the pixels of the target object, and the unsupervised training model may be, for example, a clustering algorithm.

Step S210, calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the image area where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel.

Specifically, calculate the background similarity and first probability corresponding to each pixel in the current image according to steps S202-208, and then identify the image where the target object is located in the current image by combining the background similarity and first probability corresponding to each pixel area.

In one embodiment, pixels whose background similarity is less than a preset similarity and whose first probability is greater than a preset probability may be used as target object pixels.

In an embodiment, the number of pixels of the image area corresponding to the target object may also be set, and combined with the number of pixels of the image area corresponding to the target object, the image area where the target object is located in the current image can be obtained. For example, when the number of pixels in the image area where the target object is located is preset to 8, the 10 pixels with the smallest background similarity can be obtained, and then the first probability ranking of the top 8 pixels can be obtained from these 10 pixels. Pixels are used as the pixels of the target object, and the area formed by the pixels of the target object is used as the image area where the target object is located.

In an embodiment, the image area where the target object is located in the current image can also be obtained by combining the positions of the pixel points. You can select pixels whose background similarity is less than the preset similarity and whose first probability is greater than the preset probability, and then obtain the positions of these selected pixels, and use the continuous image area composed of the selected pixels as the target object in the current image image area.

In one embodiment, the second probability can also be obtained according to the similarity of the background, and the second probability is the probability that the current pixel belongs to the pixel of the target object. Then multiply the first probability and the second probability to obtain the target probability, and use the pixel points whose target probability is greater than the preset value as the target object pixel points in the current image to obtain the image area where the target object is located.

In an embodiment, the gray value of the pixel determined as the target object in the current image may also be set to 1, and the gray value of other pixels may be set to 0, and a binarized image corresponding to the current image may be displayed.

The above image recognition method obtains the current image of the target to be recognized, obtains the current pixel from the current image, determines the corresponding background reference area according to the position of the current pixel, and calculates the background similarity corresponding to the current pixel according to the background reference area. The trained image target recognition model processes the image features corresponding to the current pixel to obtain the first probability corresponding to the current pixel. The first probability is the probability that the current pixel belongs to the pixel of the target object. The background similarity and the first probability corresponding to the pixel points are used to identify the image area where the target object is located in the current image according to the background similarity and the first probability corresponding to each pixel point. Since the background similarity can reflect whether the pixel is the background, and the probability obtained by using the model positively reflects whether the pixel is the target, combined with the background similarity corresponding to the pixel and the probability that the pixel belongs to the pixel of the target object, the location of the target object can be obtained. Image area, image recognition accuracy is high.

The method provided by the embodiment of the present invention can be applied to target object recognition of infrared images. Infrared images are images obtained by infrared imaging technology. In an environment with insufficient light intensity and poor contrast, the imaging method based on infrared detection technology can be used without relying on The image is obtained under the condition of light, and the target object in the infrared image is generally relatively small. For example, the size of the image area corresponding to the target object is generally between 1×1 pixel and 6×6 pixel, and it is usually buried in a complex background. , coupled with the inhomogeneity of atmospheric thermal radiation, atmospheric attenuation under different meteorological conditions, and the internal noise of infrared detectors, etc., lead to drastic changes in the gray scale of infrared images, which is very different from the slow change in gray scale of traditional visible light images. s difference. Therefore, the effect of identifying the target object using the traditional image recognition method is poor. However, the method provided by the embodiment of the present invention uses the similarity between the current pixel point and the pixel point in the background reference area to obtain the background similarity, and uses the image recognition model to determine the probability that the pixel point is the pixel point of the target object. The two methods are combined from one The positive and negative angles comprehensively determine whether the pixel is the pixel of the target object, so the image recognition effect is good.

In one embodiment, as shown in FIG. 4, step S206, that is, calculating the background similarity corresponding to the current pixel point according to the background reference area may include the following steps:

In step S402, the target pixel is obtained from the background reference area, and the gray value corresponding to the target pixel and the current pixel is obtained.

Specifically, there may be one or more target pixel points. The target pixels can be all the pixels in the background reference area, and of course the target pixels can also be obtained according to preset pixel filtering rules. For example, in the background reference area, the pixel point whose gray value is the median of the gray value of each pixel point in the background reference area may be acquired as the target pixel point.

In an embodiment, the background reference area may also be divided into multiple sub-areas, and target pixel points are acquired from each sub-area. For example, a pixel point whose gray value in each sub-region corresponds to the median of the gray values of the pixels in the sub-region is used as the target pixel point. As shown in Figure 5, the area composed of pixels passed by the line segment Q1 passing through the center point of the first area vertically in the horizontal direction can be used as the first sub-area, and the line segment Q1 vertically passing through the center point of the first area in the vertical direction can be used as the first sub-area. The pixels passed by the line segment Q2 are used as the second sub-area, and the areas formed by the pixels passed by the diagonal lines Q3 and Q4 of the first area are respectively used as the third sub-area and the fourth sub-area. Then, the median of the grayscale values of the pixels in each sub-region is obtained, and the pixel with the grayscale value of the median in each sub-region is used as the target pixel.

In an embodiment, when there is a second area, it can be understood that the pixels corresponding to each sub-area do not include the pixels in the second area. For example, the fourth sub-area corresponding to the line segment Q4 may be an area composed of P ₁₁ , P ₂₂ , P ₆₆ , and P ₇₇ , excluding P ₃₃ , P ₄₄ , and P ₅₅ .

Step S404, calculate the reference gray value according to the gray value of the target pixel point, perform an inverse operation on the reference gray value to obtain the corresponding complementary reference gray value, and combine the reference gray value and the complementary reference gray value to form a reference gray A vector of degree values.

Specifically, the reference grayscale value may be a median of grayscale values corresponding to each target pixel point. It may also be a gray value obtained by normalizing the gray value of the target pixel, or may be a gray value obtained by normalizing the median of the gray values corresponding to the target pixel. For example, when the grayscale value of the target pixel is 200, since the grayscale value ranges from 0 to 255, dividing 200 by 255 results in a normalized reference grayscale value of 0.784. Or when the target pixel points include 4, respectively 100, 250, 200 and 210, the median gray value obtained first is (200+210=)/2=205, and the corresponding reference gray value is 205/255= 0.804. The inverse operation is used to obtain the complementary image of the image. Complementary means that the addition of two grayscale values is equal to the maximum value of the grayscale, such as 255 or 1. For example, when the reference gray value is 0.804, the complementary reference gray value is 1−0.804=0.196. After obtaining the reference gray value and the complementary reference gray value, a reference gray value vector is formed, and the reference gray value vector is [0.804, 0.196].

Step S406, performing an inverse operation on the gray value corresponding to the current pixel to obtain a corresponding complementary current gray value, and combining the gray value corresponding to the current pixel and the complementary current gray value to form a current gray value vector.

Specifically, the gray value corresponding to the current pixel point may be a normalized gray value or an unnormalized gray value. After the gray value corresponding to the current pixel is obtained, the gray value corresponding to the current pixel is also inverted to obtain the corresponding complementary current gray value, and then the gray value corresponding to the current pixel and the complementary current gray value Make up the current gray value vector. For example, when the grayscale value corresponding to the current pixel is 0.901, the complementary current grayscale value is 1-0.901=0.099. The current test gray value vector is [0.901,0.099].

