WO2021057395A1

WO2021057395A1 - Heel type identification method, device, and storage medium

Info

Publication number: WO2021057395A1
Application number: PCT/CN2020/112536
Authority: WO
Inventors: 翟懿奎; 邓文博; 周文略; 柯琪锐; 甘俊英; 应自炉; 曾军英
Original assignee: 五邑大学
Priority date: 2019-09-29
Filing date: 2020-08-31
Publication date: 2021-04-01
Also published as: CN110705634A; CN110705634B

Abstract

A heel type identification method comprises the following steps: acquiring a heel image, performing pre-processing on the heel image, and obtaining a heel chromaticity diagram (S100); performing feature extraction on the heel chromaticity diagram by using a feature extraction network, and obtaining a feature output diagram (S200); processing the feature output diagram by using a region candidate network, and obtaining a heel candidate image (S300); performing pixel-based identification on the candidate image by means of an output network, and obtaining a heel height and a heel shape of the candidate image (S400); and performing, by means of a heel database, identification on the heel height and the heel shape, and obtaining a heel type (S500). The identification method improves the speed of heel identification, and increases accuracy of heel type identification, thereby reducing the workload of merchants.

Description

Method, device and storage medium for identifying shoe heel model

Technical field

The present invention relates to the technical field of image processing, in particular to a method, a device and a storage medium for identifying a heel model.

Background technique

Traditional shoe heel model recognition refers to the fact that businesses can vaguely judge the category and model of the heel provided by the customer through the visual judgment and search in memory through the image or physical object of the heel provided by the customer. With the booming economy, people’s living standards have been significantly improved, and the types of shoes have been greatly enriched. The corresponding heel market has also undergone great changes. According to incomplete statistics, every shoe heel nationwide The order of magnitude of the monthly update has reached 10,000, and the market demand can no longer be solved through visual identification and memory search. Secondly, the heel image provided by the customer has not been regulated by a unified collection system and collection standard. In the traditional heel recognition, it has increased the difficulty of the merchant, which consumes manpower and time, and cannot guarantee the heel model recognition. Accuracy.

Summary of the invention

In order to solve the above problems, the purpose of the present invention is to provide a heel model recognition method, device and storage medium, which can increase the speed of heel recognition, improve the accuracy of heel model recognition, and greatly reduce the workload of merchants.

The technical solution adopted by the present invention to solve the problem is: In the first aspect, an embodiment of the present invention proposes a method for identifying a heel model, which includes the following steps:

Acquiring a heel image, and preprocessing the heel image to obtain a heel chromaticity diagram;

Using a feature extraction network to perform feature extraction on the heel chromaticity map to obtain a feature output map;

Use the regional candidate network to process the feature output image to obtain a candidate image of the heel;

Perform pixel-level recognition on the candidate image through the output network to obtain the heel height and heel shape of the candidate image;

Identify the height of the heel and the shape of the heel through the heel database to obtain the model of the heel.

Further, collecting a heel image, and preprocessing the heel image to obtain a heel chromaticity diagram includes the following steps:

Obtain the heel image of the side of the heel at a horizontal angle within the range of the camera distance and the camera brightness range;

Using high-pass filtering to sharpen the heel image to obtain a sharpened image;

A bilateral filter is used to perform highlight denoising processing on the sharpened image to obtain a heel chromaticity diagram.

Further, within the range of the camera distance and the camera brightness, obtaining a heel image of the side of the heel at a horizontal angle includes the following steps:

Acquire the camera distance of the heel, and if the camera distance is not within the camera distance range, return the camera distance error message, and the camera distance range is 10 cm to 30 cm;

Acquire the camera brightness of the heel, if the camera brightness is not within the camera brightness range, return the camera brightness error message, and the camera brightness range is that the brightness superimposed value of the three color channels of red, green and blue is not less than 0.4;

Obtain the camera focal length of the heel and the heel image at the horizontal angle of the side of the heel.

