CN109145901A

CN109145901A - Item identification method, device, computer readable storage medium and computer equipment

Info

Publication number: CN109145901A
Application number: CN201810923633.8A
Authority: CN
Inventors: 刘强; 戴宇荣; 沈小勇
Original assignee: Tencent Technology Shenzhen Co Ltd; Tencent Cloud Computing Beijing Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd; Tencent Cloud Computing Beijing Co Ltd
Priority date: 2018-08-14
Filing date: 2018-08-14
Publication date: 2019-01-04
Anticipated expiration: 2038-08-14
Also published as: CN109145901B

Abstract

This application involves a kind of item identification method, device, computer readable storage medium and computer equipment, method includes: to obtain at least one target item in the target item image of article placement space；Target item image is input to image detection model；The characteristics of image distributed intelligence of image detection model output is obtained, characteristics of image distributed intelligence includes the corresponding images of items feature of each target item and each images of items feature corresponding feature locations information in target item image；Each images of items feature is matched with the standard article feature in item database, obtains corresponding matching degree；The corresponding target item type of each images of items feature is confirmed according to matching degree；According to characteristics of image distributed intelligence and target item type, article position and corresponding type of goods of each target item in article placement space are identified.This method can not only improve recognition efficiency, additionally it is possible to promote recognition accuracy.

Description

Item identification method, device, computer readable storage medium and computer equipment

Technical field

This application involves field of computer technology, more particularly to a kind of item identification method, device, computer-readable deposit Storage media and computer equipment.

Background technique

It is also more and more extensive for the application of computer technology with the development of computer technology.In the conventional technology, right Hardware plan, such as weighing sensor, RFID (radio frequency identification) and video identification are largely still relied in the identification of article Deng.But this article knows method for distinguishing, and hardware cost and operation cost are all higher.And this traditional identification technology, detection effect Rate and Detection accuracy are all lower.

Summary of the invention

Based on this, it is necessary in view of the above technical problems, provide a kind of detection efficiency and Detection accuracy of can be improved Item identification method, device, computer readable storage medium and computer equipment.

A kind of item identification method, comprising:

At least one target item is obtained in the target item image of article placement space；

The target item image is input to image detection model；

The characteristics of image distributed intelligence of described image detection model output is obtained, described image feature distribution information includes every A corresponding images of items feature of the target item and each images of items feature are right in the target item image The feature locations information answered；

Each images of items feature is matched with the standard article feature in item database, is obtained corresponding Matching degree；

The corresponding target item type of each images of items feature is confirmed according to matching degree；

According to described image feature distribution information and the target item type, identify each target item in institute State the article position and corresponding type of goods in article placement space.

A kind of item identification devices, described device include:

Image collection module, for obtaining at least one target item in the target item image of article placement space；

Detection module, for the target item image to be input to image detection model；It obtains described image and detects mould The characteristics of image distributed intelligence of type output, described image feature distribution information includes the corresponding article figure of each target item As feature and each images of items feature corresponding feature locations information in the target item image；

A matching module, for carrying out the standard article feature in each images of items feature and item database Match, obtains corresponding matching degree；The corresponding target item type of each images of items feature is confirmed according to matching degree；

Type of goods determining module, for according to described image feature distribution information and the target item type, identification Article position and corresponding type of goods of each target item in the article placement space out.

A kind of computer equipment can be run on a memory and on a processor including memory, processor and storage Computer program, the processor perform the steps of when executing the computer program

The target item image is input to image detection model；

A kind of computer readable storage medium, is stored thereon with computer program, and the computer program is held by processor It is performed the steps of when row

The target item image is input to image detection model；

Above-mentioned item identification method, device, computer readable storage medium and computer equipment pass through image detection model Target item image is extracted and distinguish wherein include article target image characteristics, then by item database to packet Target image characteristics comparison match containing article, so that it is determined that type of goods in the presence of target item image and specific Location information, this mode combine image detection model with item database, belong to article by the differentiation of image detection model Characteristics of image, i.e. it is special can to export the images of items feature for belonging to target item and each images of items image detection model Sign in target item image corresponding feature locations information, further by with the standard article feature in item database Retrieval matching is carried out, by retrieving matched mode instead of directly in such a way that image detection model is classified, substantially Degree improves recognition efficiency, and it is quasi- to promote identification in such a way that image detection model and item database are combined True rate.

Detailed description of the invention

Fig. 1 is the applied environment figure of item identification method in one embodiment；

Fig. 2 is the flow diagram of item identification method in one embodiment；

Fig. 3 is to carry out feature to target item image by the feature extraction network of image detection model in one embodiment The flow diagram of extraction step；

Fig. 4 is in one embodiment by the standard article feature progress in each images of items feature and item database The flow diagram for the step of matching, obtaining corresponding matching degree；

Fig. 5 is the flow diagram in one embodiment before obtaining target item image step；

Fig. 6 A is the interface schematic diagram of article placement space reference point in one embodiment；

Fig. 6 B is the interface schematic diagram in article placement space line of demarcation in one embodiment；

Fig. 7 is the flow diagram of the establishment step of item database in one embodiment；

Fig. 8 is the interface schematic diagram divided in one embodiment to article placement space；

Fig. 9 is the interface schematic diagram for carrying out Image Acquisition in one embodiment to standard article；

Figure 10 is the flow diagram of the training step of image detection model in one embodiment；

Figure 11 is the flow diagram in one embodiment after step 212；

Figure 12 is the flow diagram of the also included step of item identification method in one embodiment；

Figure 13 is the flow diagram of the also included step of item identification method in one embodiment；

Figure 14 is the flow diagram of item identification method in another embodiment；

Figure 15 is the interface schematic diagram of the recommendation placement location of image capture device in one embodiment；

Figure 16 A is the flow diagram being input to target item image in one embodiment in image detection model；

Figure 16 B is the schematic diagram of the second convolution module in one embodiment；

Figure 17 is the structural block diagram of item identification devices in one embodiment；

Figure 18 is the structural block diagram of computer equipment in one embodiment.