Step S408, calculating the background similarity corresponding to the current pixel according to the reference gray value vector and the current gray value vector.

Specifically, the method of calculating the background similarity according to the reference gray value vector and the current gray value vector can be set according to actual needs. For example, the cosine similarity algorithm may be used, or the Euclidean distance calculation method may be used for calculation.

In one embodiment, when there are multiple reference grayscale value vectors, the similarity between each reference grayscale value vector and the current grayscale value vector can be calculated respectively, and then according to each reference grayscale value vector and the current grayscale value vector The similarity between the value vectors yields the background similarity. For example, one of the median value, the average value, the maximum value and the minimum value of the similarity between each reference gray value vector and the current gray value vector may be taken as the background similarity.

In one embodiment, the method for calculating the background similarity includes: comparing the reference gray value vector with the vector value at the same position in the current gray value vector, and obtaining the minimum vector value corresponding to each position. Combine the minimum values to obtain the recombination vector, calculate the square of the modulus of the recombination vector, and obtain the background blur membership corresponding to the current pixel according to the square of the modulus of the recombination vector, the modulus of the current gray value vector, and the modulus of the reference gray value vector degree, the background similarity is obtained according to the background fuzzy membership degree.

Specifically, the background blur membership degree is used to indicate the degree to which the current pixel belongs to the background, and the algorithm for obtaining the background similarity according to the background blur membership degree can be set as required. For example, when there is only one background fuzzy membership degree, the fuzzy membership degree can be used as the background similarity degree. When there are multiple background fuzzy membership degrees, one of the middle value, average value, maximum value, and minimum value of the fuzzy membership degrees can be used as the background similarity. An example of how to obtain the reorganization vector is as follows. If the current gray value vector is [0.901,0.099] and the reference gray value vector is [0.804,0.196], then the reorganization vector is [0.804,0.099]. Assuming that there are 4 target pixels, then The above method of calculating background similarity is expressed as follows with formulas (1)-(4):

I=[F _t (x, y), F _t (x, y) ^c ], F _t (x, y) ^c = 1-F _t (x, y) (1)

Bg(x,y)=max{P _j |j=1...4} (4)

Among them, in the above formula, F _t (x, y) is the gray value corresponding to the current pixel point t, F _t (x, y) ^c is the complementary current gray value of the current pixel point t, and I is the current gray value vector. w _j is the reference gray value corresponding to the jth target pixel, w _j ^c is the complementary reference gray value of the jth target pixel, W _j is the reference gray value vector of the jth target pixel, P _j is the fuzzy membership degree between the current pixel and the jth target pixel, The parameters set to prevent the fuzzy membership degree from being equal to 1 can be set as required. "^" is a fuzzy intersection operator, and its calculation result is the minimum value of the vector values at the same position between two vectors. "||" represents the modulus of the vector, Bg(x, y) represents the background similarity, and max represents the maximum value, that is, P _j with the maximum background similarity.

In one embodiment, the image object recognition model is a support vector clustering model, as shown in FIG. 6 , step S208 is to process the image features corresponding to the current pixel according to the trained image object recognition model, and obtain the current pixel corresponding to The specific steps of the first probability may include:

Step S602, obtaining a feature mapping function corresponding to the image target recognition model, and obtaining a central value of a feature space corresponding to the feature mapping function.

Specifically, the basic idea of the support vector clustering model is: for the input features used for model training, a feature mapping function can be used to map the input features to the feature space to obtain the feature map value. The feature space is the smallest possible The feature space of the mapped value obtained after the overlay mapping, and the center value of the feature space is the center value of the feature map value obtained after the feature map function is mapped. For example, the feature space can be a hypersphere, and the center of the feature space is the center of the hypersphere. During training, since finding a minimum feature space that completely covers all feature vectors will result in a relatively large feature space, the model training conditions can be set, and when the model training conditions are met, the obtained feature space will be used as the minimum feature that satisfies the conditions Space, the method of model training is described later. Therefore, for the trained image target recognition model, the corresponding feature mapping function can be obtained, and the central value of the feature space corresponding to the feature mapping function can be obtained.

Step S604, calculate the image feature according to the feature mapping function, and obtain the mapping value corresponding to the image feature.

Specifically, after the feature mapping function is obtained, the feature mapping function is used to map the image feature into the feature space to obtain the corresponding mapping value.

Step S606, calculating a first distance between the mapping value and the central value.

Specifically, after the mapping value is obtained, the distance between the mapping value and the central value in the feature space is obtained according to the mapping value. Suppose the image feature is s(i), the feature mapping function is Φ, and the center value is a, then the first distance between the mapped value and the center value can be expressed as |Φ(s(i))-a|, where “||” Indicates that the Euclidean distance is calculated.

Step S608, calculating a first probability corresponding to the current pixel point according to the first distance, wherein the first distance has a negative correlation with the first probability.

Specifically, the first distance has a negative correlation with the first probability, that is, the first probability becomes smaller as the first distance increases. For example, the first probability may be the inverse of the first distance.

In an embodiment, a second distance from the center of the feature space to the boundary of the feature space may also be obtained. Calculating the first probability that the current pixel is a pixel corresponding to the target according to the first distance includes: calculating a ratio of the first distance to the second distance. The first probability corresponding to the current pixel point is calculated according to the proportional value, wherein the proportional value and the first probability are negatively correlated.

Specifically, the second distance is the distance from the boundary of the feature space to the center of the feature space. When the feature space is a hypersphere, the distance from the boundary to the center of the feature space is the distance from the center of the sphere to the surface of the sphere, that is, the distance of the sphere radius. The corresponding relationship between the proportional value and the first probability can be set. For example, when the proportional value is 0-10%, the corresponding first probability can be set as 0.8, and when the proportional value is 10-20%, the corresponding first probability can be set as 0.6.

In one embodiment, the first probability is 1 minus the proportional value, expressed as follows with formula (5), where H _t (x, y) is the first probability, and Φ(s(i)) is the image of the current pixel The mapping value corresponding to the feature, a is the central value, R is the second distance, and "||" indicates that the Euclidean distance is calculated.

In the embodiment of the present invention, by calculating the first distance between the mapping value of the current pixel point and the central value corresponding to the support vector clustering model, the first distance is negatively correlated with the first probability, that is, the farther away from the center of the feature space The pixel corresponding to the feature map value is less likely to belong to the pixel of the target object. Therefore, the probability that the current pixel is the pixel of the target object can be accurately quantified, which further improves the accuracy of image recognition.

Fig. 7 shows an embodiment to obtain the implementation flowchart of the image target recognition model, which may specifically include the following steps:

Step S702, acquiring a training image, and acquiring a training area corresponding to a target object in the training image.

Specifically, training images are used for model training. The training area is the image area where the target object is located in the training image. The training area may be manually marked, as shown in FIG. 7A , the area surrounded by the rectangular frame in FIG. 7A is the area corresponding to the target object obtained through manual recognition.

Step S704, acquiring training image features corresponding to each pixel in the training area.

Specifically, the image features are used to represent the properties corresponding to the image, such as one or more of the contrast features, color features, or grayscale features corresponding to the pixels, which can be selected according to needs.