Further, the feature extraction network includes: a residual network and a feature pyramid network; the residual network includes a number of residual blocks, the feature pyramid network includes a number of feature pyramid network layers, and the residual block is connected to some The feature pyramid network layer.

Further, using the regional candidate network to process the feature output map to obtain the candidate image of the heel includes the following steps:

Recognizing the heel of the feature output map by using the regional candidate network to obtain several candidate regions;

Obtaining the confidence of the candidate region by using a classifier, and screening the candidate region according to the confidence to obtain the confidence candidate region;

Acquiring the degree of overlap between the candidate confidence regions to obtain a data set of the degree of overlap between the candidate confidence regions;

Using non-maximum value suppression to process the region overlap degree data set to obtain an optimal candidate region;

The bilinear interpolation method is used to perform alignment processing on the optimal candidate region to obtain a candidate image.

Further, performing pixel-level recognition on the candidate image through the output network to obtain the heel height and heel shape of the candidate image includes the following steps:

Performing pixel-level recognition on the candidate image by using a segmentation network to obtain heel pixels in the candidate image;

Obtain the actual heel height according to the heel pixel points and the camera focal length;

A classification network is used to classify the heel pixels to obtain the heel shape of the candidate image.

Further, recognizing the height of the heel and the shape of the heel through the heel database to obtain the model of the heel includes the following steps:

Input the heel height and the heel shape into the heel database;

Obtain the heel height and the heel overlap degree between the heel shape and the data in the heel database, and filter and arrange the heel overlap degree to obtain the heel overlap degree according to the heel overlap degree Several heel models arranged by value.

In the second aspect, an embodiment of the present invention also provides a heel model recognition device, which includes at least one control processor and a memory for communicating with the at least one control processor; At least one instruction executed by the control processor, the instruction being executed by the at least one control processor, so that the at least one control processor can execute the method for identifying a heel model as described in any one of the above.

In the third aspect, the embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute any of the above A method for identifying the heel model described.

The technical solutions provided in the embodiments of the present invention have at least the following beneficial effects: collecting heel images, preprocessing the heel images, and enhancing the clarity and resolution of the heel images; using the feature extraction network to analyze the heel chromaticity diagram Perform feature extraction to improve the speed of feature extraction and enhance the resolution of features; use the regional candidate network to identify, filter, and classify the feature output map to obtain a normalized candidate image, which improves the accuracy of candidate region selection , Reduce the overlap between the candidate regions; through the output network to perform pixel-level recognition of the candidate image, improve the accuracy of the heel height and shape acquisition; through the heel database to store the heel information, the candidate image of the shoe The recognition of heel height and shape improves the speed and efficiency of heel recognition.

Description of the drawings

The present invention will be further explained below with reference to the drawings and examples.

Fig. 1 is an overall flowchart of an embodiment of a method for identifying a heel model of the present invention.

detailed description

With the booming economy, people’s living standards have been significantly improved, and the types of shoes have been greatly enriched; the corresponding heel market has also undergone great changes, and the national heels have been updated by the order of magnitude every month. Up to 10,000 levels, through visual identification and memory search, it is no longer able to solve the market demand; and the heel image provided by the customer has not been standardized by a unified collection system and collection standard, which increases the difficulty of merchant identification, which consumes manpower and time. And the accuracy of the heel model identification cannot be guaranteed.

Based on this, the present invention provides a heel model recognition method, device and storage medium, which can increase the speed of heel recognition, improve the accuracy of heel model recognition, and greatly reduce the workload of merchants.

The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

1, an embodiment of the present invention provides a method for identifying a heel model, which includes the following steps:

Step S100: Collect a heel image, and preprocess the heel image to obtain a heel chromaticity diagram;

Step S200: Perform feature extraction on the heel chromaticity map by using a feature extraction network to obtain a feature output map;

Step S300: Use the regional candidate network to process the feature output map to obtain a candidate image of the heel;

Step S400: Perform pixel-level recognition on the candidate image through the output network to obtain the heel height and heel shape of the candidate image;

Step S500: Recognizing the height of the heel and the shape of the heel through the heel database to obtain the model of the heel.

