CN115375412A

CN115375412A - Intelligent commodity recommendation processing method and system based on image recognition

Info

Publication number: CN115375412A
Application number: CN202211200192.1A
Authority: CN
Inventors: 郭哲; 邱雯婷
Original assignee: Individual
Current assignee: Shangcha Industrial Co ltd
Priority date: 2022-09-29
Filing date: 2022-09-29
Publication date: 2022-11-22
Anticipated expiration: 2042-09-29
Also published as: CN115375412B

Abstract

The invention provides an intelligent commodity recommendation processing method and system based on image recognition, and relates to the technical field of image processing. In the invention, the commodity recommendation request is analyzed to output a corresponding commodity request description image sequence. And for each candidate commodity image sequence, calculating the image sequence matching degree of the candidate commodity image sequence and the commodity request description image sequence by using an image matching neural network so as to output the corresponding image sequence matching degree. And screening at least one target commodity image sequence from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence and the commodity request description image sequence, and sending the target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal equipment. Based on the method, the reliability of commodity recommendation can be improved to a certain extent.

Description

Intelligent commodity recommendation processing method and system based on image recognition

Technical Field

The invention relates to the technical field of image processing, in particular to a commodity intelligent recommendation processing method and system based on image recognition.

Background

Commodity recommendation is an important link in internet behaviors (such as online shopping, commodity inquiry and the like), wherein the reliability of commodity recommendation directly influences the stickiness of a user to a corresponding platform. Therefore, it is necessary to provide a scheme that can reliably perform recommendation of an article. However, in the prior art, the recommendation of the product is generally limited to text search, and then the content or information described by the text generally has a large limitation, so that the limitation of the reliability is large.

Disclosure of Invention

In view of the above, the present invention provides a method and a system for intelligent commodity recommendation processing based on image recognition, so as to improve the reliability of commodity recommendation to a certain extent.

In order to achieve the above purpose, the embodiment of the invention adopts the following technical scheme:

an intelligent commodity recommendation processing method based on image recognition is applied to a commodity recommendation server and comprises the following steps:

when a commodity recommendation request of a target user terminal device is received, analyzing the commodity recommendation request to output a corresponding commodity request description image sequence;

for each candidate commodity image sequence in the configured multiple candidate commodity image sequences, calculating the image sequence matching degree of the candidate commodity image sequence and the commodity request description image sequence by utilizing an image matching neural network formed by training so as to output the image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence;

and screening at least one target commodity image sequence from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence in the candidate commodity image sequences and the commodity request description image sequence, and sending target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal equipment, wherein the target commodity description information at least comprises the corresponding target commodity image sequence, and the target commodity image sequence is formed by carrying out image acquisition on the corresponding to-be-recommended commodity.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, when a commodity recommendation request of a target user terminal device is received, the step of parsing the commodity recommendation request to output a corresponding commodity request description image sequence includes:

when a commodity recommendation request of a target user terminal device is received, analyzing the commodity recommendation request to determine each frame of commodity request description image contained in the commodity recommendation request and image sequencing number information corresponding to each frame of commodity request description image;

and sequencing and combining each frame of commodity request description image contained in the commodity recommendation request according to the image sequencing number information corresponding to each frame of commodity request description image so as to form a commodity request description image sequence corresponding to the commodity recommendation request.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, the step of calculating, for each candidate commodity image sequence in the configured plurality of candidate commodity image sequences, an image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence by using an image matching neural network formed by training to output an image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence includes:

for each candidate commodity image sequence in the configured multiple candidate commodity image sequences, utilizing an image matching neural network formed by pre-training to respectively calculate the image matching degree of each frame of candidate commodity image included in the candidate commodity image sequence and each frame of commodity request description image included in the commodity request description image sequence so as to output the image matching degree between each frame of candidate commodity image and each frame of commodity request description image;

and for each candidate commodity image sequence in the configured multiple candidate commodity image sequences, performing data fusion processing on the image matching degree between each frame of candidate commodity image included in the candidate commodity image sequence and each frame of the commodity request description image to output the image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, the step of performing, by using an image matching neural network formed by pre-training for each candidate commodity image sequence in the configured plurality of candidate commodity image sequences, image matching degree calculation for each candidate commodity image included in the candidate commodity image sequence and each commodity request description image included in the commodity request description image sequence respectively to output an image matching degree between each candidate commodity image and each commodity request description image includes:

respectively performing information conversion processing on the candidate commodity image and the commodity request description image to output a candidate commodity image block ordered set corresponding to the candidate commodity image and a request description image block ordered set corresponding to the commodity request description image;

respectively carrying out pixel point identification processing on the candidate commodity image and the commodity request description image based on a pre-configured important pixel point identification algorithm so as to identify a candidate important pixel point ordered set corresponding to the candidate commodity image and a request important pixel point ordered set corresponding to the commodity request description image from the candidate commodity image;