Specific embodiment

It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, and It is not used in restriction the application.

Fig. 1 is the applied environment figure of item identification method in one embodiment.Referring to Fig.1, the item identification method application In article identification system.The article identification system includes terminal 110 and server 120.It wherein include that image is adopted in terminal 110 Collect equipment 102 and processor 104.Terminal 110 can be terminal corresponding to article placement space, image capture device 102 It can be the equipment for having image collecting function, such as video camera, camera, mobile phone, tablet computer, laptop etc..Figure As acquisition equipment 102 can carry out Image Acquisition to the article being placed in article placement space.It may include in processor 104 There is image detection model, for carrying out feature extraction to the collected images of items of image capture device 102.110 kimonos of terminal Device 120 be engaged in by network connection, is can wrap in server 120 containing item database, is stored in item database each The corresponding multiple standard article features of a type of goods.Processor 104 may be mounted in terminal 110, can also be located at service In device 120, for carrying out feature extraction to the collected images of items of image capture device 102, and the object for belonging to article is exported Product characteristics of image.

As shown in Fig. 2, in one embodiment, providing a kind of item identification method.Referring to Fig. 2, the article identification side Method specifically comprises the following steps:

Step 202, at least one target item is obtained in the target item image of article placement space.

Step 204, target item image is input to image detection model.

Target item image is the image for needing to identify included type of goods and position, i.e., in target item image It can wrap containing one or more target items, target item has can in corresponding type of goods, such as target item image To include target item A, target item B, target item C, the corresponding type of goods of target item A, B, C can be beverage, Snacks etc..After getting target item image, the target item image that can be will acquire be input to image detection model into Row feature-extraction analysis.Image detection model is can to carry out feature extraction to the image of input and judge that the image extracted is special Sign whether include article model.Image detection model may include there are two network, be feature extraction network and inspection respectively Survey grid network.Feature extraction network can carry out feature extraction to the image of input, and detection network can mention feature extraction network The characteristics of image got is judged, judges which characteristics of image is the feature for including article, which is not include to have article Feature.That is detection network can filter out the characteristics of image for belonging to article.So as to the figure for belonging to article that will be filtered out As feature output, matched with the standard article feature in item database.

Step 206, the characteristics of image distributed intelligence of image detection model output is obtained, characteristics of image distributed intelligence includes every A corresponding images of items feature of target item and each images of items feature corresponding feature locations in target item image Information.

Feature extraction network in image detection model can be remembered when carrying out feature extraction to the target item image of input Location information of the feature extracted in target item image is recorded, location information can be seat of the feature in target item image Mark information.Detection network in image detection model is available to extract target item image to feature extraction network The characteristics of image arrived, and whether detection image feature belongs to the feature of article, i.e., in detection target image characteristics whether include Article, if so, then the characteristics of image is the feature for belonging to article, i.e. images of items feature；If it is not, the characteristics of image It is not belonging to the feature of article.After image detection model inspection goes out target item corresponding characteristics of image, i.e. image detection model After detecting images of items feature, corresponding characteristics of image distributed intelligence can be exported.Characteristics of image distributed intelligence includes every Images of items feature corresponding to a target item and each images of items feature corresponding feature in target item image Location information.

Step 208, each images of items feature is matched with the standard article feature in item database, is obtained pair The matching degree answered.

The corresponding standard article feature of each type of goods is stored in item database.Each type of goods can correspond to Multiple standard article features, standard article feature, which refers to, carries out the feature that feature extraction obtains, root to the article of the type of goods The article of the article of the type of goods and other type of goods can be subjected to area according to the standard article feature of each type of goods Point.Type of goods corresponds to a kind of article, for example canned cola and bottled cola respectively correspond a type of goods, it is canned can Pleasure is corresponding to can be canned laughable class, and bottled cola is corresponding to can be bottled laughable class.When there are not for canned cola With packaging or different designs when, it is also desirable to establish corresponding type of goods respectively according to for each canned cola, and extract To each canned cola feature as the corresponding standard article feature of each canned cola, be used for and other articles carry out area Point.

It, can be by article figure after the images of items feature and corresponding location information for getting the output of image detection model It is available to images of items feature and each article kind as feature is matched with the standard article feature in item database Matching degree between the corresponding standard article feature of class.Matching degree represents each images of items feature and each standard article is special Similarity degree between sign, similarity degree is higher, and to represent a possibility that the two features belong to the same type of goods bigger.

Step 210, the corresponding target item type of each images of items feature is confirmed according to matching degree.

Step 212, according to characteristics of image distributed intelligence and target item type, identify that each target item is put in article Article position and corresponding type of goods between emptying.

It, can after getting the matching degree between images of items feature standard article feature corresponding with each type of goods Using the target item type by the corresponding type of goods of the highest standard article feature of matching degree as target image characteristics.When depositing In multiple images of items features, available target species type corresponding with each images of items feature.That is, working as Originally there are when multiple target items, it is corresponding that image detection model can then export multiple target items in target item image Central Plains Images of items feature.When being matched with the standard article feature in item database, then each article can be got The matching degree of characteristics of image and each standard article feature, so as to obtain article kind corresponding to each images of items feature Class.Due to having recorded spy of each characteristics of image in target item image when carrying out feature extraction to target item image Levy location information.Therefore after the images of items feature for getting the output of image detection model, it may be determined that in target item image Article position of the type of goods and each target item for each target item for including in article placement space.It can Determination includes to belong to the target item of which type of goods, and can know that each target item exists in target item image Specific location in article placement space.