In this embodiment, step S100 collects a heel image and preprocesses the heel image to enhance the clarity and resolution of the heel image; wherein, the preprocessing can be set to sharpening processing, denoising processing, and brightness adjustment. , Saturation adjustment, etc. Step S200 uses the feature extraction network to perform feature extraction on the heel chromaticity diagram, which improves the speed of feature extraction and enhances the resolution of features; Step S300 uses the regional candidate network to identify, filter, and classify the feature output map to obtain the classification The unified candidate image improves the accuracy of candidate region selection and reduces the repetition between candidate regions; step S400 performs pixel-level recognition on the candidate image through the output network, which improves the heel height and heel shape of the candidate image Accuracy of acquisition: Step S500 stores the heel information through the heel database, and recognizes and compares the heel height and heel shape of the candidate image, which improves the speed and efficiency of heel model recognition.

Further, another embodiment of the present invention also provides a method for recognizing a heel model, in which collecting a heel image and preprocessing the heel image to obtain a heel chromaticity diagram includes the following steps:

Step S110: Acquire a heel image of the side of the heel at a horizontal angle within the camera distance range and the camera brightness range;

Step S120: Use high-pass filtering to perform sharpening processing on the heel image to obtain a sharpened image;

Step S130: Use a bilateral filter to perform highlight denoising processing on the sharpened image to obtain a heel chromaticity diagram.

In this embodiment, step S110 collects images of the heel within the range of the camera distance and the scope of the camera brightness to ensure the clarity of the heel image; the image collection is performed at a horizontal angle to avoid the difference in the shape of the heel. , To ensure the accuracy of the shape of the heel in the heel image; to acquire the image of the side of the heel, so that the image can acquire more characteristic points of the heel, for example, to acquire the image of the side of the heel of a high-heeled shoe. The heel slope and thickness of the heel can be obtained from the side image of the heel of a high-heeled shoe, but the above data of the heel cannot be obtained on the front, back and bottom of the heel of a high-heeled shoe. Therefore, the image of the side of the heel is acquired. It can improve the accuracy of heel recognition.

Step S120 High-pass filtering is a filtering method. The rule is that high-frequency signals can pass normally, while low-frequency signals below the set critical value are blocked and attenuated. That is, high-pass filtering is only for those below a given frequency. The frequency component has an attenuation effect, and the frequency component above the cutoff frequency is allowed to pass, and there is no phase shift filtering process; it is mainly used to eliminate low frequency noise, also called a low cut filter.

Input the heel image into the following formula for high-pass filtering:

y(n,m)=x(n,m)+λz(n,m);

Among them, x(n,m) is the heel image, y(n,m) is the sharpened image after high-pass filtering, and z(n,m) is the correction signal, which is generally obtained by high-pass filtering on x. λ is a scaling factor used to control the enhancement effect. The high-frequency part of the heel image is extracted through high-pass filtering, and the high-frequency part is superimposed on the heel image, thereby enhancing the edge information of the heel image, achieving the effect of sharpening the heel image, and improving the quality of the heel image. Clarity.

Step S130 Bilateral filtering is a nonlinear filtering method that combines the spatial proximity of the image and the similarity of the value domain. It also considers the spatial information and the maximum diffuse reflection chromaticity similarity to achieve edge preservation and denoising. The purpose is simple, non-iterative, and the output depends on the weighted combination of neighboring pixel values. Use the estimated maximum diffuse reflection chromaticity value of the pixel as a weighted combination of the value domain and the spatial domain to guide smoothing, and perform denoising and edge protection on the sharpened image, and then retrieve the maximum chromaticity value of each pixel, which is about to sharpen the image Enter into the following formula for bilateral filtering:

Among them, D is the spatial weight function, R is the estimated maximum diffuse reflection chromaticity similarity weight function, p is the pixel point after bilateral filtering, q is the pixel point of the sharpened image, and σ _max is the highlight Maximum diffuse chromaticity,

Is the maximum diffuse chromaticity in the state of no highlight, Λmax(x) is the estimated maximum diffuse chromaticity.