Step S706, perform model training according to the training image features, and obtain a feature mapping function that maps the training image features to the smallest feature space and a central value of the feature space.

Specifically, the minimum feature space is based on the model training conditions. The model training conditions can be set, and when the model training conditions are met, the obtained feature space is used as the minimum feature space. For the support vector clustering model, the optimization goal of model training can be expressed as the following formula, where min means seeking the minimum value, st means subject to, and means that the formula for finding the minimum value is constrained by the formula after st. R represents the radius of the hypersphere, a represents the center of the hypersphere, Φ is the feature mapping function, and ξ _i is the relaxation variable, which means that the mapping value corresponding to some training samples can be located outside the hypersphere, and the relaxation variable can be specified according to the needs To set, n is the number of training samples, x(i) represents image features. C is a penalty function, which is used to adjust the balance between the error and the boundary of the hypersphere. Therefore, the optimization objective can be summarized as: under the condition of the penalty function and the set slack variable, the feature mapping function and the minimum hypersphere are obtained according to the training samples. The smallest hypersphere has a center value a and a radius value R.

St|Φ((x _i )-a) ² |≤R ² +ξ _i and ξ _i ≥0

In one embodiment, as shown in FIG. 8, the image feature includes a contrast feature, and the step of obtaining the contrast feature corresponding to the current pixel includes:

Step S802, acquiring a third area and a fourth area on the current image according to the position of the current pixel, wherein the fourth area is a sub-area of the third area, and the current pixel is located inside the fourth area.

Specifically, the contrast refers to the degree of light and dark contrast of the image, and the size of the third area and the fourth area can be set according to needs, for example, the third area can include 9*9 pixels, and the fourth area can include 3*3 pixels point. It can be understood that since the current pixel point is located in the fourth area, and the fourth area is a sub-area of the third area, the current pixel point is also located in the third area.

In one embodiment, the fourth area may be the same as the second area, and the third area may be the same as the first area.

In one embodiment, at least one of the fourth area and the third area takes the current pixel as a symmetrical center.

Step S804, acquiring non-overlapping image areas between the third area and the fourth area.

Specifically, since the fourth area is a sub-area of the third area, the non-overlapping image area between the third area and the fourth area is an area in the third area except the fourth area.

Step S806 , performing statistics on the grayscale values of the pixels corresponding to the non-overlapping image regions to obtain a first statistical result, and performing statistics on the grayscale values of the pixels corresponding to the fourth region to obtain a second statistical result.

Specifically, the first statistical result and the second statistical result may be a sum of gray values or an average value of gray values. For example, the gray value of each pixel in the non-overlapping area can be added and summed, and then divided by the number of pixels in the non-overlapping area to obtain the average gray value of the pixel corresponding to the non-overlapping area, as the first statistical results. The gray value of each pixel in the fourth area can be added and summed, and then divided by the number of pixels in the fourth area to obtain the average gray value of the pixel corresponding to the fourth area as the second statistical result .

Step S808, obtaining the contrast feature according to the first statistical result and the second statistical result.

Specifically, after the first statistical result and the second statistical result are obtained, the contrast feature is obtained by combining the first statistical result and the second statistical result. For example, the contrast feature may be the ratio of the first statistical result to the second statistical result, or the contrast feature may also be the difference between the first statistical result and the second statistical result.

In one embodiment, when the statistical result is the average value of the gray value, and the contrast feature is the difference between the first statistical result and the second statistical result, the calculation formula of the contrast feature is formula (6), wherein lmci is the ith Contrast features corresponding to current pixels, n _in is the number of pixels in the fourth area, n _out is the number of pixels in the third area, F(x, y) represents the gray value of the pixels.

In one embodiment, as shown in FIG. 9, the image feature includes a grayscale gradient feature, and the step of obtaining the grayscale gradient feature corresponding to the current pixel includes:

Step S902, acquiring a third area and a fourth area on the current image according to the position of the current pixel, wherein the fourth area is a sub-area of the third area, and the current pixel is located inside the fourth area.

Specifically, the grayscale gradient feature refers to a feature related to the grayscale difference between pixels. The third area and the fourth area can be obtained by referring to the method in step S802, and details are not repeated here.

Step S904, acquiring non-overlapping image areas between the third area and the fourth area.

Specifically, since the fourth area is a sub-area of the third area, the non-overlapping image area between the third area and the fourth area is an area in the third area except the fourth area.

Step S906, acquiring a first grayscale difference between each pixel in the non-overlapping image area and adjacent pixels, and acquiring a second grayscale difference between each pixel in the fourth area and adjacent pixels.

Specifically, an adjacent pixel point refers to a pixel point that has an overlapping boundary with the pixel point, and the grayscale difference value with all adjacent pixel points may be calculated, or the grayscale difference value between some adjacent pixel points may be calculated. In one embodiment, the grayscale difference can be divided into the grayscale difference in the horizontal direction and the grayscale difference in the vertical direction. The grayscale difference can be the grayscale difference in the horizontal direction and the grayscale difference in the vertical direction. One of the differences, or the grayscale difference, may be the sum of the grayscale difference in the horizontal direction and the grayscale difference in the vertical direction. Taking the calculation of the first grayscale difference corresponding to pixel point _P44 in Figure 3A as an example, the absolute value of the grayscale difference between _P44 and _P45 can be used as the grayscale difference in the horizontal direction corresponding to _P44 , the absolute value of the grayscale difference between _P44 and _P54 is taken as the grayscale difference in the vertical direction corresponding to _P44 , and then the grayscale difference in the horizontal direction and the grayscale difference in the vertical direction Add them together to get the first grayscale difference corresponding to P ₄₄ . The calculation method of the gray level difference is expressed by the formula:

G _h (x,y)=|F(x,y)-F(x+1,y)| (7)

G _v (x,y)=|F(x,y)-F(x,y+1)| (8)

G(x,y)=G _h (x,y)+G _v (x,y) (9)

Among them, G _h (x, y) represents the gray level difference in the horizontal direction, G _v (x, y) represents the gray level difference in the vertical direction, and G(x, y) is the gray level difference corresponding to the pixel Value, F(x, y) refers to the gray value corresponding to the pixel point (x, y), x can represent a row, and y represents a column.

Step S908: Perform statistics on the first grayscale difference corresponding to each pixel in the non-overlapping image area to obtain a third statistical result, and perform statistics on the second grayscale difference corresponding to each pixel in the fourth area to obtain the first grayscale difference 4. Statistical results.

Specifically, the third statistical result and the fourth statistical result may be a sum of grayscale differences or an average value of grayscale differences. For example, the first grayscale difference of each pixel in the non-overlapping area can be added and summed, and then divided by the number of pixels in the non-overlapping area to obtain the average grayscale difference of the pixels corresponding to the non-overlapping area , as the third statistical result. The second grayscale difference of each pixel in the fourth area can be added and summed, and then divided by the number of pixels in the fourth area to obtain the average grayscale difference of the pixel corresponding to the fourth area, as The fourth statistical result.

In step S910, the gray gradient feature is obtained according to the third statistical result and the fourth statistical result.