Above-mentioned item identification method, is extracted and is distinguished to target item image by image detection model and wherein include There are the target image characteristics of article, then by item database to the target image characteristics comparison match for including article, thus Determine the type of goods and specific location information in the presence of target item image, this mode is by image detection model and object Product database combines, and the characteristics of image for belonging to article is distinguished by image detection model, i.e. image detection model can export category In the images of items feature of target item and each images of items feature in target item image corresponding Q-character confidence Breath, further by carrying out retrieval matching with the standard article feature in item database, by retrieving matched mode generation It has replaced directly in such a way that image detection model is classified, recognition efficiency is greatly improved, passed through image detection mould The mode that type and item database are combined can also promote recognition accuracy.

In one embodiment, Image Acquisition is carried out by the image capture device in article placement space and obtains object Product image；Standard article in item database is characterized according to the consistent Image Acquisition of setting parameter with image capture device Equipment carries out the image that Image Acquisition obtains and carries out what feature extraction obtained.

Article placement space refers to the space that can place article, and article placement space can be cabinet, shelf, rack Deng.In article placement space, image capture device is housed, when needing to obtain the images of items in article placement space, can be led to Image capture device is crossed to be acquired to obtain.In establishing item database, it can be joined by the setting with image capture device The consistent image capture device of number collects corresponding image, establishes often so as to carry out feature extraction to acquired image Standard article feature corresponding to a type of goods.Setting parameter unanimously includes at least one below: putting height and position Unanimously, device parameter is consistent, the environmental parameter where equipment is consistent etc..It puts height and position consistency refers to image capture device Height and position are put apart from the height or the image capture device in position, with article placement space for needing the article acquired It is consistent.Device parameter unanimously refers to that the parameter of image capture device is consistent, for example, focal length, camera lens size and Screening-mode etc. is consistent about the relevant parameter of equipment.Environmental parameter where equipment is consistent, can refer to that image is adopted Environment where collection equipment is consistent, for example the counter size where image capture device installation is consistent or image The light of environment is consistent etc. where acquisition equipment.It follows that target item image and establishing item database when institute The image used is collected by the way that the consistent image capture device of parameter is arranged, therefore can be by target item image In the feature extracted be compared with the standard article feature in item database, to determine included in target item image Type of goods.

Further, the quantity of image capture device can be determined according to the size of article placement space.For example article is placed When space is a biggish counter A, then multiple images acquisition equipment can be installed in counter A, so that each area in counter A The article that domain is placed can be collected by image capture device, and the clear and energy in image capture device acquired image It is enough that corresponding feature is extracted by feature extraction network.Image capture device, which can be, various has setting for image collecting function It is standby, in the present embodiment, in order to ensure camera can capture the object of the different height and size placed in article placement space 180 ° of fish-eye camera can be used using fish-eye camera as image capture device, specifically in product.It so both can be with Guarantee to capture the region that complete article is placed, can also reduce article placement space fringe region generation it is abnormal Become, in order to avoid influence that error occurs when carrying out feature extraction to image capture device acquired image, to influence to target The identification accuracy for the type of goods for including in images of items.

In one embodiment, the characteristics of image distributed intelligence of image detection model output is obtained, comprising: examine by image The feature extraction network for surveying model carries out feature extraction to target item image, obtains corresponding images of items feature and each object Feature locations information of the product characteristics of image in target item image；The detection network of image detection model is obtained to images of items Feature is detected, and the corresponding images of items feature of each target item of output and each images of items feature are in target item Corresponding feature locations information in image.

Include there are two network in image detection model, is feature extraction network and detection network respectively.Feature extraction is used In carrying out feature extraction to the image of input, to obtain the location information of characteristics of image and characteristics of image in the picture.Detect net The characteristics of image that network is extracted for inspection feature extraction network whether belong to include article feature, i.e. detection network uses It include that the characteristics of image of article has the characteristics of image of article with not including, and exports the object for determining each target item in distinguishing Product characteristics of image and each images of items feature corresponding location information in target item image.

In one embodiment, as shown in figure 3, by the feature extraction network of image detection model to target item image Carry out feature extraction, comprising:

Step 302, convolution algorithm is carried out to target item image by the first convolution module of feature extraction network.

Step 304, the fisrt feature of the first convolution module output is normalized.

Step 306, the input pond layer progress pond effect of the fisrt feature after normalized will have been carried out, pond is obtained As a result.

Step 308, pond result is input to the second convolution module in feature extraction network and carries out cyclic convolution operation.

Step 310, the second feature for obtaining the output of the second convolution module, it is corresponding using second feature as target item image Characteristics of image.

It include the first convolution module and the second convolution module in feature extraction network.First convolution module for pair The target item image of input carries out feature extraction and obtains corresponding characteristics of image, and characteristics of image is converted into image vector, Convolution algorithm is carried out to image vector.Obtained fisrt feature after first convolution module output convolution algorithm, i.e. fisrt feature to Amount.First eigenvector is normalized and is referred to, first eigenvector is handled, so that first eigenvector is protected It holds in the regional scope of setting.The fisrt feature after normalized will be carried out and is input to pond layer to carry out pond effect. Pond layer can traverse each first eigenvector, using the maximum value of each first eigenvector as output, be input to feature It extracts and carries out cyclic convolution operation in the second convolution module of network.Cyclic convolution operation refers to through institute in the second convolution module The multilayer for including repeats convolution block and single laminate roll block, can carry out cyclic convolution operation to first eigenvector.

Multilayer repeats convolution block refers to include the same convolutional layer of multiple concatenated convolution nuclear parameters, can by multilayer The number of plies of convolutional layer can be increased by repeating convolution module, enhance convolution algorithm.Multilayer repeats convolution module can be more effectively sharp It is also simpler to realize more complicated convolution algorithm with system resource.Single laminate roll block is non-duplicate convolution block, single layer convolution The convolution nuclear parameter that block repeats convolution block with multilayer can have differences, in order that the different convolution numbers of plies can calculate difference The feature of fineness, and convolution nuclear parameter has differences and can also carry out edge filling to feature that convolution algorithm obtains, can be with The feature that the convolutional layer small to convolution kernel is calculated is filled, and the length of feature and width is avoided to change, thus can be with Repeat convolution algorithm operation always to this feature.