After bilateral filtering is performed on the maximum chromaticity value of the pixel with specular reflection, the chromaticity value will be reduced, so that the maximum chromaticity of the filtered pixel is closer to the true maximum diffuse chromaticity. At the same time, the estimated chromaticity of pixels that only contain diffuse reflection will also be affected by pixels that contain specular reflection and become smaller. Therefore, in order to reduce the influence of a pixel with specular components on the chromaticity of a pixel with only diffuse reflection, the maximum diffuse reflection chromaticity σ _max when the pixel contains highlights can be compared with the estimated maximum diffuse reflection chromaticity σ max under the state of no highlights. Diffuse chromaticity

And take the maximum value as the maximum chromaticity value of each pixel:

_{The σ max} in the above formula is iterated using bilateral filtering to make the maximum diffuse reflection chromaticity diagram of the same color smooth. This article compares the filtered value after each iteration

And σ _max , when their difference is less than the threshold at each pixel, the filter value is considered to converge and the iteration is completed; wherein the threshold at the pixel can be set according to the actual situation, for example, set to 0.02.

Perform highlight denoising processing on sharpened images through bilateral filters, and filter values

Iteratively compare processing with σ _max to obtain the heel chromaticity diagram containing the maximum chromaticity value of each pixel of the heel, which improves the effect of heel pixel recognition, thereby improving the recognition effect of the heel image.

Further, another embodiment of the present invention also provides a method for recognizing a heel model, in which, within the camera distance range and the camera brightness range, acquiring the heel image of the side of the heel at a horizontal angle includes the following steps:

Step S111: Obtain the camera distance of the heel, and if the camera distance is not within the camera distance range, return the camera distance error message, and the camera distance range is 10 cm to 30 cm;

Step S112: Obtain the camera brightness of the heel. If the camera brightness is not within the camera brightness range, return the camera brightness error message, and the camera brightness range is that the brightness superimposed value of the three color channels of red, green, and blue is not Less than 0.4;

Step S113: Obtain the camera focal length of the heel and the heel image of the side of the heel at a horizontal angle.

In this embodiment, the camera distance range of step S111 is set to be 10 cm to 30 cm, so that the size of the heel image acquired by the camera device is within a certain range, so that the camera device can completely obtain the heel image without avoiding the heel The size of the image is too small or too large, and the definition of the heel image obtained by the camera equipment is guaranteed; when the user collects the heel image, the size of the camera distance is obtained. When the camera distance is not within the camera distance range, Return the error information of the camera distance, allowing the user to adjust the camera distance through the error information of the distance, ensuring the accuracy of the heel image acquisition.

In step S112, the camera brightness of the heel can be embodied according to the superimposed values of the brightness of the heel image in the three color channels of red, green, and blue, that is, the calculation formula of the camera brightness is:

Brightness (RGB = 0.26 red (R + 0.67 green (G + 0.07 blue (B;

Among them, brightness (RGB is the brightness superposition value of the three color channels of red, green and blue, red (R is the brightness value of the heel image in the red channel, and green (G is the brightness value of the heel image in the green channel, Blue (B is the brightness value of the heel image in the blue channel. When the brightness overlay value is less than 0.4, the brightness of the heel image is darker, which is not conducive to image processing. Therefore, the user collects the heel image When the camera brightness is obtained, when the camera brightness is less than 0.4, the camera brightness error message is returned, so that the user can adjust the camera brightness through the brightness error message, which ensures the accuracy of the heel image acquisition.