Specifically, after the third statistical result and the fourth statistical result are obtained, the gray gradient feature is obtained by combining the third statistical result and the fourth statistical result. For example, the contrast feature may be the ratio of the third statistical result to the fourth statistical result, or the gray gradient feature may also be the difference between the third statistical result and the fourth statistical result. When the grayscale gradient feature is the difference between the third statistical result and the fourth statistical result, and the third statistical result and the fourth statistical result are the average value of the grayscale difference, the calculation formula of the grayscale gradient feature is as follows, where lmgi is The grayscale gradient feature corresponding to the i-th current pixel, n _in is the number of pixels in the fourth area, n _out is the number of pixels in the third area, G(x, y) represents the grayscale difference.

In the embodiment of the present invention, when calculating the image feature corresponding to the current pixel point, the image feature of the area corresponding to the current pixel point is calculated, and it is divided into two areas for calculation, so the obtained image feature can reflect the image feature of the current pixel point. environment, the accuracy of the obtained first probability is high.

In an embodiment, the image features may include one or both of contrast features and grayscale gradient features.

In one embodiment, as shown in FIG. 10 , step S210 is to identify the image area where the target object is located in the current image according to the background similarity corresponding to each pixel point and the first probability, including:

Step S1002, obtain a second probability corresponding to the current pixel according to the background similarity, wherein the second probability is the probability that the current pixel belongs to the pixel of the target object, and the second probability is negatively correlated with the background similarity.

Specifically, the negative correlation between the background similarity and the second probability means that as the background similarity increases, the second probability becomes smaller. For example, the second probability may be the inverse of the background similarity. Or you can set the corresponding relationship between the range of background similarity and the second probability. For example, when the background similarity can be set to 0-10%, the corresponding second probability is 90%. When the background similarity is 11-20%, the corresponding The second probability is 80%. Alternatively, the second probability may be represented by the following formula, where Bg(x, y) is the background similarity, and H _b (x, y) represents the second probability corresponding to the pixel point (x, y).

_Hb (x,y)=1-Bg(x,y) (11)

Step S1004, determine the current target probability corresponding to the current pixel point according to the first probability and the second probability, the current target probability is the probability that the current pixel point belongs to the target object pixel point.

Specifically, the method of determining the current target probability corresponding to the current pixel point according to the first probability and the second probability can be set as required, for example, the first probability can be multiplied by the second probability, and the obtained product can be used as the current target probability. Alternatively, the average value of the first probability and the second probability is used as the current target probability. Of course, it is also possible to further set the corresponding relationship between the product and the target probability, and obtain the target probability according to the product. The target probability can also be obtained by combining the third probability that the current pixel point belongs to the pixel point of the target object obtained by another method. For example, the target probability may be the product of the first probability, the second probability, and the third probability.

Step S1006, according to the target probability corresponding to each pixel in the current image, identify the image area where the target object is located in the current image.

Specifically, the current image includes multiple pixels, and it is necessary to obtain the image region where the target object is located according to the target probability corresponding to each pixel. For example, pixels whose target probability is greater than a preset probability may be used as target object pixels, and an area composed of target object pixel points may be used as an image area where the target object is located.

In one embodiment, the image area where the target object is located in the current image can also be obtained according to the positional relationship of the pixel points. For example, for a target object whose corresponding area is a continuous area, if there is a pixel point that exists alone although the target probability is greater than the preset probability, that is, there is no pixel around the target probability greater than the preset probability, then the pixel point is also It may not be the pixel of the target object.

In one embodiment, as shown in FIG. 11 , step S1006 is to identify the image area where the target object is located in the current image according to the target probability corresponding to each pixel in the current image, including:

Step S1102, acquiring the first pixel in the current image whose target probability is greater than the first threshold.

Specifically, the first threshold can be set according to needs, for example, requirements for recognition accuracy. Through experiments, it is found that when the first threshold is 0.85, the recognition accuracy is high.

In one embodiment, for a pixel point whose target probability is less than the first threshold, the pixel point may be used as a background pixel point, and its corresponding target probability is updated to 0.

Step S1104, acquiring the distribution feature of the target probability corresponding to the first pixel.

Specifically, the distribution feature reflects the distribution of the obtained target probability, and may include one or more of the distribution ratio or number in the maximum value, minimum value, average value, and in each numerical range of the target probability. Specifically, it can be set as required. Each numerical range may be preset, for example, 0.85-0.88 may be set as the first numerical range, and 0.89-0.92 may be set as the second numerical range.

Step S1106, obtaining a second threshold according to the distribution feature.

Specifically, after the distribution features are obtained, the second threshold is obtained according to each distribution feature. The method for obtaining the second threshold can be set as required. For example, the image threshold segmentation algorithm can be used to calculate the second threshold. The image threshold segmentation algorithm can include, for example, OTSU (maximum inter-class variance algorithm), iterative maximum variance method and maximum entropy method One or more methods in etc., wherein when using the image segmentation algorithm to obtain the second threshold, the gray value feature is replaced by the target probability corresponding to the first pixel point for calculation to obtain the second threshold.

Step S1108, acquiring an area obtained by combining the first pixel points in the current image whose target probability is greater than the second threshold, as the image area where the target object is located in the current image.

Specifically, after the second threshold is obtained, the first pixel whose target probability is greater than the second threshold is used as the target object pixel, and the area obtained by combining the target object pixel points is used as the image area where the target object is located in the current image.

In the embodiment of the present invention, the second threshold is obtained according to the distribution characteristics of the target probability, and the image is segmented again according to the second threshold, since the pixels can be preliminarily screened through the first threshold, and then according to the specific image corresponding The probability distribution of the pixel points further screens out the pixel points of the target object, thus further improving the image recognition accuracy.

The method provided by the embodiment of the present invention is described with a specific embodiment, including the following steps:

1. Obtain the current image of the target to be identified. Assume that the current image is an image of 7*7 pixels, that is, it includes 7*7 pixels, and there are 7 rows and 7 columns in total.

2. Using each pixel as a current pixel, calculate the background similarity and corresponding first probability corresponding to each current pixel according to the method provided by the embodiment of the present invention. Wherein, when calculating the background similarity, the size of the first area and the third area is 5*5 pixels, and the size of the second area and the fourth area is 3*3 pixels.

3. Calculate the second probability corresponding to each pixel according to the background similarity corresponding to each pixel, where the second probability=1-background similarity.

4. Obtain the target probability corresponding to each pixel according to the first probability and the second probability, wherein the target probability is the product of the first probability and the second probability. As shown in FIG. 12 , a square in FIG. 12 represents a pixel, and the number in the pixel is the target probability corresponding to the pixel.

5. Obtain the first pixel in the current image whose target probability is greater than the first threshold. Assuming the first threshold is 0.85, then P ₁₆ , P ₂₅ , P ₂₆ , P ₃₄ , P ₃₅ , P ₃₆ , P ₄₅ and P ₄₆ are the first pixel.