Single laminate roll block and multilayer repeat convolution block when carrying out convolution algorithm, are using the volume pre-set Product collecting image carries out feature extraction and obtains characteristics of image.Multilayer repeat convolution block due to convolution nuclear parameter it is consistent, When carrying out convolution algorithm, image size can be kept constant, and extract different types of characteristics of image.During this, May need to change the size of characteristics of image, thus can by the convolution operation of the different single laminate roll block of convolution nuclear parameter into Row is realized.Convolution algorithm is carried out to image when repeating convolution block with multilayer by the single laminate roll block in the second convolution module Afterwards, the available second feature exported to the second convolution module, using second feature as the corresponding image of target item image Feature.

In one embodiment, as shown in figure 4, it is the standard article in each images of items feature and item database is special Sign is matched, and corresponding matching degree is obtained, comprising:

Step 402, the corresponding multiple standard article features of each type of goods in item database are obtained.

Step 404, the Europe between images of items feature multiple standard article features corresponding with each type of goods is obtained Formula distance value.

Step 406, according to Euclidean distance value determine the standard article feature in images of items feature and item database it Between matching degree.

It, can be by each object after getting image detection model and exporting the corresponding images of items feature of each target item Product characteristics of image is matched with the standard article feature got from item database respectively.It can be according to images of items spy It levies the Euclidean distance value between corresponding with each type of goods multiple standard article features and determines target image characteristics and each Matching degree between standard article feature.Euclidean distance is the distance definition generallyd use, refers to two points in an n-dimensional space Between actual distance or vector natural length, i.e. the distance of the point to origin.Euclidean in two and three dimensions space Distance is exactly the actual range between two o'clock.When the Euclidean distance of two features is smaller, i.e. Euclidean distance between two features When being worth smaller, then the similitude between the two features is bigger, and matching degree is also higher.If conversely, the Europe between two features Family name's distance value is larger, then the differentiation between two features is also larger, and matching degree is lower.

The target image characteristics of image detection model output can have it is multiple, and in item database, each type of goods Multiple standard article features are corresponding with, therefore when comparing, each images of items feature will be corresponding with each type of goods Standard article feature compared one by one.So as to obtain each images of items feature mark corresponding with each type of goods Euclidean distance value between quasi- article characteristics, it can obtain the matching between each images of items feature and each type of goods Degree.

In one embodiment, the corresponding target item type of each images of items feature is confirmed according to matching degree, comprising: By the Euclidean distance value of images of items feature corresponding with each type of goods standard article feature respectively according to from small to large Sequence is ranked up to obtain ranking results；The standard article of preset quantity is obtained from ranking results according to direction from small to large Feature；The quantity for belonging to the standard article feature of the same type of goods in the standard article feature of preset quantity is counted, according to The population size of the corresponding standard article feature of each type of goods determines the corresponding target item type of images of items feature.

It, can when by each images of items feature, standard article feature corresponding with each type of goods is compared respectively Get the Euclidean distance value between multiple standard article features and each images of items feature.That is, being directed to each object For product characteristics of image, the standard article characteristic sequence being ranked up according to Euclidean distance value can be obtained.It can be according to Each standard article feature is ranked up by the sequence from small to large of Euclidean distance value, i.e., will be between images of items feature The smallest standard article feature of Euclidean distance value comes the front end of sequence.So standard article can be obtained according to Euclidean distance value The ranking results of feature.It, can be in corresponding standard article characteristic sequence, according to Europe for each images of items feature The direction of formula distance value from small to large selects the standard article feature of preset quantity.For example preset quantity is set as 15, then It can then select and the lesser 15 standard article features of the Euclidean distance value of images of items feature.

The number for belonging to the standard article feature of the same type of goods in the standard article feature of preset quantity can be counted Amount, so as to determine the corresponding mesh of images of items feature according to the population size of the corresponding standard article feature of each type of goods Mark type of goods.Such as target image characteristics A, the number of type of goods X1 is belonged in 15 standard article features selecting Amount is 8, and the quantity for belonging to type of goods X2 is 5, and the quantity for belonging to type of goods X3 is 2, is being chosen then would know that In the standard article feature of preset quantity out, the quantity of the corresponding standard article feature of type of goods X1 is most, then can be true The corresponding target item type of the characteristics of image A that sets the goal is X1.It can determine that each target image characteristics institute is right in this way The type of goods answered.

In one embodiment, as shown in figure 5, further including determining article placement space before obtaining target item image Line of demarcation the step of, comprising:

Step 502, the corresponding reference point of article placement space where image capture device is obtained.

Step 504, the corresponding spatial image of article placement space of image capture device acquisition is obtained.

Step 506, spatial image is pre-processed.

Step 508, the boundary point of pretreated spatial image is obtained according to reference point.

Step 510, the line of demarcation of article placement space is determined according to boundary point.

In order to reduce the influence that the environment outside article placement space identifies article, point of article placement space can be determined Boundary line.Specifically, technical staff can analyze article placement space and determine reference point.As shown in FIG, exist There are two small square Q1 and Q2, the as determining reference point of technical staff in the upper left corner and the upper right corner of figure.It is available to arrive The corresponding spatial image of article placement space of image capture device acquisition can be in order to improve the determination accuracy rate in line of demarcation The spatial image got is pre-processed.Pretreatment refers to the tune that the parameters such as brightness, color value or gray scale are carried out to image Section.Pretreatment can be binary conversion treatment, refer to after the image that will acquire carries out gray processing processing, i.e., by colored space After image is converted to gray level image, gray level image is subjected to binary conversion treatment, that is, by the gray value of the point on gray level image Processing is 0 or 255.