Step S113 The camera focal length of the heel is automatically adjusted by the camera equipment according to the actual shooting environment. By obtaining the camera focal length of the heel, it is helpful to calculate the actual height of the heel; image acquisition of the heel at a horizontal angle , To ensure the accuracy of the heel shape in the heel image; image acquisition of the side of the heel, so that the image can acquire more characteristic points of the heel, and improve the accuracy of heel recognition.

Further, another embodiment of the present invention also provides a method for recognizing a shoe heel model, wherein the feature extraction network includes: a residual network and a feature pyramid network; the residual network includes a plurality of residual blocks, the The feature pyramid network includes several feature pyramid network layers, and the feature pyramid network layer is connected behind the residual block.

In this embodiment, the characteristic of the residual network is that it is easy to optimize and can increase the accuracy by increasing a considerable depth. The internal residual block uses jump connections, which alleviates the increase in depth in the deep neural network. The problem of gradient disappearance; the feature pyramid is used to detect and recognize objects of different scales, and the inherent multi-scale pyramid hierarchy of deep convolutional networks is used to construct feature gold characters with marginal additional losses, so that the network has a horizontal connection from top to bottom The architecture of, can build high-level semantic feature maps on all scales. The residual network includes several residual blocks, and a characteristic pyramid network layer can be connected behind any residual block, that is, in several residual blocks, at most, a characteristic pyramid network layer can be connected after each residual block. . By inputting the heel chromaticity diagram into the feature extraction network for feature extraction, the speed of heel feature extraction is improved, and the resolution of heel features is enhanced. Among them, the number of layers of the feature pyramid network is not limited, and it is set according to the actual number of residual blocks.

Further, another embodiment of the present invention also provides a method for recognizing a shoe heel model, wherein, using a regional candidate network to process the feature output map to obtain a candidate image of the heel includes the following steps:

Step S310: Recognizing the heel of the characteristic output map by using the region candidate network to obtain several candidate regions;

Step S320: Obtain the confidence of the candidate region by using a classifier, and screen the candidate region according to the confidence to obtain the candidate confidence region;

Step S330: Obtain the area overlap degree between the confidence candidate areas, and obtain the area overlap degree data set of the confidence candidate area;

Step S340: Use non-maximum value suppression to process the region overlap degree data set to obtain an optimal candidate region;

Step S350: Perform alignment processing on the optimal candidate region using a bilinear interpolation method to obtain a candidate image.

In this embodiment, in step S310, the area candidate network uses a sliding window to traverse all points on the feature output map, judges all regions of interest on the feature output map, and obtains several candidate regions. Step S320 uses the classifier to calculate the confidence of the candidate regions, and according to the size of the confidence, selects a number of candidate regions with the highest confidence, which are recorded as confidence candidate regions; Step S330 obtains the degree of overlap between the confidence candidate regions Obtain the area overlap data set of the confidence candidate area, that is, the area overlap data set of each confidence candidate area contains the data of the area overlap with the candidate area and other candidate areas.

Step S340 Non-maximum suppression is to suppress elements that are not maximum values, that is, to select a local maximum search. This local represents a neighborhood. The neighborhood has two variable parameters, one is the dimension of the neighborhood, and the other is The size of the neighborhood. The specific search steps for the regional overlap data set are: start with the confidence candidate region with the highest confidence, record it as the first confidence candidate region, and filter from the region overlap data set of the first confidence candidate region Select the first-type confidence candidate region whose region overlap value is not greater than the threshold; then select the second confidence candidate region with the highest regional confidence from the first-type confidence candidate region, and select the second confidence candidate region in the second confidence candidate region. In the regional overlap degree data group, filter out the third-type confidence candidate regions whose regional overlap value is not greater than the threshold; continue the above-mentioned screening until the largest confidence candidate region among all the Nth-type confidence candidate regions It is selected and merged to obtain the optimal candidate area of the heel, which improves the accuracy of candidate area selection and reduces the overlap between candidate areas. Wherein, the range of the N-1th type of confidence candidate area includes the Nth type of confidence candidate area.