performing corresponding substitution processing on the candidate commodity image block ordered set and the request description image block ordered set based on a preset image block corresponding relation to form image block substitution information corresponding to the candidate commodity image block ordered set and the request description image block ordered set;

loading the image block substitution information into an image matching neural network formed by pre-training for feature mining processing so as to output image block feature distribution corresponding to the image block substitution information;

respectively carrying out feature mining processing on the candidate important pixel point ordered set and the request important pixel point ordered set based on the image block feature distribution so as to output the candidate important pixel point feature distribution corresponding to the candidate important pixel point ordered set and the request important pixel point feature distribution corresponding to the request important pixel point ordered set;

respectively loading the candidate important pixel point feature distribution and the request important pixel point feature distribution to a weight attention sub-network included in the image matching neural network for processing so as to output candidate important pixel point weight feature distribution corresponding to the candidate important pixel point feature distribution and request important pixel point weight feature distribution corresponding to the request important pixel point feature distribution;

and matching the weight characteristic distribution of the candidate important pixel points and the weight characteristic distribution of the request important pixel points based on an image matching sub-network included in the image matching neural network so as to output the image matching degree between the candidate commodity image and the commodity request description image.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, the step of performing information conversion processing on the candidate commodity image and the commodity request description image respectively to output an ordered candidate commodity image block set corresponding to the candidate commodity image and an ordered request description image block set corresponding to the commodity request description image includes:

respectively carrying out identification screening processing on interference pixels in the candidate commodity image and the commodity request description image so as to output a candidate commodity screening image corresponding to the candidate commodity image and a commodity request description screening image corresponding to the commodity request description image;

and respectively carrying out information conversion processing on the candidate commodity screening image and the commodity request description screening image to form a candidate commodity image block ordered set corresponding to the candidate commodity image and a request description image block ordered set corresponding to the commodity request description image.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, the step of performing corresponding substitution processing on the ordered candidate commodity image block set and the ordered request description image block set based on a preconfigured image block correspondence relationship to form image block substitution information corresponding to the ordered candidate commodity image block set and the ordered request description image block set includes:

merging the image blocks of the candidate commodity image block ordered set and the request description image block ordered set to form a corresponding image block merged ordered set;

performing corresponding substitution processing on the image block merged ordered set based on a preset image block corresponding relation so as to output image block corresponding substitution data corresponding to each image block included in the image block merged ordered set;

respectively determining alternative data corresponding to the set position corresponding to each image block based on the image block set position corresponding to each image block included in the image block merging ordered set;

and carrying out merging operation on the alternative data corresponding to the image block and corresponding to each image block in the image block merging ordered set and the alternative data corresponding to the set position so as to output the image block alternative information corresponding to the image block merging ordered set.

In some preferred embodiments, in the above commodity intelligent recommendation processing method based on image recognition, the loading the image block replacement information into an image matching neural network formed by pre-training for feature mining processing to output an image block feature distribution corresponding to the image block replacement information includes:

loading a first feature expression parameter preset in a weight attention sub-network and each corresponding substitute data in the image block substitute information included in an image matching neural network formed by pre-training into the weight attention sub-network for processing so as to output a weight coefficient of each corresponding substitute data;

comparing the execution quantity of the repeated feedback processing to determine whether the execution quantity of the repeated feedback processing meets the execution quantity requirement configured in advance;

splicing the weight coefficients output by performing the repeated feedback processing each time to form image block feature distribution corresponding to the image block substitution information on the assumption that the execution number of the repeated feedback processing meets the execution number requirement;

and if the execution quantity of the repeated feedback processing does not meet the execution quantity requirement, performing data fusion calculation on the weight coefficients and the first feature representation parameters to output corresponding weight coefficients, loading the weight coefficients into a recursion sub-network included in the image matching neural network to perform repeated feedback processing to output corresponding repeated feedback processing results, taking the repeated feedback processing results as new first feature representation parameters, and performing rotation to load each corresponding substitute data in the first feature representation parameters and the image block substitute information preset in a weight attention sub-network included in the pre-trained image matching neural network into the weight attention sub-network to perform processing so as to output the weight coefficients of each corresponding substitute data.

In some preferred embodiments, in the above method for intelligent recommendation processing of a commodity based on image recognition, the step of respectively performing feature mining processing on the ordered set of candidate important pixels and the ordered set of requested important pixels based on the image block feature distribution to output feature distribution of candidate important pixels corresponding to the ordered set of candidate important pixels and feature distribution of requested important pixels corresponding to the ordered set of requested important pixels includes:

combining to form a candidate feature distribution parameter ordered set corresponding to the candidate important pixel point ordered set based on the feature distribution parameter of each important pixel point in the image block feature distribution in the candidate important pixel point ordered set;

based on the characteristic distribution parameters of each image block in the request important pixel ordered set in the image block characteristic distribution, combining to form a request characteristic distribution parameter ordered set corresponding to the request important pixel ordered set;

according to whether the candidate important pixel point ordered sets belong to the same image block or not, performing fusion processing on the feature distribution parameters corresponding to each important pixel point in the candidate feature distribution parameter ordered sets to output the feature distribution of the candidate important pixel points corresponding to the candidate important pixel point ordered sets;

and performing fusion processing on the feature distribution parameters corresponding to each important pixel point included in the request feature distribution parameter ordered set according to whether the feature distribution parameters belong to the same image block or not so as to output the feature distribution of the request important pixel points corresponding to the request important pixel point ordered set.