6. Obtain the second threshold. The second threshold can be obtained by using the iterative method to select the threshold value. The iterative method selects the threshold value algorithm first selects an initial threshold value T, divides the image into two parts: R1 and R2, calculates the mean values u1 and u2 of the regions R1 and R2, and then Select a new threshold T=(u1+u2)/2, and repeat the above process until the new threshold does not change from the previous threshold or the change is smaller than the set threshold. In the embodiment of the present invention, first obtain the corresponding maximum target probability and minimum target probability in the first pixel point, which are 0.86 and 0.96 respectively, so the initial threshold value is 0.86+0.96=0.91, and 0.91 is used as the threshold value, for the first pixel point After classification, it can be obtained that P ₂₆ , P ₃₆ , and P ₄₅ are target pixel points, and the others are background pixel points. Calculate the target probability average value of the target object pixel points P ₂₆ , P ₃₆ , and P ₄₅ , which is (0.96+0.91+0.96)/3=0.94. Calculate the average target probability of background pixels, which is 0.89+0.86+0.87+0.87+0.89=0.876, so the new threshold is (0.876+0.94)/2=0.913, and the change value between the new threshold and the initial threshold is 0.913- 0.91=0.003, assuming that the change value is less than the set threshold, then 0.913 is the second threshold, assuming that the change value is greater than the set threshold, you can continue to use 0.913 as the threshold to classify the first pixel, and then repeat the above steps , until the new threshold does not change from the previous threshold or the change is smaller than the set threshold.

7. Obtain an area obtained by combining the first pixel points in the current image whose target probability is greater than the second threshold, and use it as the image area where the target object is located in the current image. Assuming that the second threshold is finally 0.87, in FIG. 12 , the area composed of four pixels P ₁₆ , P ₂₆ , P ₃₆ , P ₄₅ , and P ₄₆ is the image area where the target object is located in the current image of the target object.

In one embodiment, after performing image recognition on nine segments of infrared video by adopting Feature Selective Filtering (CSF), Maximum-Minimum Difference (DMMF) and Multi-Scale Gradient (MSG) and the method provided by the embodiment of the present invention Taking background suppression as an example, the effect of the method provided by the embodiment of the present invention is further described, and three indexes, signal-to-noise ratio gain (ISNR), contrast gain (ISCR) and background suppression factor (BSF), are used to evaluate the background suppression result. Among them, the larger the background suppression factor is, the better the global background smoothing effect is. The greater the SNR gain and contrast gain, the stronger the ability of the method to suppress clutter and enhance weak targets. The specific index results are shown in Table 1, where seq in the table represents the video, and the number after seq is the video label. It can be seen from Table 1 that the BSF, ISNR and ISCR of the feature selective filtering method (CSF) are relatively high Low, indicating that it has good global background smoothing performance, but the clutter suppression effect is not good. The ISCR of the maximum and minimum difference method (DMMF) is high, but the ISNR and BSF are low, indicating that the target signal-to-noise ratio is improved. Not strong, poor global background suppression. The three indicators of the multi-scale gradient method (MSG) are all low, indicating that its background suppression performance is average. However, this method has higher ISNR, ISCR and BSF indicators at the same time, and its global background smoothing and local clutter suppression effects are better, which can effectively suppress complex background clutter and highlight the target.

Table 1 Image recognition evaluation indicators of four detection algorithms

In one embodiment, detection rate (r), accuracy rate (p) and comprehensive index (F1) can also be used to evaluate the recognition effect of weak and small targets. Among them, r is the ratio of the number of detected correct targets to the total number of real targets, p is the ratio of the number of detected correct targets to the total number of detected targets, F1 is the comprehensive index of r and p indicators, which can be r and p The weighted harmonic mean of . Having a higher r value while maintaining a higher p value, a higher F1 value also means good recognition performance.

The detection and evaluation results of the four methods for the nine-segment infrared video are shown in Table 2. It can be seen from Table 2 that the feature selective filtering method (CSF) has a high detection rate, but the accuracy rate is low, and the overall detection performance is average. The multi-scale gradient method (MSG) and the maximum-minimum difference method (DMMF) have a high detection rate, but the accuracy rate is low, while this method has a high detection rate and accuracy rate at the same time, and its F1 index is as high as 97.7%. It has good detection stability.

Table 2 The average target detection evaluation index of the four detection methods

In addition, three images are selected, the background of the first image is cloudless and the sky color is dark, the background of the second image is cloudy and the camera distance is far away, the background of the third image is cloudy and the camera The distance is short, as shown in Figure 13, the image in the first column is the original image, the dots in the box indicate the actual position of the target object, and the above four methods are used to identify the target object in the image, the second column in Figure 13 The images in columns 4 to 4 are respectively the detection results of the target object corresponding to the feature selective filter method (CSF), the maximum and minimum difference method (DMMF), the multi-scale gradient method (MSG) and the method provided by the embodiment of the present invention, It can be seen from FIG. 13 that the method provided by the embodiment of the present invention can accurately identify the image position where the target object is located.

As shown in FIG. 14, in one embodiment, an image recognition device is provided, which can be integrated into the above-mentioned computer device 120, and specifically can include a current image acquisition module 1402, a background area determination module 1404, similar A degree calculation module 1406 , a first probability obtaining module 1408 and a target area identification module 1410 .

The current image acquiring module 1402 is configured to acquire the current image of the target to be identified.

The background area determination module 1404 is configured to obtain the current pixel point from the current image, and determine the corresponding background reference area according to the position of the current pixel point.

A similarity calculation module 1406, configured to calculate the background similarity corresponding to the current pixel point according to the background reference area.

The first probability obtaining module 1408 is used to process the image feature corresponding to the current pixel according to the trained image target recognition model to obtain the first probability corresponding to the current pixel. The first probability is that the current pixel belongs to the pixel of the target object The probability.

The target area identification module 1410 is used to calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the image where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel area.

In one of the embodiments, the background area determination module 1404 includes:

The first area acquisition unit is configured to acquire a first area and a second area on the current image according to the position of the current pixel, wherein the second area is a sub-area of the first area, and the current pixel is located inside the second area.

The first area determining unit is configured to use the non-overlapping image area between the first area and the second area as a background reference area.

In one of the embodiments, as shown in FIG. 15, the similarity calculation module 1406 includes:

The gray value obtaining unit 1406A is configured to obtain the target pixel from the background reference area, and obtain the gray value corresponding to the target pixel and the current pixel.

The reference vector composition unit 1406B is used to calculate the reference gray value according to the gray value of the target pixel point, perform an inverse operation on the reference gray value to obtain the corresponding complementary reference gray value, and combine the reference gray value and the complementary reference gray value Intensity values constitute the reference gray value vector.

The current vector composition unit 1406C is used to invert the gray value corresponding to the current pixel to obtain the corresponding complementary current gray value, and to combine the gray value corresponding to the current pixel and the complementary current gray value to form the current gray value vector.

The similarity calculation unit 1406D is configured to calculate the background similarity corresponding to the current pixel according to the reference gray value vector and the current gray value vector.

In one of the embodiments, the image target recognition model is a support vector clustering model, and the first probability obtaining module includes:

The model parameter acquisition unit is configured to acquire a feature mapping function corresponding to the image target recognition model, and acquire a central value of a feature space corresponding to the feature mapping function.

The mapping value calculation unit is used to calculate the image feature according to the feature mapping function to obtain the mapping value corresponding to the image feature.

A first distance calculation unit, configured to calculate a first distance between the mapped value and the central value.