After pre-processing to spatial image, the boundary point of pretreated spatial image can be obtained according to reference point. For example it can determine the boundary point of spatial image according to the reference point in pretreated spatial image.It is by two boundary point connections The line of demarcation of article placement space can be obtained.As shown in Figure 6B, the Q1 grey horizontal line that connects with Q2 in top is in image According to the line of demarcation for the article placement space that two boundary points obtain.Therefore when carrying out article identification, it can be used and detect Line of demarcation, remove influence of the environment to article recognition accuracy outside article placement space.

Target item image is input to image detection model, comprising: determine in target item image and belong to according to line of demarcation Target image is input to image detection model by the target image in article placement space.

When target item image to be input in image detection model, object can be determined according to determining line of demarcation Belong to the image-region in article placement space in product image, and belongs to the image-region outside article placement space.And it can be with The image-region outside line of demarcation is removed according to line of demarcation, that is, belongs to the image-region outside article placement space, and article will be belonged to Target image in placement space is input to the operations such as the extraction that characteristics of image is carried out in image detection model and analysis.

In one embodiment, as shown in fig. 7, the establishment step of item database, comprising:

Step 702, standard article image is obtained.

It include the corresponding multiple standard article features of each type of goods in item database.Establishing item database When, need to get multiple standard article images corresponding with each type of goods.

In one embodiment, standard article image is obtained, comprising: obtain according to the type of goods of standard article to standard The position of article and angle carry out the standard article image that Image Acquisition obtains after being adjusted.

It, can be according to the type of goods of standard article, to mark when image capture device carries out Image Acquisition to standard article The position of quasi- article and angle are adjusted, and Image Acquisition are carried out to the standard article of different angle and position, can obtain Different standard article images, so that each feature of each standard article can clearly be shown on standard article image, from And the feature of standard article can be extracted to.

Step 704, image procossing is carried out to standard article image, the quantity of images of items is extended to obtain reference substance Product image collection.

It can be by images of items after carrying out image procossing to the standard article image got for each type of goods Quantity be extended.Image procossing can be data enhancing processing, for example gamma correction processing (Gamma correction), the spiced salt are made an uproar Sonication etc..Gamma correction processing, which refers to, edits the gamma curve of image, to carry out non-linear tone editor to image Method, the dark parts and light-colored part in picture signal are detected, and increase ratio, to improve picture contrast The processing method of effect.Salt-pepper noise is a kind of white point or stain occurred at random, it may be possible to which there is black picture element in bright region Or have a white pixel in dark region, or both.Salt-pepper noise can be removed using median filter.

Step 706, feature extraction is carried out to each image in standard article image collection.

Step 708, duplicate removal screening is carried out to the feature that extraction obtains.

Step 710, the feature after screening is deposited according to the corresponding type of goods of standard item each in standard article image Storage is into item database.

In collected standard article image, data are more single, for example standard article image is all in same ring It is collected under border.Picture is carried out upper and lower, left and right by the picture that different brightness can be generated after handling by gamma correction The operation such as overturning, single image can be extended.Salt-pepper noise processing etc., which is added, can simulate more actual scenes In the variation that is likely to occur, in this way, the quantity of collected standard article image can be carried out at double Amplification, can be obtained standard article image collection.Each standard article image in standard article image collection is input to feature It extracts in network and carries out feature extraction, characteristics of image corresponding with each standard article image can be obtained.Due to each article kind The corresponding standard article image of class have it is multiple, therefore to the characteristics of image that the standard article image zooming-out under each type of goods arrives In, there are the higher characteristics of image of similarity for meeting.In order to reduce unnecessary characteristics of image, the redundancy of data is reduced, it can To carry out duplicate removal screening operation to the image extracted.

When carrying out duplicate removal screening operation, can be screened according to the Euclidean distance value between two features.When two When Euclidean distance value between characteristics of image is very close, it may be considered that the two characteristics of image are duplicate.It can also set A fixed Euclidean distance threshold value can when Euclidean distance threshold value of the Euclidean distance value between two characteristics of image lower than setting To think that the two characteristics of image are duplicate.In this way, one of them in the two characteristics of image can be deleted, retain wherein One characteristics of image.Characteristics of image after duplicate removal screening then represents the feature of each type of goods, can be according to each object The characteristics of image of kind class distinguishes the type of goods and other type of goods.It can will be after screening according to type of goods Characteristic storage is into item database.Assuming that being directed to type of goods A, 50 standard article images are collected, this 50 standards The feature that images of items is extracted is 2000, and after duplicate removal is screened, remaining 1000 features then can be by institute after screening 1000 obtained features are saved as the feature of type of goods A into item database.

In one embodiment, standard article is placed on the first cargo path region of article placement space.

When article placement space is larger, article placement space can be subjected to the processing of layering and zoning domain.Such as by conduct The counter of article placement space is divided into 4 layers, all configures an image capture device for each layer.It further, can also be every One layer of division cargo path, it can each layer region is divided into multiple row, so can be convenient and put standard article.As shown in figure 8, Article placement space is divided into 5 column in figure, when putting standard article, can according to ready-portioned region to standard article into Row is put, and what can so be put standard article is more neat, then is shown in the collected standard article image of image capture device The feature of the standard article shown is also obvious.The standard article for needing to acquire image can be placed on article placement space First cargo path region.

Acquisition carries out image after being adjusted according to the type of goods of standard article to the position of standard article and angle and adopts Collect obtained standard article image, comprising:

Mode one: when the height of standard article is lower than article height threshold value, in the second cargo path area of article placement space Domain placing height is higher than the article of article height threshold value, is adjusted according to the first preset rules to standard article, and carry out figure As collecting corresponding first standard article image collection.

Article height threshold value is pre-set height value, when the height of article is more than the article height threshold value of setting, It is on the contrary then think that the article is short article it is believed that the article is high article.When the standard article for needing to acquire image When height is lower than article height threshold value, that is, when the standard article for needing to acquire image is short article, it can be placed in article empty Between the second cargo path region place high article.For example article placement space is divided into 6 column, standard article can be placed on First cargo path region of article placement space, such as 2-5 arrange, and are higher than article height threshold value in the 1st, the 6th column placing height Other articles.