In some preferred embodiments, in the above intelligent commodity recommendation processing method based on image recognition, the step of screening at least one target commodity image sequence from the multiple candidate commodity image sequences according to an image sequence matching degree between each candidate commodity image sequence in the multiple candidate commodity image sequences and the commodity request description image sequence, and then sending target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal device includes:

screening each candidate commodity image sequence of which the corresponding image sequence matching degree is greater than or equal to a configured image sequence matching degree reference value from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence of the candidate commodity image sequences and the commodity request description image sequence so as to output at least one target commodity image sequence;

and for each target commodity image sequence in the at least one target commodity image sequence, combining to form target commodity description information corresponding to the target commodity image sequence according to the target commodity image sequence and the commodity description text data of the commodity to be recommended corresponding to the target commodity image sequence, and then sending the target commodity description information to the target user terminal equipment.

The embodiment of the invention also provides an intelligent commodity recommendation processing system based on image recognition, which is applied to a commodity recommendation server, and the intelligent commodity recommendation processing system based on image recognition comprises:

the recommendation request analysis module is used for analyzing the commodity recommendation request when receiving the commodity recommendation request of the target user terminal equipment so as to output a corresponding commodity request description image sequence;

the matching degree calculation module is used for calculating the matching degree of the candidate commodity image sequence and the commodity request description image sequence by utilizing an image matching neural network formed by training for each candidate commodity image sequence in the configured multiple candidate commodity image sequences so as to output the matching degree of the candidate commodity image sequence and the commodity request description image sequence;

and the commodity recommending module is used for screening at least one target commodity image sequence from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence in the candidate commodity image sequences and the commodity request description image sequence, and then sending the target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal equipment, wherein the target commodity description information at least comprises the corresponding target commodity image sequence, and the target commodity image sequence is formed based on image acquisition of the corresponding to-be-recommended commodity.

The embodiment of the invention also provides a server, which comprises a processor and a memory; the processor is connected in communication with the memory, and the processor is configured to read the computer program from the memory and execute the computer program to implement the method of any one of claims 1 to 3.

The embodiment of the invention provides a commodity intelligent recommendation processing method and system based on image recognition, which are used for analyzing a commodity recommendation request so as to output a corresponding commodity request description image sequence. And for each candidate commodity image sequence, calculating the image sequence matching degree of the candidate commodity image sequence and the commodity request description image sequence by using an image matching neural network so as to output the corresponding image sequence matching degree. And screening at least one target commodity image sequence from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence and the commodity request description image sequence, and sending the target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal equipment. Through the content, the commodity search based on the image can be realized, wherein the commodity information generally contained in the image is rich, so that the reliability of the commodity search recommendation based on the commodity search recommendation can be higher, the reliability of the commodity recommendation is improved to a certain extent, and the problem of low reliability in the prior art is solved.

In order to make the aforementioned and other objects, features and advantages of the present invention comprehensible, preferred embodiments accompanied with figures are described in detail below.

Drawings

Fig. 1 is a block diagram of a structure of a product recommendation server according to an embodiment of the present invention.

Fig. 2 is a schematic flowchart illustrating steps of a commodity intelligent recommendation processing method based on image recognition according to an embodiment of the present invention.

Fig. 3 is a schematic diagram of modules included in the intelligent commodity recommendation processing system based on image recognition according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to fig. 1, an embodiment of the present invention provides a product recommendation server. Wherein the goods recommendation server may include a memory and a processor.

It will be further appreciated that in some possible embodiments, the memory and the processor are electrically connected, directly or indirectly, to enable data transfer or interaction. For example, they may be electrically connected to each other via one or more communication buses or signal lines. The memory may have stored therein at least one software function, which may be in the form of software or firmware. The processor may be configured to execute the executable computer program stored in the memory, so as to implement the intelligent commodity recommendation processing method based on image recognition provided by the embodiment of the present invention.

It is further noted that in some possible embodiments, the Memory may be, but is not limited to, random Access Memory (RAM), read Only Memory (ROM), programmable Read-Only Memory (PROM), erasable Read-Only Memory (EPROM), electrically Erasable Read-Only Memory (EEPROM), and the like. The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), a System on Chip (SoC), and the like; but may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components.

It may be further explained that, in some possible embodiments, the structure shown in fig. 1 is only an illustration, and the product recommendation server may further include more or fewer components than those shown in fig. 1, or have a different configuration than that shown in fig. 1, for example, include a communication unit for information interaction with other devices (for example, a user terminal device, such as a mobile phone, for making a product recommendation request).