For example, the confidence candidate regions include A, B, C, D, E, and F, and the region confidence level of the confidence candidate region is A<B<C<D<E<F, first extract the F with the highest regional confidence , Mark F as the first confidence candidate region; in the F’s region overlap data set, filter out the confidence candidate regions whose region overlap is not greater than the threshold, assuming that A and F, B and F, C and F If the regional overlap value is not greater than the threshold, then A, B, and C will be screened out, and record the first type of confidence candidate regions; because the regional confidence of C is greater than the regional confidence of A and B, mark C as the second confidence Candidate area. At this time, in the area overlap degree data set of C, only the area overlap degree of C and A, and C and B are the area overlap degrees. The overlap degree of these two areas is filtered, and the area overlap degree value is not greater than the threshold A; then the most The optimal candidate area is the merged area of F, C, and A.

In step S350, the bilinear interpolation is a linear interpolation extension of the interpolation function of two variables, and its core idea is to perform linear interpolation in two directions respectively. Select four fixed-position pixels in the optimal candidate area, and perform bilinear interpolation on these four fixed-position pixels. The bilinear interpolation process is: for each fixed-position pixel, the optimal The four heel pixels adjacent to it are selected in the candidate area, and the four heel pixels are linearly interpolated in the horizontal and vertical directions. That is, according to the fixed position of the pixel and its four heel pixels. The distance between the two determines the corresponding weight, so as to calculate the interpolation position of the pixel at a fixed position. That is, the principle of bilinear interpolation is: taking the distance from the pixel at a fixed position to the four nearest heel pixels as the reference weight, and after two linear interpolations, the interpolation position of the pixel at the fixed position is obtained; According to the interpolation position of the four fixed pixel points, the optimal candidate area is aligned to obtain a normalized candidate image, which improves the accuracy of the heel image recognition.

Further, another embodiment of the present invention also provides a method for recognizing a heel model, wherein the candidate image is identified at the pixel level through an output network to obtain the heel height and heel shape of the candidate image, Including the following steps:

Step S410: Perform pixel-level recognition on the candidate image by using a segmentation network to obtain heel pixels in the candidate image;

Step S420: Obtain the actual heel height according to the heel pixel points and the camera focal length;

Step S430: Use a classification network to classify the heel pixels to obtain the heel shape of the candidate image.

In this embodiment, there are many types of segmentation networks in step S410. Commonly used segmentation networks are: FCN, UNet, SegNet, DeepLab, etc. The segmentation network is to identify and classify candidate images at the pixel level to obtain the pixels of the candidate image. The types of points are: heel pixels and non-heel pixels. At the same time, the pixel points of the candidate image are filtered to obtain the heel pixels.

In step S420, the height of the heel in the candidate image can be calculated according to the position of the pixel point of the heel, and the actual height of the heel can be calculated by the following formula:

Among them, f is the camera focal length, h is the heel height of the candidate image, D is the camera distance, and H is the actual heel height.

Step S430 There are many types of classification networks. Commonly used classification networks are: LeNet-5, AlexNet, ZFNet, VGGNet, GoogLeNet, ResNet, etc.; classification networks mainly use convolution, parameter sharing, pooling and other operations to extract features, and use The fully connected neural network classifies and recognizes features, reducing a large amount of calculations between data. Among them, a classification network is used to classify the heel pixels to obtain a shape composed of all pixels of the same heel, that is, to obtain the heel shape of the candidate image.

Further, another embodiment of the present invention also provides a method for identifying a heel model, wherein the heel height and the shape of the heel are identified through a heel database to obtain the model of the heel, including the following step:

Step S510: Input the height of the heel and the shape of the heel into the heel database;

Step S520: Obtain the heel height and the heel overlap between the heel shape and the data in the heel database, and filter and arrange the heel overlap to obtain the heel A number of heel models arranged by the value of the degree of overlap.