It, can be according to the first preset rules to standard article in order to which the article pattern to standard article is completely acquired It is adjusted.First preset rules, which refer to, successively rotates standard article according to the first predetermined angle, and removes the first cargo path region In a sub-regions in article.Tentative standard article is placed on the first cargo path region of article placement space, such as 2-5 Column, then the direction of the standard article of changeable 2-4 column, and standard article is rotated according to the first predetermined angle.First is default Angle is preset angle value, can be set to 30-60 degree.

As shown in figure 9, standard article is rotated 30-60 degree, figure can be used every time after changing the direction of standard article Corresponding multiple standard article images are obtained as acquisition equipment carries out Image Acquisition to standard article adjusted.It removes respectively again The article in a sub-regions in first cargo path region, for example the reference substance of the first row in the first cargo path region is removed respectively Product carry out Image Acquisition obtain several standard article images, removal the first cargo path region in second row standard article carry out Image Acquisition obtain several standard article images, removal the first cargo path region in third row standard article carry out image adopt Collection obtains several standard article images etc..To which corresponding first standard article image collection can be obtained, in the first standard It include to carry out multiple first standard article images that Image Acquisition obtains according to the first preset rules in images of items set.

Mode two: when standard article corresponds to multiple type of goods, standard article is adjusted according to the second preset rules It is whole, and carry out Image Acquisition and obtain corresponding second standard article image collection.

When the standard article for needing to acquire image has corresponded to multiple type of goods, that is, need to acquire the standard article of image More than one, and when having corresponded to different type of goods, multiple standard articles can be adjusted according to the second preset rules And acquire corresponding image data.Second preset rules, which refer to, to be rotated standard article according to the second predetermined angle and carries out corresponding Image Acquisition operation.The multiple standard articles for needing to acquire image can be placed in the 2-4 row in article placement space, and The standard article of placement is rotated into the second predetermined angle and carries out Image Acquisition, the second predetermined angle can be set to the 0- that turns left 30 degree and the 0-30 degree that turns right.To which corresponding second standard article image collection can be obtained, in the second standard article figure Image set includes to carry out multiple second standard article images that Image Acquisition obtains according to the second preset rules in closing.

Mode three: when standard article is the article of default type of goods, according to third preset rules to standard article into Row adjustment, and carry out Image Acquisition and obtain corresponding third standard article image collection.

It, can be default according to third when the corresponding type of goods of the standard article for needing to acquire image is default type of goods Rule is adjusted the standard article put, to carry out Image Acquisition operation.Third preset rules refer to standard article It puts in a particular area, and successively removes the article in the sub-regions in the specific region.For example it can be put in article 2-4 column between emptying are well placed standard article, successively take the article that first row, second row, third are arranged away and carry out Image Acquisition, To obtain corresponding third standard article image collection.It include default according to third in third standard article image collection Rule carries out multiple third standard article images that Image Acquisition obtains.

Above first preset rules, the second preset rules, third preset rules are in different situations to mark respectively Quasi- article carries out Image Acquisition.In order to allow image capture device that can photograph each article put, reference substance can put When product, according to the disposing way of U-shaped, image capture device can be made to can capture more article details and article in this way Information, when acquiring images of items, it is possible to reduce the case where article overlaps, but enable to each object as far as possible Product are shown completely in the acquired images.In this way, when carrying out feature extraction to image capture device acquired image, then The case where there are the overlappings of two or more article characteristics in the feature extracted can be reduced.When feature overlapping occurs, then can The case where both having included the feature of article A in appearance feature X, and also having included article B.And the disposing way of U row is taken then can The generation for avoiding such case, due to by the laying for goods of "high" in two sides, and by the laying for goods of " short " "high" article Centre so that U-shaped is presented in the article put.In this way, enabling to the article of " short " can't be by "high" when acquisition image Article shelter from, will not cause in the article of acquisition " short ", can be sheltered from by the article of "high" so that extraction " short " Article article characteristics when, occur part objects feature by the article of "high" article characteristics cover the case where.Therefore, it takes It more completely includes the whole features for needing the article acquired that the disposing way of U-shaped, which enables to the images of items of acquisition, from And when carrying out feature extraction to images of items, the article characteristics of article whole can be extracted, and extract Feature more completely and accurately, then can store the article characteristics extracted to item database, and it is corresponding to establish the article Standard article feature set.It is each in target item image to which when carrying out article identification, image detection model be exported A target item is matched include in article placement space to determine each with the standard article feature in item database When the type of target item and position, the recognition accuracy of article also can be improved.

I.e. by higher laying for goods in two sides, by shorter laying for goods in centre.Image is carried out to standard article to adopt Image Acquisition is carried out to standard article when the mode of collection is applied not only to establish item database, is having new article online, is needing to put It is placed on when being sold or shown on article placement space, three kinds of above-mentioned rules can be taken to carry out image to new article and adopted Collection carries out feature extraction so as to the image to new article and saves into item database, so can be by new article Feature is stored, to the identification of new article after being also convenient for.

In one embodiment, as shown in Figure 10, the training step of image detection model, comprising:

Step 1002, Image Acquisition is carried out to article sample, obtains article sample image.

Step 1004, feature extraction is carried out to article sample image, obtains multiple sample image features.

Step 1006, the image-region and corresponding markup information divided to article sample image is obtained.

Article sample can be located in each article in article placement space, can be by article placement space Image capture device to article sample carry out Image Acquisition, be also possible to by other image collecting devices to article sample into Row Image Acquisition.After obtaining the corresponding article sample image of article sample, it can be mentioned by the feature in image detection model It takes network to carry out feature extraction to article sample image, obtains the corresponding multiple sample image features of each article sample image. On the other hand, image division manually can be carried out to article sample image, obtains multiple figures corresponding with each article sample image Corresponding markup information is added as region, and to image-region.Markup information refers to that marking the image-region is to belong to article area Domain still falls within non-article region.It can be according to the image-region divided to article sample image got and every The corresponding markup information of a image-region, determines the sample image feature for belonging to article in sample image feature.