With reference to fig. 2, an embodiment of the present invention further provides an intelligent commodity recommendation processing method based on image recognition, which is applicable to the commodity recommendation server. The method steps defined by the flow related to the intelligent commodity recommendation processing method based on image recognition can be realized by the commodity recommendation server. The specific process shown in FIG. 2 will be described in detail below.

Step S110, when a commodity recommendation request of a target user terminal device is received, the commodity recommendation request is analyzed, and a corresponding commodity request description image sequence is output.

In the embodiment of the present invention, the product recommendation server may analyze the product recommendation request when receiving the product recommendation request from the target user terminal device, so as to output a corresponding product request description image sequence.

Step S120, for each candidate commodity image sequence in the configured multiple candidate commodity image sequences, calculating the image sequence matching degree of the candidate commodity image sequence and the commodity request description image sequence by using an image matching neural network formed by training so as to output the image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence.

In this embodiment of the present invention, the product recommendation server may perform, for each configured candidate product image sequence in the multiple candidate product image sequences, calculation of an image sequence matching degree on the candidate product image sequence and the product request description image sequence by using an image matching neural network formed by training, so as to output an image sequence matching degree between the candidate product image sequence and the product request description image sequence.

Step S130, according to an image sequence matching degree between each candidate product image sequence in the multiple candidate product image sequences and the product request description image sequence, screening at least one target product image sequence from the multiple candidate product image sequences, and then sending target product description information of at least one to-be-recommended product corresponding to the at least one target product image sequence to the target user terminal device.

In the embodiment of the present invention, the product recommendation server may screen at least one target product image sequence from the multiple candidate product image sequences according to an image sequence matching degree between each candidate product image sequence in the multiple candidate product image sequences and the product request description image sequence, and then send target product description information of at least one to-be-recommended product corresponding to the at least one target product image sequence to the target user terminal device (to implement corresponding product recommendation). The target commodity description information at least comprises a corresponding target commodity image sequence, and the target commodity image sequence is formed by carrying out image acquisition on a corresponding commodity to be recommended.

Through the content, the commodity search based on the image can be realized, wherein the commodity information generally contained in the image is rich, so that the reliability of the commodity search recommendation based on the image can be higher, the reliability of the commodity recommendation is improved to a certain extent, and the problem of low reliability in the prior art (due to insufficient information richness described in the text search) is solved.

It can be further explained that, in some possible embodiments, to facilitate understanding of step S110 in the above description, the following specific contents may be further included:

when a commodity recommendation request of a target user terminal device is received, analyzing the commodity recommendation request to determine each frame of commodity request description image contained in the commodity recommendation request and image sequencing number information corresponding to each frame of commodity request description image (the image sequencing number information can be timestamp information of the commodity request description image, namely acquisition time and the like);

and sequencing and combining each frame of commodity request description image contained in the commodity recommendation request according to the image sequencing number information corresponding to each frame of commodity request description image to form a commodity request description image sequence corresponding to the commodity recommendation request.

It may be further explained that, in some possible embodiments, the step S110 may further include a step of performing image similarity calculation processing on any two frames of product request description images, so as to perform deduplication screening on product request description images with a greater similarity based on the image similarity, where the step of performing image similarity calculation processing may further include the following specific contents:

performing feature point extraction processing (which may be any feature point extraction manner in the prior art) on a first frame of commodity request description image to form a first image feature point set corresponding to the first frame of commodity request description information, and then performing feature point extraction processing on a second frame of commodity request description image to form a second image feature point set corresponding to the second frame of commodity request description information;

traversing each first image feature point included in the first image feature point set in sequence, screening a plurality of relevant first image feature points (for example, a pixel position distance is smaller than a first threshold value, and a difference value between pixel values is smaller than a second threshold value) corresponding to the currently traversed first image feature point from the first image feature point set according to a corresponding pixel position relationship and a pixel value difference relationship for the currently traversed first image feature point, wherein the plurality of relevant first image feature points include corresponding first image feature points currently traversed, respectively screening a target first image feature point corresponding to the currently traversed first image feature point from the plurality of relevant first image feature points for a pixel value difference relationship between each relevant first image feature point and each other first pixel point in an adjacent region (for example, a circular region with a specified size taking the relevant first image feature point as a central point), and the target first image feature point corresponding to the currently traversed first image feature point has a first image feature point average value of a corresponding pixel value (i.e., a first image feature point difference value) between each other first pixel point in the relevant first image feature point and the adjacent region, and the target first image feature point has a first image feature point average value of the corresponding to the other first image feature point;

traversing each second image feature point included in the second image feature point set in sequence, screening a plurality of related second image feature points corresponding to the currently traversed second image feature point from the second image feature point set according to a corresponding pixel position relationship and a pixel value difference relationship, wherein the plurality of related second image feature points include the corresponding currently traversed second image feature point, respectively screening a target second image feature point corresponding to the currently traversed second image feature point from the plurality of related second image feature points for the pixel value difference relationship between each related second image feature point and each other second pixel point in an adjacent area, and the target second image feature point is a related second image feature point in the plurality of related second image feature points, wherein the average value of the corresponding relationship values of the pixel value difference relationship between each other second pixel point in the adjacent area and the currently traversed second image feature point has a maximum value (the same step);