The first probability obtaining unit is configured to calculate and obtain the first probability corresponding to the current pixel point according to the first distance, wherein the first distance has a negative correlation with the first probability.

In one of the embodiments, the image recognition device further includes:

A second distance acquisition module, configured to acquire a second distance from the center of the feature space to the boundary of the feature space.

The first probability obtaining unit is used for: calculating a ratio value of the first distance to the second distance. The first probability corresponding to the current pixel point is calculated according to the proportional value, wherein the proportional value and the first probability are negatively correlated.

In one of the embodiments, the image recognition device further includes:

The training area acquiring module is configured to acquire the training image, and acquire the training area corresponding to the target object in the training image.

The training feature acquisition module is used to acquire training image features corresponding to each pixel in the training area.

The training module is used for performing model training according to the training image features, and obtaining a feature mapping function for mapping the training image features to the smallest feature space and a central value of the feature space.

In one of the embodiments, the image device further includes:

The second area acquisition unit is configured to acquire a third area and a fourth area on the current image according to the position of the current pixel, wherein the fourth area is a sub-area of the third area, and the current pixel is located inside the fourth area.

The second area determination unit is configured to acquire non-overlapping image areas between the third area and the fourth area.

The first statistical unit is configured to perform statistics on the grayscale values of pixels corresponding to the non-overlapping image regions to obtain a first statistical result, and to collect statistics on the grayscale values of pixels corresponding to the fourth region to obtain a second statistical result.

The contrast feature obtaining unit is configured to obtain the contrast feature according to the first statistical result and the second statistical result.

In one of the embodiments, as shown in FIG. 16 , the target area identification module 1410 includes:

The second probability obtaining unit 1410A is configured to obtain the second probability corresponding to the current pixel point according to the background similarity, wherein the second probability is the probability that the current pixel point belongs to the target object pixel point, and the second probability is negatively correlated with the background similarity relation.

The target probability obtaining unit 1410B is configured to determine the current target probability corresponding to the current pixel point according to the first probability and the second probability, and the current target probability is the probability that the current pixel point belongs to the target object pixel point.

The target area identification unit 1410C is configured to identify the image area where the target object is located in the current image according to the target probability corresponding to each pixel in the current image.

In one of the embodiments, the target area identifying unit 1410C is configured to: acquire a first pixel point in the current image whose target probability is greater than a first threshold. Obtain the distribution feature of the target probability corresponding to the first pixel point. The second threshold is obtained according to the distribution characteristics. An area obtained by combining the first pixel points in the current image with a target probability greater than the second threshold is acquired as the image area where the target object is located in the current image.

Figure 17 shows an internal block diagram of a computer device in one embodiment. The computer device may specifically be the calculator device 120 in FIG. 1 . As shown in FIG. 17 , the computer equipment includes a processor, a memory, a network interface, and an input device connected through a system bus. Wherein, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system, and may also store a computer program. When the computer program is executed by the processor, the processor can realize the image recognition method. A computer program may also be stored in the internal memory, and when the computer program is executed by the processor, the processor may execute the image recognition method. The input device of the computer equipment may be a touch layer covered on the display screen, or a button, a trackball or a touch pad provided on the casing of the computer equipment, or an external keyboard, touch pad or mouse.

Those skilled in the art can understand that the structure shown in Figure 17 is only a block diagram of a partial structure related to the solution of this application, and does not constitute a limitation on the computer equipment on which the solution of this application is applied. The specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.

In one embodiment, the image recognition apparatus provided in the present application can be implemented in the form of a computer program, and the computer program can be run on the computer device as shown in FIG. 17 . Various program modules that make up the image recognition device can be stored in the memory of the computer equipment, such as the current image acquisition module 1402 shown in FIG. Region identification module 1410 . The computer program constituted by each program module enables the processor to execute the steps in the image recognition method of each embodiment of the application described in this specification.

For example, the computer equipment shown in FIG. 17 can obtain the current image of the target to be recognized through the current image acquisition module 1402 in the image recognition device as shown in FIG. 14; obtain the current pixel from the current image through the background area determination module 1404, Determine the corresponding background reference area according to the position of the current pixel point; calculate the background similarity corresponding to the current pixel point by the similarity calculation module 1406 according to the background reference area; obtain the current pixel point by the first probability module 1408 according to the trained image target recognition model The image feature corresponding to the pixel is processed to obtain the first probability corresponding to the current pixel, and the first probability is the probability that the current pixel belongs to the pixel of the target object; through the calculation of the target area recognition module 1410, each pixel in the current image corresponds to The background similarity and first probability corresponding to each pixel point are identified to obtain the image area where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel.

In one embodiment, a computer device is proposed. The computer device includes a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the computer program, the following steps are implemented: acquiring the target to be identified the current image; obtain the current pixel from the current image, and determine the corresponding background reference area according to the position of the current pixel; calculate the background similarity corresponding to the current pixel according to the background reference area; The image features corresponding to the pixels are processed to obtain the first probability corresponding to the current pixel. The first probability is the probability that the current pixel belongs to the pixel of the target object; the background similarity and the first probability corresponding to each pixel in the current image are calculated. One probability, according to the background similarity corresponding to each pixel and the first probability, the image area where the target object is located in the current image is identified.

In one embodiment, the determination of the corresponding background reference area according to the position of the current pixel by the processor includes: acquiring a first area and a second area on the current image according to the position of the current pixel, wherein the second area is the first A sub-area of the first area, the current pixel is located inside the second area; the non-overlapping image area between the first area and the second area is used as the background reference area.

In one embodiment, the calculation of the background similarity corresponding to the current pixel point according to the background reference area performed by the processor includes: obtaining the target pixel point from the background reference area, and obtaining the gray value corresponding to the target pixel point and the current pixel point; The gray value of the target pixel is calculated to obtain a reference gray value, and the reference gray value is inverted to obtain a corresponding complementary reference gray value, and the reference gray value and the complementary reference gray value are composed of a reference gray value vector; The gray value corresponding to the current pixel is inverted to obtain the corresponding complementary current gray value, and the gray value corresponding to the current pixel and the complementary current gray value form the current gray value vector; according to the reference gray value vector Calculate the background similarity corresponding to the current pixel with the current gray value vector.

In one embodiment, the image target recognition model is a support vector clustering model, and the processor executes processing the image features corresponding to the current pixel according to the trained image target recognition model, and obtaining the first probability corresponding to the current pixel includes: Obtain the feature mapping function corresponding to the image target recognition model, and obtain the central value of the feature space corresponding to the feature mapping function; calculate the image feature according to the feature mapping function, and obtain the mapping value corresponding to the image feature; calculate the first mapping value and the central value A distance: calculating a first probability corresponding to the current pixel point according to the first distance, wherein the first distance is negatively correlated with the first probability.

In one embodiment, the computer program further enables the processor to perform the following steps: obtain the second distance from the center of the feature space to the boundary of the feature space; calculate the first probability that the current pixel point is the pixel point corresponding to the target according to the first distance The method includes: calculating a ratio value between the first distance and the second distance; calculating a first probability corresponding to the current pixel point according to the ratio value, wherein the ratio value and the first probability have a negative correlation.