Step 1008, the sample image feature for belonging to article is determined according to image-region and corresponding markup information.

In one embodiment, determine that the sample image for belonging to article is special according to image-region and corresponding markup information Sign, comprising: obtain the degree of overlapping of sample image feature and image-region；When degree of overlapping reaches anti-eclipse threshold, then according to image The corresponding markup information in region determines the sample image feature for belonging to article；When the corresponding markup information of image-region shows to belong to When article, then sample image feature belongs to the sample image feature of article；When the corresponding markup information of image-region shows not belong to When article, then sample image feature is not belonging to the sample image feature of article.

It can be compared respectively according to sample image feature with obtained image-region is divided, when there are sample image spies When sign and the degree of overlapping for dividing between obtained image-region reach preset anti-eclipse threshold, then the figure that can be obtained according to division As the markup information in region determines whether sample image feature belongs to the sample image feature of article.When the image district that division obtains When the mark in domain indicates that this image-region is the image-region for belonging to article, then sample image feature belongs to the sample image of article Feature；When the corresponding markup information of image-region shows to be not belonging to article, then sample image feature is not belonging to the sample of article Characteristics of image.

Step 1010, label information is added to sample image feature according to the information of article.

When the sample image feature for belonging to article has been determined, it is equivalent to and also determines the sample image for belonging to non-article spy Sign so can add label information to both sample image features.Such as to belong to the sample image feature of article addition mark Label 1 add label 0 to belong to the sample image feature of non-article.So it is according to the label information of each sample image feature Would know that whether be the characteristics of image for belonging to article.

Step 1012, the weight of image detection model is adjusted, article is belonged to according to weight adjusted output Sample image feature and corresponding location information.

Step 1014, the predictablity rate of image detection model is obtained according to the label information of sample image feature.

Step 1016, when predictablity rate reaches accuracy rate threshold value, trained image detection model is obtained.

It, can be defeated by the sample image feature for being added to label information after being added to label information for each sample image feature Enter in the detection network into image detection model, detection network is trained.Training process is the adjustment process of weight, It can also be had differences according to the sample image feature for belonging to article that weight adjusted exports.The exportable sample for belonging to article The label information of this characteristics of image determines the predictablity rate of image detection model.Assuming that 0.9 is set by accuracy rate threshold value, when 100 sample image features are inputted into image detection model, output 50 sample image features for belonging to article, and this 50 In a sample image feature, the feature quantity that label is 1 is 30, i.e., label information shows the characteristics of image quantity for belonging to article It is 30, then predictablity rate is 30/50=0.6.Therefore, the predictablity rate 0.6 of image detection model is not up to accuracy rate Threshold value 0.9, then training does not complete, and needs the weight to image detection model to be adjusted, until the prediction of image detection model Accuracy rate reaches accuracy rate threshold value, and trained image detection model just can be obtained.

When being trained to image detection model, it is equivalent to the feature extraction network and detection net to image detection model Network is trained.It therefore had both included being adjusted to the weight of feature extraction network when adjusting the weight of image detection model, So that the ability in feature extraction of feature extraction network enhances, it also include being adjusted to the weight of detection network, so that detection net Network also further enhances the separating capacity of characteristics of image.

In one embodiment, as shown in figure 11, after step 212, further includes:

Step 1102, when identifying mistake to the type of goods in target item image, the article kind of identification mistake is obtained Class object region and the corresponding correct type of goods markup information of object region in target item image.

Target item image is input in image detection model after obtaining images of items feature and corresponding location information, Further images of items feature can be matched with the standard article feature in item database, so that it is determined that target item figure The target item type for including as in and corresponding location information.When to the type of goods identification mistake in target item image When, the standard article feature in item database can be updated according to the situation of identification mistake.

Specifically, available to target item image corresponding when occurring to identify wrong, and get identification mistake Type of goods in target item image object region and the corresponding correct type of goods mark letter of object region Breath.Identifying the type of goods of mistake, object region and the type of goods markup information determined can be by target item image Artificial progress standard, such as auditor are labeled region of the type of goods of identification mistake in target species image, And indicate correct type of goods corresponding to the type of goods of the identification mistake.

Step 1104, the feature extraction network that object region is input in image detection model feature is carried out to mention It takes.

Step 1106, the target item feature that feature extraction network extracts object region is obtained.

Step 1108, target item feature is updated in item database according to correct type of goods markup information corresponding Type of goods in.

Get identification mistake the corresponding object region of type of goods and correct type of goods mark letter After breath, the feature extraction network that object region can be input in image detection model carries out feature extraction, so as to The target item feature extracted in the object region is obtained, the target item feature extracted can be saved to article number According to type of goods corresponding with correct type of goods markup information in library.Such as the corresponding correct type of goods of object region Markup information is type of goods A1, then can save the target item feature that object region extracts to item database The corresponding feature of type of goods A1 in, the update to the standard article feature in item database can be fast implemented, also into One step improves the efficiency of identification.

In one embodiment, target item image is to carry out image by the image capture device in article placement space It collects.As shown in figure 12, the above method further include:

Step 1202, comparison images of items is obtained, comparison images of items is become to the article in article placement space Carry out what Image Acquisition obtained by image capture device more afterwards.

It compares images of items and target item image is to carry out figure by the image capture device in article placement space As collecting.Images of items is compared unlike target item image, comparison images of items is to article placement space What interior article was obtained after changing by image capture device progress Image Acquisition.It can be placed at the beginning by article empty Interior image capture device carries out first time Image Acquisition, obtains target item image.Article in article placement space After being changed, second of Image Acquisition can be carried out by the image capture device in article placement space, obtain comparison Product image.Article change there may be the case where be that the position of the article in article placement space is changed, and is counted Any change does not occur in amount.It is also likely to be that number of articles in article placement space is changed, for example has lacked 1 object Product or multiple articles.