counting the number of times that the first image feature point is determined as a target first image feature point for each first image feature point included in the first image feature point set, and then according to the counted number of times, performing determination processing on an importance coefficient for the first image feature point to output a first importance coefficient corresponding to the first image feature point (for example, the first importance coefficient may have a positive correlation with the number of times), and according to a pixel correlation between each other first image feature point included in the first image feature point set and the first image feature point, constructing a first pixel correlation network corresponding to the first image feature point (for example, in the first pixel correlation network, the position coordinate of each other first image feature point is the coordinate difference between the position coordinate of the other first image feature point in the first frame commodity request description image and the position coordinate of the first image feature point in the first frame commodity request description image);

counting the number of times that the second image feature point is determined as the target second image feature point for each second image feature point included in the second image feature point set, and then according to the counted number of times, performing determination processing on an importance coefficient for the second image feature point to output a second importance coefficient corresponding to the second image feature point (for example, the second importance coefficient may have a positive correlation with the number of times), and constructing a second pixel correlation relationship network corresponding to the second image feature point according to a pixel correlation relationship between each other second image feature point included in the second image feature point set and the second image feature point (for example, in the second pixel correlation relationship network, the position coordinate of each other second image feature point is the coordinate difference between the position coordinate of the other second image feature point in the second frame commodity request description image and the position coordinate of the second image feature point in the second frame commodity request description image);

the method includes performing network similarity calculation processing on a first pixel correlation network corresponding to each first image feature point and a second pixel correlation network corresponding to each second image feature point respectively (for example, the first pixel correlation network and the second pixel correlation network may be traversed respectively to form a corresponding first traversal path set and a corresponding second traversal path set, then calculating path similarity between a first traversal path and a second traversal path between the first traversal path set and the second traversal path set, that is, trajectory similarity, and then merging path similarity between each first traversal path and each second traversal path to obtain corresponding network similarity, which may be mean or weighted mean calculation, and the like), then performing fusion processing on the network similarity between the first pixel correlation network corresponding to each first image feature point and the second pixel correlation network corresponding to each second image feature point according to the corresponding first importance coefficient and second importance coefficient (for example, the corresponding first importance coefficient and second importance coefficient may be used as a corresponding average of the commodity image, and the commodity image is calculated by using a corresponding first weighted mean or a corresponding weighted mean of the commodity image, and the commodity image is calculated by using a corresponding first importance coefficient and a corresponding second image as a weighted mean value, and the commodity image is calculated by the commodity image request frame.

It can be further explained that, in some possible embodiments, to facilitate understanding of step S120 in the above description, it may further include the following specific contents:

for each candidate commodity image sequence in the configured multiple candidate commodity image sequences, performing data fusion processing on the image matching degree between each frame of candidate commodity image included in the candidate commodity image sequence and each frame of the commodity request description image (for example, a huge image matching degree therein may be marked as a corresponding image sequence matching degree) to output the image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence.

It may be further explained that, in some possible embodiments, to facilitate understanding of the step of performing, by using a pre-trained image matching neural network, an image matching degree calculation on each candidate product image included in the candidate product image sequence and each product request description image included in the product request description image sequence to output an image matching degree between each candidate product image and each product request description image in the above description for each candidate product image sequence of the configured plurality of candidate product image sequences, the method may further include the following specific contents:

respectively performing pixel point identification processing on the candidate commodity image and the commodity request description image based on a pre-configured important pixel point identification algorithm (for example, the existing identification technology about image feature points can be referred to, for example, difference comparison is performed between a pixel point and an adjacent pixel point, so that a pixel point with a larger difference with the adjacent pixel point is marked as an important pixel point), so as to identify a candidate important pixel point ordered set corresponding to the candidate commodity image and a request important pixel point ordered set corresponding to the commodity request description image from the candidate commodity image;

loading the feature distribution of the candidate important pixel points and the feature distribution of the request important pixel points to a weight attention subnetwork included in the image matching neural network for processing, so as to output a weight feature distribution of the candidate important pixel points corresponding to the feature distribution of the candidate important pixel points and a weight feature distribution of the request important pixel points corresponding to the feature distribution of the request important pixel points (for example, a preset first feature representation parameter and a feature distribution of the candidate important pixel points or a feature distribution parameter of the request important pixel points can be input into the weight attention subnetwork, so that a weight coefficient of each dimension in the feature distribution parameter can be calculated to serve as a weight vector of the feature distribution parameter, wherein a specific calculation process of the weight coefficient can be calculated based on an attention mechanism, and a value range of the weight coefficient can be [0,1], so that each feature distribution parameter in the feature distribution of the candidate important pixel points can be calculated to obtain a weight vector correspondingly, and an average value of a plurality of weight vectors corresponding to the feature distribution of the candidate important pixel points can be calculated to obtain a weight feature distribution corresponding to the important pixel points corresponding to the weight distribution of the candidate important pixel points;