In this embodiment, a large amount of heel information is stored in the heel database of step S520. For the same heel, the heel information includes the real heel height and the real heel shape; by calculating the heel height and the shoe heel Heel shape, the heel overlap degree between the real heel height and the real heel shape in the heel database, sorted according to the degree of heel overlap, and selected a number of heels with larger heel overlap Model, output several heel models according to the degree of overlap value, so that users can obtain the several heel models with the highest similarity in the heel image, improve the accuracy of heel model recognition, and greatly reduce the business’s burden. Workload. Among them, a number of heel models with greater heel overlap can be set to ten heel models with the largest heel overlap.

In addition, referring to Fig. 1, another embodiment of the present invention also provides a method for identifying a heel model, which includes the following steps:

Step S112: Obtain the camera brightness of the heel. If the camera brightness is not within the camera brightness range, return the camera brightness error message. The camera brightness range is that the brightness superimposed values of the three color channels of red, green, and blue are not Less than 0.4;

Step S113: Obtain the camera focal length of the heel and the heel image of the side of the heel at a horizontal angle;

Step S130: Use a bilateral filter to perform highlight denoising processing on the sharpened image to obtain a heel chromaticity diagram;

Step S350: aligning the optimal candidate area by using a bilinear interpolation method to obtain a candidate image;

Step S430: Use a classification network to classify the heel pixels to obtain the heel shape of the candidate image;

In this embodiment, the heel image is collected, and the heel image is preprocessed to enhance the clarity and resolution of the heel image; the feature extraction network is used to perform feature extraction on the heel chromaticity diagram, which improves the speed of feature extraction , To enhance the resolution of features; use the regional candidate network to identify, filter, and classify the feature output map to obtain a normalized candidate image, which improves the accuracy of candidate region selection and reduces the overlap between candidate regions Pixel-level recognition of candidate images through the output network improves the accuracy of heel height and shape acquisition; through the storage of heel information in the heel database, the recognition of the height and shape of the heel of the candidate image improves Speed and efficiency of heel recognition.

In addition, another embodiment of the present invention also provides a heel model identification device, which includes at least one control processor and a memory for communicating with the at least one control processor; The instructions executed by the at least one control processor, the instructions are executed by the at least one control processor, so that the at least one control processor can execute the method for identifying a heel model as described in any one of the above.

In this embodiment, the identification device includes: one or more control processors and memories, and the control processors and memories may be connected by a bus or in other ways.

As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions/modules corresponding to the identification method in the embodiment of the present invention. The control processor executes various functional applications and data processing of the identification device by running the non-transitory software programs, instructions, and modules stored in the memory, that is, realizes the identification method of the foregoing method embodiment.

The memory may include a program storage area and a data storage area, where the program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created according to the use of the identification device and the like. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely provided with respect to the control processor, and these remote memories may be connected to the identification device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

The one or more modules are stored in the memory, and when executed by the one or more control processors, the identification method in the above method embodiment is executed, for example, the steps S100 to S500, S110 to S130, S111 to S113, S310 to S350, S410 to S430, and S510 to S520 functions.

The embodiment of the present invention also provides a computer-readable storage medium, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by one or more control processors, for example, a control processor Execution may cause the above-mentioned one or more control processors to execute the identification method in the above-mentioned method embodiment, for example, execute the above-described method steps S100 to S500, S110 to S130, S111 to S113, S310 to S350, S410 to S430, And the functions of S510 to S520.

The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

Through the description of the above implementation manners, those skilled in the art can clearly understand that each implementation manner can be implemented by means of software plus a general hardware platform. Those skilled in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by computer programs instructing relevant hardware. The programs can be stored in a computer readable storage medium. At this time, it may include the flow of the embodiment of the above-mentioned method. Wherein, the storage medium may be a magnetic disk, an optical disc, a read-only memory (Read Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.