In one embodiment, the step of obtaining the image target recognition model performed by the processor includes: obtaining a training image, obtaining a training area corresponding to a target object in the training image; obtaining training image features corresponding to each pixel in the training area; The image features are used for model training, and the feature mapping function that maps the training image features to the smallest feature space and the center value of the feature space are obtained.

In one embodiment, the image features executed by the processor include contrast features, and the step of obtaining the contrast feature corresponding to the current pixel point includes: acquiring the third area and the fourth area on the current image according to the position of the current pixel point, wherein, The fourth area is a sub-area of the third area, and the current pixel is located inside the fourth area; the non-overlapping image area between the third area and the fourth area is obtained; the gray value of the pixel point corresponding to the non-overlapping image area is calculated Statistics, to obtain a first statistical result, and perform statistics on the gray values of pixels corresponding to the fourth area to obtain a second statistical result; obtain a contrast feature according to the first statistical result and the second statistical result.

In one embodiment, the process performed by the processor to identify the image region where the target object is located in the current image according to the background similarity corresponding to each pixel and the first probability includes: obtaining the second probability corresponding to the current pixel according to the background similarity, Wherein, the second probability is the probability that the current pixel belongs to the pixel of the target object, and the second probability is negatively correlated with the background similarity; the current target probability corresponding to the current pixel is determined according to the first probability and the second probability, and the current target probability is the probability that the current pixel belongs to the pixel of the target object; the image area where the target object is located in the current image is identified according to the target probability corresponding to each pixel in the current image.

In one embodiment, the identifying the image region where the target object is located in the current image according to the target probability corresponding to each pixel in the current image performed by the processor includes: acquiring a first pixel in the current image with a target probability greater than a first threshold; Obtain the distribution feature of the target probability corresponding to the first pixel point; obtain the second threshold according to the distribution feature; obtain the area obtained by combining the first pixel points in the current image whose target probability is greater than the second threshold value, and use it as the image where the target object is located in the current image area.

In one embodiment, a computer-readable storage medium is provided. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, the processor is made to perform the following steps: acquire the current image of the target to be identified; Obtain the current pixel in the current image, determine the corresponding background reference area according to the position of the current pixel; calculate the background similarity corresponding to the current pixel according to the background reference area; Features are processed to obtain the first probability corresponding to the current pixel point, the first probability is the probability that the current pixel point belongs to the target object pixel point; the background similarity and the first probability corresponding to each pixel point in the current image are calculated, and according to each The background similarity corresponding to the pixel points and the first probability identify the image area where the target object is located in the current image.

In one embodiment, the determination of the corresponding background reference area according to the position of the current pixel by the processor includes: acquiring a first area and a second area on the current image according to the position of the current pixel, wherein the second area is the first A sub-area of the first area, the current pixel is located inside the second area; the non-overlapping image area between the first area and the second area is used as the background reference area.

In one embodiment, the calculation of the background similarity corresponding to the current pixel point according to the background reference area performed by the processor includes: obtaining the target pixel point from the background reference area, and obtaining the gray value corresponding to the target pixel point and the current pixel point; The gray value of the target pixel is calculated to obtain a reference gray value, and the reference gray value is inverted to obtain a corresponding complementary reference gray value, and the reference gray value and the complementary reference gray value are composed of a reference gray value vector; The gray value corresponding to the current pixel is inverted to obtain the corresponding complementary current gray value, and the gray value corresponding to the current pixel and the complementary current gray value form the current gray value vector; according to the reference gray value vector Calculate the background similarity corresponding to the current pixel with the current gray value vector.

In one embodiment, the image target recognition model is a support vector clustering model, and the processor executes processing the image features corresponding to the current pixel according to the trained image target recognition model, and obtaining the first probability corresponding to the current pixel includes: Obtain the feature mapping function corresponding to the image target recognition model, and obtain the central value of the feature space corresponding to the feature mapping function; calculate the image feature according to the feature mapping function, and obtain the mapping value corresponding to the image feature; calculate the first mapping value and the central value A distance: calculating a first probability corresponding to the current pixel point according to the first distance, wherein the first distance is negatively correlated with the first probability.

In one embodiment, the computer program further enables the processor to perform the following steps: obtain the second distance from the center of the feature space to the boundary of the feature space; calculate the first probability that the current pixel point is the pixel point corresponding to the target according to the first distance The method includes: calculating a ratio value between the first distance and the second distance; calculating a first probability corresponding to the current pixel point according to the ratio value, wherein the ratio value and the first probability have a negative correlation.

In one embodiment, the step of obtaining the image target recognition model performed by the processor includes: obtaining a training image, obtaining a training area corresponding to a target object in the training image; obtaining training image features corresponding to each pixel in the training area; The image features are used for model training, and the feature mapping function that maps the training image features to the smallest feature space and the center value of the feature space are obtained.

In one embodiment, the image features executed by the processor include contrast features, and the step of obtaining the contrast feature corresponding to the current pixel point includes: acquiring the third area and the fourth area on the current image according to the position of the current pixel point, wherein, The fourth area is a sub-area of the third area, and the current pixel is located inside the fourth area; the non-overlapping image area between the third area and the fourth area is obtained; the gray value of the pixel point corresponding to the non-overlapping image area is calculated Statistics, to obtain a first statistical result, and perform statistics on the gray values of pixels corresponding to the fourth area to obtain a second statistical result; obtain a contrast feature according to the first statistical result and the second statistical result.

In one embodiment, the process performed by the processor to identify the image region where the target object is located in the current image according to the background similarity corresponding to each pixel and the first probability includes: obtaining the second probability corresponding to the current pixel according to the background similarity, Wherein, the second probability is the probability that the current pixel belongs to the pixel of the target object, and the second probability is negatively correlated with the background similarity; the current target probability corresponding to the current pixel is determined according to the first probability and the second probability, and the current target probability is the probability that the current pixel belongs to the pixel of the target object; the image area where the target object is located in the current image is identified according to the target probability corresponding to each pixel in the current image.

In one embodiment, the identifying the image region where the target object is located in the current image according to the target probability corresponding to each pixel in the current image performed by the processor includes: acquiring a first pixel in the current image with a target probability greater than a first threshold; Obtain the distribution feature of the target probability corresponding to the first pixel point; obtain the second threshold according to the distribution feature; obtain the area obtained by combining the first pixel points in the current image whose target probability is greater than the second threshold value, and use it as the image where the target object is located in the current image area.

It should be understood that although the various steps in the flow charts of the embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in each embodiment may include multiple sub-steps or multiple stages, these sub-steps or stages are not necessarily executed at the same time, but may be executed at different times, the sub-steps or stages The order of execution is not necessarily performed sequentially, but may be performed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be realized through computer programs to instruct related hardware, and the programs can be stored in a non-volatile computer-readable storage medium When the program is executed, it may include the processes of the embodiments of the above-mentioned methods. Wherein, any references to memory, storage, database or other media used in the various embodiments provided in the present application may include non-volatile and/or volatile memory. Nonvolatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The technical features of the above-mentioned embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, should be considered as within the scope of this specification.