based on the image matching sub-network included in the image matching neural network, the candidate important pixel point weight feature distribution and the request important pixel point weight feature distribution are subjected to matching processing to output an image matching degree between the candidate commodity image and the commodity request description image (for example, the candidate important pixel point weight feature distribution and the request important pixel point weight feature distribution may be subjected to feature distribution similarity calculation, for example, the candidate important pixel point weight feature distribution and the request important pixel point weight feature distribution may be represented in a vector form, and thus, the image matching degree may be vector similarity).

It may be further explained that, in some possible embodiments, to facilitate understanding of the step of performing information conversion processing on the candidate product image and the product request description image respectively to output an ordered set of candidate product image blocks corresponding to the candidate product image and an ordered set of request description image blocks corresponding to the product request description image in the foregoing description, the following specific contents may be further included:

respectively carrying out identification screening processing on interference pixels (such as background area pixels in a commodity image) in the candidate commodity image and the commodity request description image so as to output a candidate commodity screening image corresponding to the candidate commodity image and a commodity request description screening image corresponding to the commodity request description image (so that only foreground area pixels are included in the candidate commodity screening image and the commodity request description image);

and performing information conversion processing (the information conversion processing may be to encode included pixel values) on the candidate commodity screening image and the commodity request description screening image respectively to form a candidate commodity image block ordered set corresponding to the candidate commodity image and a request description image block ordered set corresponding to the commodity request description image.

It may be further explained that, in some possible embodiments, to facilitate understanding of the steps of performing corresponding replacement processing on the ordered candidate product image block set and the ordered request description image block set based on the preconfigured image block correspondence in the above description to form image block replacement information corresponding to the ordered candidate product image block set and the ordered request description image block set, the following specific contents may be further included:

performing corresponding substitution processing on the image block merged ordered set based on a preconfigured image block correspondence relationship to output image block corresponding substitution data corresponding to each image block included in the image block merged ordered set (for example, a preconfigured identification number or symbol may be used to substitute an image block to form image block corresponding substitution data);

respectively determining alternative data corresponding to the set position of each image block (for example, a preset identification number or symbol may be used to replace the set position to form set position corresponding alternative data) based on the image block set position corresponding to each image block included in the image block merged ordered set;

and carrying out merging operation (such as adding numbers or converting symbols into binary data and then adding the binary data) on the alternative data corresponding to the image blocks and the alternative data corresponding to the set positions, which are included in the image block merging ordered set, so as to output the image block alternative information corresponding to the image block merging ordered set.

It may be further explained that, in some possible embodiments, to facilitate understanding of the step of loading the image block replacement information into a pre-trained image matching neural network for performing a feature mining process to output an image block feature distribution corresponding to the image block replacement information in the above description, it may further include the following specific contents:

loading a first feature representation parameter preset in a weight attention sub-network and each corresponding substitute data in the image block substitute information, which are included in an image matching neural network formed by pre-training, into the weight attention sub-network for processing so as to output a weight coefficient of each corresponding substitute data;

comparing the execution number of the repeated feedback processing to determine whether the execution number of the repeated feedback processing meets the execution number requirement configured in advance (if the execution number is more than the preset number);

As an exemplary embodiment, the following steps may be included:

the first feature representation parameter (i.e., default value) pre-configured in the weight attention subnetwork and each corresponding substitute data in the image block substitute information may be loaded into the weight attention subnetwork, so as to perform self-attention calculation, so as to output a weight coefficient of each corresponding substitute data, where a value range of each weight coefficient is [0,1], so that the weight coefficient of each corresponding substitute data in the image block substitute information may be obtained, then, the corresponding weighted image block substitute information is obtained by performing weighted summation on the corresponding image block substitute information according to the weight coefficient, the weighted image block substitute information and the first feature representation parameter are used as input parameters of a recursive subnetwork, so as to perform calculation to obtain a corresponding repeated feedback processing result, then, the repeated feedback processing result is used as a new first feature representation parameter, and the above steps are executed in a loop until an execution number requirement is met, and then the execution number of the repeated feedback processing is stopped, for example, the execution number of the repeated feedback processing may be greater than or equal to the number of corresponding substitute data included in the image block substitute information, and the like. The image block replacement information is obtained by performing the process of the image block feature distribution, wherein one weight coefficient of the image block replacement information can be obtained by performing the process once in each cycle, a plurality of groups of corresponding weight coefficients can be obtained by performing the process for a plurality of times, and then the groups of weight coefficients are combined to form the image block feature distribution corresponding to the image block replacement information.