The above is a detailed description of the preferred implementation of the present invention, but the present invention is not limited to the above implementation

In this way, those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

A method for identifying a heel model, which is characterized in that it comprises the following steps:

Acquiring a heel image, and preprocessing the heel image to obtain a heel chromaticity diagram;

Using a feature extraction network to perform feature extraction on the heel chromaticity map to obtain a feature output map;

Use the regional candidate network to process the feature output image to obtain a candidate image of the heel;

Perform pixel-level recognition on the candidate image through the output network to obtain the heel height and heel shape of the candidate image;

Identify the height of the heel and the shape of the heel through the heel database to obtain the model of the heel.
The method for recognizing a shoe heel model according to claim 1, characterized in that: collecting a heel image and preprocessing the heel image to obtain a heel chromaticity diagram comprises the following steps:

Obtain the heel image of the side of the heel at a horizontal angle within the range of the camera distance and the camera brightness range;

Using high-pass filtering to sharpen the heel image to obtain a sharpened image;

A bilateral filter is used to perform highlight denoising processing on the sharpened image to obtain a heel chromaticity diagram.
The method for recognizing a shoe heel model according to claim 2, characterized in that: acquiring a heel image of the side of the heel at a horizontal angle within the range of the camera distance and the camera brightness includes the following steps:

Acquire the camera distance of the heel, and if the camera distance is not within the camera distance range, return the camera distance error message, and the camera distance range is 10 cm to 30 cm;

Acquire the camera brightness of the heel, if the camera brightness is not within the camera brightness range, return the camera brightness error message, and the camera brightness range is that the brightness superimposed value of the three color channels of red, green and blue is not less than 0.4;

Obtain the camera focal length of the heel and the heel image at the horizontal angle of the side of the heel.
The method for recognizing a shoe heel model according to claim 1, wherein the feature extraction network includes: a residual network and a feature pyramid network; the residual network includes a plurality of residual blocks, and the feature pyramid The network includes several feature pyramid network layers, and the feature pyramid network layer is connected behind the residual block.
A method for recognizing a shoe heel model according to claim 1, characterized in that: using a regional candidate network to process the feature output map to obtain a candidate image of the heel includes the following steps:

Recognizing the heel of the feature output map by using the regional candidate network to obtain several candidate regions;

Obtaining the confidence of the candidate region by using a classifier, and screening the candidate region according to the confidence to obtain the confidence candidate region;

Acquiring the degree of overlap between the candidate confidence regions to obtain a data set of the degree of overlap between the candidate confidence regions;

Using non-maximum value suppression to process the region overlap degree data set to obtain an optimal candidate region;

The bilinear interpolation method is used to perform alignment processing on the optimal candidate region to obtain a candidate image.
The method for recognizing a heel model according to claim 3, characterized in that: performing pixel-level recognition on the candidate image through an output network to obtain the heel height and heel shape of the candidate image, comprising the following steps:

Performing pixel-level recognition on the candidate image by using a segmentation network to obtain heel pixels in the candidate image;

Obtain the actual heel height according to the heel pixel points and the camera focal length;

A classification network is used to classify the heel pixels to obtain the heel shape of the candidate image.
The method for recognizing a heel model according to claim 1, characterized in that: recognizing the height of the heel and the shape of the heel through a heel database to obtain the model of the heel comprises the following steps:

Input the heel height and the heel shape into the heel database;

Obtain the heel height and the heel overlap degree between the heel shape and the data in the heel database, and filter and arrange the heel overlap degree to obtain the heel overlap degree according to the heel overlap degree Several heel models arranged by value.
A shoe heel model recognition device, which is characterized in that it comprises at least one control processor and a memory for communicating with the at least one control processor; the memory stores the memory that can be executed by the at least one control processor. Instructions, the instructions are executed by the at least one control processor, so that the at least one control processor can execute the heel model identification method according to any one of claims 1-7.
A computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute the shoe of any one of claims 1-7 Follow the model identification method.