The above-mentioned embodiments only express several implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but should not be construed as limiting the patent scope of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (15)

1. An image recognition method, said method comprising: Obtain the current image of the target to be identified; Obtaining a current pixel point from the current image, and determining a corresponding background reference area according to the position of the current pixel point; calculating the background similarity corresponding to the current pixel according to the background reference area; According to the trained image target recognition model, the image feature corresponding to the current pixel is processed to obtain the first probability corresponding to the current pixel, and the first probability is that the current pixel belongs to the pixel of the target object probability; Calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the image area where the target object is located in the current image according to the background similarity and first probability corresponding to each pixel. 2. The method according to claim 1, wherein the determining the corresponding background reference area according to the position of the current pixel comprises: Acquire a first area and a second area on the current image according to the position of the current pixel, wherein the second area is a sub-area of the first area, and the current pixel is located in the second inside the area; A non-overlapping image area between the first area and the second area is used as the background reference area. 3. The method according to claim 1, wherein the calculating the background similarity corresponding to the current pixel according to the background reference area comprises: acquiring a target pixel from the background reference area, and acquiring gray values corresponding to the target pixel and the current pixel; Calculate the reference gray value according to the gray value of the target pixel point, perform an inverse operation on the reference gray value to obtain a corresponding complementary reference gray value, and combine the reference gray value and the complementary reference gray value The gray value constitutes the reference gray value vector; performing an inversion operation on the gray value corresponding to the current pixel to obtain a corresponding complementary current gray value, and combining the gray value corresponding to the current pixel and the complementary current gray value to form a current gray value vector; The background similarity corresponding to the current pixel is calculated according to the reference gray value vector and the current gray value vector. 4. The method according to claim 1, wherein the image target recognition model is a support vector clustering model, and the image feature corresponding to the current pixel is processed according to the trained image target recognition model , obtaining the first probability corresponding to the current pixel includes: Obtaining the feature mapping function corresponding to the image target recognition model, and acquiring the central value of the feature space corresponding to the feature mapping function; calculating the image feature according to the feature mapping function to obtain a mapping value corresponding to the image feature; calculating a first distance between the mapped value and the central value; A first probability corresponding to the current pixel point is calculated according to the first distance, where the first distance is negatively correlated with the first probability. 5. method according to claim 4, is characterized in that, described method also comprises: obtaining a second distance from the center of the feature space to the boundary of the feature space; The calculation according to the first distance to obtain the first probability that the current pixel point is the pixel point corresponding to the target includes: calculating a ratio of the first distance to the second distance; A first probability corresponding to the current pixel point is calculated according to the proportional value, where the proportional value is negatively correlated with the first probability. 6. The method according to claim 4, wherein the step of obtaining the image target recognition model comprises: Obtain a training image, and obtain a training area corresponding to the target object in the training image; Acquiring training image features corresponding to each pixel in the training area; Model training is performed according to the training image features, and a feature mapping function for mapping the training image features to a minimum feature space and a central value of the feature space are obtained. 7. The method according to claim 1, wherein the image feature comprises a contrast feature, and the step of obtaining the contrast feature corresponding to the current pixel comprises: Acquire a third area and a fourth area on the current image according to the position of the current pixel, wherein the fourth area is a sub-area of the third area, and the current pixel is located in the fourth inside the area; acquiring non-overlapping image areas between the third area and the fourth area; performing statistics on the grayscale values of the pixels corresponding to the non-overlapping image regions to obtain a first statistical result, and performing statistics on the grayscale values of the pixels corresponding to the fourth region to obtain a second statistical result; The contrast feature is obtained according to the first statistical result and the second statistical result. 8. The method according to claim 1, wherein said identifying the image area where the target object is located in the current image according to the background similarity corresponding to each pixel point and the first probability comprises: Obtain the second probability corresponding to the current pixel according to the background similarity, wherein the second probability is the probability that the current pixel belongs to the target object pixel, and the second probability is the same as the background similarity There is a negative correlation; determining a current target probability corresponding to the current pixel according to the first probability and the second probability, where the current target probability is a probability that the current pixel belongs to a pixel of a target object; The image area where the target object is located in the current image is obtained by identifying according to the target probability corresponding to each pixel in the current image. 9. The method according to claim 8, wherein the identifying the image area where the target object is located in the current image according to the target probability corresponding to each pixel in the current image comprises: Acquiring a first pixel point in the current image whose target probability is greater than a first threshold; Obtaining the distribution characteristics of the target probability corresponding to the first pixel point; obtaining a second threshold according to the distribution feature; Acquiring an area obtained by combining first pixel points in the current image with a target probability greater than the second threshold as the image area where the target object is located in the current image. 10. An image recognition device, said device comprising: The current image acquisition module is used to acquire the current image of the target to be identified; A background area determination module, configured to obtain a current pixel point from the current image, and determine a corresponding background reference area according to the position of the current pixel point; A similarity calculation module, configured to calculate the background similarity corresponding to the current pixel according to the background reference area; The first probability obtaining module is used to process the image feature corresponding to the current pixel point according to the trained image target recognition model to obtain the first probability corresponding to the current pixel point, and the first probability is the current The probability that the pixel belongs to the pixel of the target object; A target area identification module, configured to calculate the background similarity and first probability corresponding to each pixel in the current image, and identify the background similarity and first probability corresponding to each pixel to obtain the background similarity and first probability in the current image The image area where the target object is located. 11. The device according to claim 10, wherein the similarity calculation module comprises: a gray value obtaining unit, configured to obtain a target pixel from the background reference area, and obtain a gray value corresponding to the target pixel and the current pixel; The reference vector composition unit is used to calculate a reference gray value according to the gray value of the target pixel point, perform an inverse operation on the reference gray value to obtain a corresponding complementary reference gray value, and convert the reference gray value to value and the complementary reference gray value to form a reference gray value vector; The current vector composition unit is used to invert the gray value corresponding to the current pixel to obtain a corresponding complementary current gray value, and combine the gray value corresponding to the current pixel with the complementary current gray value Form the current gray value vector; A similarity calculation unit, configured to calculate the background similarity corresponding to the current pixel according to the reference gray value vector and the current gray value vector. 12. The device according to claim 10, wherein the image target recognition model is a support vector clustering model, and the first probability obtaining module includes: A model parameter acquisition unit, configured to acquire a feature mapping function corresponding to the image target recognition model, and acquire a central value of a feature space corresponding to the feature mapping function; A mapping value calculation unit, configured to calculate the image feature according to the feature mapping function, to obtain a mapping value corresponding to the image feature; a first distance calculation unit, configured to calculate a first distance between the mapping value and the central value; The first probability obtaining unit is configured to calculate and obtain a first probability corresponding to the current pixel point according to the first distance, wherein the first distance has a negative correlation with the first probability. 13. The device according to claim 12, further comprising: A second distance acquisition module, configured to acquire a second distance from the center of the feature space to the boundary of the feature space; The first probability obtaining unit is used for: calculating a ratio of the first distance to the second distance; A first probability corresponding to the current pixel point is calculated according to the proportional value, where the proportional value is negatively correlated with the first probability. 14. A computer device, characterized in that it comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the process according to claims 1 to 9. The steps of the image recognition method according to any claim. 15. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the processor can perform any one of claims 1-9. A claim to the steps of the image recognition method.