It may be further explained that, in some possible embodiments, in order to facilitate understanding of the step of respectively performing feature mining processing on the candidate important pixel point ordered set and the request important pixel point ordered set based on the image block feature distribution in the above description to output the feature distribution of the candidate important pixel point corresponding to the candidate important pixel point ordered set and the feature distribution of the request important pixel point corresponding to the request important pixel point ordered set, the following specific contents may be further included:

based on the feature distribution parameters of each important pixel point in the ordered set of candidate important pixel points in the image block feature distribution (for example, an image block may include a plurality of pixel points, and the image block feature distribution corresponding to the image block includes a plurality of feature distribution parameters corresponding to the plurality of pixel points, so that the corresponding feature distribution parameters may be determined from the image block feature distribution), combining to form an ordered set of candidate feature distribution parameters corresponding to the ordered set of candidate important pixel points (the corresponding feature distribution parameters may be ordered according to the position distribution relationship of the corresponding important pixel points to form a corresponding ordered set of candidate feature distribution parameters);

based on the characteristic distribution parameters of each image block in the request important pixel ordered set in the image block characteristic distribution, combining to form a request characteristic distribution parameter ordered set (same as above) corresponding to the request important pixel ordered set;

performing fusion processing on the feature distribution parameters corresponding to each important pixel point included in the candidate feature distribution parameter ordered set according to whether the candidate feature distribution parameters belong to the same image block (for example, if the image blocks corresponding to two important pixel points belong to the same image block, summing processing may be performed on the feature distribution parameters corresponding to the two important pixel points), so as to output the feature distribution of the candidate important pixel points corresponding to the candidate important pixel point ordered set;

and performing fusion processing (the same as above) on the feature distribution parameters corresponding to each important pixel point included in the request feature distribution parameter ordered set according to whether the feature distribution parameters belong to the same image block, so as to output the feature distribution of the request important pixel point corresponding to the request important pixel point ordered set.

It can be further explained that, in some possible embodiments, to facilitate understanding of step S130 in the above description, the following specific contents may be further included:

screening each candidate commodity image sequence with the corresponding image sequence matching degree greater than or equal to a configured image sequence matching degree reference value from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence in the candidate commodity image sequences and the commodity request description image sequence to output at least one target commodity image sequence (for example, if no candidate commodity image sequence with the corresponding image sequence matching degree greater than or equal to the image sequence matching degree reference value exists, marking the candidate commodity image sequence with the corresponding image sequence matching degree having the maximum value as the target commodity image sequence);

for each target commodity image sequence in the at least one target commodity image sequence, combining to form target commodity description information corresponding to the target commodity image sequence according to the target commodity image sequence and commodity description text data of a to-be-recommended commodity corresponding to the target commodity image sequence (exemplarily, the commodity description text data may be superimposed into a commodity image included in the target commodity image sequence), and then sending the target commodity description information to the target user terminal device.

With reference to fig. 3, an embodiment of the present invention further provides an intelligent commodity recommendation processing system based on image recognition, which is applicable to the commodity recommendation server. The intelligent commodity recommendation processing system based on image recognition can comprise a recommendation request analysis module, a matching degree calculation module and a commodity recommendation module, and can also comprise other software function modules.

It may be further explained that, in some possible embodiments, the recommendation request parsing module is configured to, when receiving a product recommendation request of a target user terminal device, parse the product recommendation request to output a corresponding product request description image sequence.

It may be further explained that, in some possible embodiments, the matching degree calculation module is configured to, for each candidate product image sequence in the configured multiple candidate product image sequences, perform image sequence matching degree calculation on the candidate product image sequence and the product request description image sequence by using an image matching neural network formed by training, so as to output an image sequence matching degree between the candidate product image sequence and the product request description image sequence.

It may be further explained that, in some possible embodiments, the product recommendation module is configured to screen at least one target product image sequence from the multiple candidate product image sequences according to an image sequence matching degree between each candidate product image sequence in the multiple candidate product image sequences and the product request description image sequence, and then send target product description information of at least one to-be-recommended product corresponding to the at least one target product image sequence to the target user terminal device, where the target product description information at least includes a corresponding target product image sequence, and the target product image sequence is formed based on image acquisition on the corresponding to-be-recommended product.

Based on this, the embodiment of the present invention further provides a server, including a processor and a memory; the processor is communicatively connected to the memory, and the processor is configured to read the computer program from the memory and execute the computer program to implement the method.

In summary, the method and system for processing intelligent recommendation of commodities based on image recognition provided by the present invention analyze the commodity recommendation request to output the corresponding commodity request description image sequence. And for each candidate commodity image sequence, calculating the image sequence matching degree of the candidate commodity image sequence and the commodity request description image sequence by using an image matching neural network so as to output the corresponding image sequence matching degree. And screening at least one target commodity image sequence from the candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence and the commodity request description image sequence, and sending the target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal equipment. Through the content, the commodity search based on the image can be realized, wherein the commodity information generally contained in the image is rich, so that the reliability of the commodity search recommendation based on the image can be higher, the reliability of the commodity recommendation is improved to a certain extent, and the problem of low reliability in the prior art is solved.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The intelligent commodity recommendation processing method based on the image recognition is characterized by being applied to a commodity recommendation server and comprises the following steps:

when a commodity recommendation request of the target user terminal equipment is received, analyzing the commodity recommendation request to output a corresponding commodity request description image sequence;

2. The intelligent commodity recommendation processing method based on image recognition as claimed in claim 1, wherein the step of parsing the commodity recommendation request to output a corresponding commodity request description image sequence when receiving the commodity recommendation request of the target user terminal device comprises:

3. The intelligent commodity recommendation processing method based on image recognition according to claim 1, wherein the step of performing image sequence matching degree calculation on each of the candidate commodity image sequences and the commodity request description image sequence by using an image matching neural network formed by training to output the image sequence matching degree between the candidate commodity image sequence and the commodity request description image sequence comprises:

4. The intelligent commodity recommendation processing method based on image recognition according to claim 3, wherein the step of performing image matching degree calculation on each candidate commodity image included in the candidate commodity image sequence and each commodity request description image included in the commodity request description image sequence by using a pre-trained image matching neural network for each candidate commodity image sequence in the configured multiple candidate commodity image sequences to output the image matching degree between each candidate commodity image and each commodity request description image comprises:

5. The intelligent commodity recommendation processing method based on image recognition according to claim 1, wherein the step of performing information conversion processing on the candidate commodity image and the commodity request description image respectively to output a candidate commodity image block ordered set corresponding to the candidate commodity image and a request description image block ordered set corresponding to the commodity request description image comprises:

6. The intelligent commodity recommendation processing method based on image recognition according to claim 5, wherein the step of performing corresponding substitution processing on the ordered candidate commodity image block set and the ordered request description image block set based on a preconfigured image block correspondence relationship to form image block substitution information corresponding to the ordered candidate commodity image block set and the ordered request description image block set includes:

carrying out corresponding substitution processing on the image block merged ordered set based on a pre-configured image block corresponding relation so as to output substitution data corresponding to each image block included in the image block merged ordered set;

7. The commodity intelligent recommendation processing method based on image recognition as claimed in claim 1, wherein said loading the image block substitution information into an image matching neural network formed by pre-training for feature mining processing to output an image block feature distribution corresponding to the image block substitution information comprises:

comparing the execution number of the repeated feedback processing to determine whether the execution number of the repeated feedback processing meets the execution number requirement configured in advance;

splicing the weight coefficients output by the repeated feedback processing on the basis of each time to form image block feature distribution corresponding to the image block substitute information, assuming that the execution number of the repeated feedback processing meets the execution number requirement;

and if the execution quantity of the repeated feedback processing does not meet the execution quantity requirement, performing data fusion calculation on the weight coefficient and the first feature representation parameter to output a corresponding weight coefficient, loading the weight coefficient into a recursive sub-network included in the image matching neural network to perform repeated feedback processing to output a corresponding repeated feedback processing result, taking the repeated feedback processing result as a new first feature representation parameter, and performing a step of loading each corresponding substitute data in the image block substitute information and the first feature representation parameter preset in the weight concerned sub-network included in the image matching neural network formed by pre-training into the weight concerned sub-network to perform processing so as to output the weight coefficient of each corresponding substitute data.

8. The intelligent commodity recommendation processing method based on image recognition as claimed in claim 4, wherein said step of performing feature mining processing on said ordered set of candidate important pixels and said ordered set of requested important pixels based on said image block feature distribution, respectively, to output feature distribution of candidate important pixels corresponding to said ordered set of candidate important pixels and feature distribution of requested important pixels corresponding to said ordered set of requested important pixels, comprises:

performing fusion processing on the feature distribution parameters corresponding to each important pixel point in the candidate feature distribution parameter ordered set according to whether the candidate feature distribution parameters belong to the same image block or not so as to output the feature distribution of the candidate important pixel points corresponding to the candidate important pixel point ordered set;

and performing fusion processing on the feature distribution parameters corresponding to each important pixel point included in the request feature distribution parameter ordered set according to whether the image blocks belong to the same image block or not so as to output the feature distribution of the request important pixel points corresponding to the request important pixel point ordered set.

9. The intelligent commodity recommendation processing method based on image recognition according to any one of claims 1 to 8, wherein the step of screening out at least one target commodity image sequence from the plurality of candidate commodity image sequences according to the image sequence matching degree between each candidate commodity image sequence in the plurality of candidate commodity image sequences and the commodity request description image sequence, and then sending the target commodity description information of at least one to-be-recommended commodity corresponding to the at least one target commodity image sequence to the target user terminal device comprises:

10. The intelligent commodity recommendation processing system based on the image recognition is applied to a commodity recommendation server and comprises the following components:

11. A server, comprising a processor and a memory; the processor is in communication connection with the memory, and the processor is configured to read the computer program from the memory and execute the computer program to implement the method of any one of claims 1 to 3.