Step 1204, comparison images of items is input to image detection model, until obtaining including in comparison images of items Comparison article and compare article position of the article in article placement space.

According to the recognition methods to the type of goods for including in target item image, same method can be used to comparison Images of items is identified, so as to identify the type of goods of comparison article and comparison included in comparison images of items Article corresponding location information in article placement space.Comparison images of items can be input in image detection model, be schemed As detection model can will compare images of items in belong to article contrast images feature and corresponding location information output, will Belong to the contrast images feature of article to be matched with the standard article feature in item database, may thereby determine that comparison The location information of the type of goods for the comparison article for including in product image and each comparison article in article placement space.

Step 1206, according to target item and target item in the article position of article placement space and comparison article and right Than article position of the article in article placement space, the article changed in article placement space is determined.

The type of goods for the target item for having included in target item image has been determined and each target item are put in article The type of goods and each comparison for the comparison article for including in corresponding location information between emptying, and comparison images of items Product after corresponding location information, that is, can determine whether there is article change in article placement space in article placement space, and The type of goods of the article of change and the quantity of change article.Such as comparison images of items in compared to target item image and Speech, has lacked 2 article A, then illustrates that the article changed in article placement space is article A, and quantity is 2.This identification side Formula remains on even if large-scale change has occurred in the article position in article placement space and can recognize that article places sky The interior type of goods changed and corresponding quantity realize accurately identification operation.

In one embodiment, as shown in figure 13, above-mentioned item identification method further include:

Step 1302, the article acquisition request that terminal is sent is obtained, carries user identifier in article acquisition request.

Step 1304, corresponding pay invoice is generated according to the article changed in user identifier, article placement space.

Step 1306, pay invoice is sent to terminal, completes corresponding delivery operation for terminal.

The available article acquisition request sent to terminal of article placement space, to be obtained according to the article received Request opens article placement space to user, and user is allowed freely to obtain the article in article placement space.? Corresponding user identifier is carried in the article acquisition request that terminal is sent, the article changed in article placement space is being determined Afterwards, corresponding pay invoice can be generated according to the article changed in user identifier, article placement space.For example article places sky The article of interior change is cola, quantity 2, then can be determining corresponding with the user identifier according to two bottles of colas of change Pay invoice, then the corresponding account of user identifier then needs to complete corresponding resource numerical value transfer, i.e. terminal according to pay invoice Need to complete corresponding delivery operation.Further, when article placement space gets the article acquisition request of terminal transmission, It can send and exempt from close obligation authority to terminal, terminal after agreeing to that close obligation authority opening is exempted from opening, just may be used by article placement space Corresponding branch can be generated to the open permission of user, the article so as to change according to user identifier, article placement space After paying order, operation of withholing directly is carried out from the corresponding account of user identifier.Exempt from close payment to refer to without user's confirmation branch Operation of withholing is carried out from the account of user in the case where paying.

In one embodiment, a kind of item identification method is provided.Referring to Fig.1 4, which specifically includes Following steps:

Step 1402, it obtains to the corresponding article sample image of article sample, image detection model is trained.

Before by image detection model use into actual project, image detection model can be trained.It can To get a large amount of article sample, Image Acquisition is carried out to article sample and obtains the corresponding sample image of each article sample. The corresponding sample image of each article sample can have multiple, after carrying out feature extraction to article sample image, then can be obtained every Multiple sample image features corresponding to a article sample.Sample image feature can be input in image detection model and be instructed Practice, in the training process, can constantly adjust the weight of image detection model, belongs to article so as to the output of image detection model The accuracy of sample image feature is higher.Therefore in the training process, need to add label for each sample image feature, so as to In the prediction accuracy for determining image detection model.

When adding label for sample image feature, can be incited somebody to action after getting the corresponding sample image of article sample Sample image is copied, and carries out region division to sample image by technical staff, for example each sample image is divided into 16 impartial image-regions.And it is corresponded to according to the case where including article in each image-region for the addition of each image-region Markup information, such as when including article in image-region A, it is believed that image-region A is the region for belonging to article, can be figure It is 1 as region A adds markup information, and is to add markup information 0 not comprising the image-region B for having article.

In this way, being directed to each article sample image, there are the image districts with markup information that artificial division obtains Domain, there is also extract the sample image feature for carrying out obtaining after feature extraction by feature extraction.It can be by sample image feature and figure As the comparison of region progress degree of overlapping, when the degree of overlapping of some sample image feature and a certain image-region reaches anti-eclipse threshold When, it can be using the markup information of the image-region as the markup information of sample image feature.Such as sample image feature X1 and figure It, can be using the markup information of image-region B1 as the mark of sample image feature X1 as the degree of overlapping of region B1 has reached anti-eclipse threshold Infuse information.I.e. when the markup information of image-region B1 shows that B1 is to belong to include the region of article, then illustrate sample image Feature X1 is the sample image feature for belonging to article.Conversely, when the corresponding markup information of image-region shows to be not belonging to article, Then illustrate that sample image feature is not belonging to the sample image feature of article.

Step 1404, the target item image that will acquire is input to image detection model, and it is defeated to obtain image detection model Characteristics of image distributed intelligence out.

Step 1406, each images of items feature is matched with the standard article feature in item database, is determined The object of the corresponding type of goods of the target item for including in target item image and each target item in article placement space Grade is set.

Step 1408, comparison images of items is obtained, comparison images of items is input to image detection model.

Step 1410, the type of goods and each comparison article for determining the comparison article for including in comparison images of items are in object Comparison article position in product placement space.

Step 1412, according to the type of goods of target item and corresponding article position and compare article type of goods and Corresponding comparison article position determines the article changed in article placement space.

Target item image and the comparison images of items images of items that be two different, are by article placement space Image capture device collect.It, can be to image in order to guarantee that the distortion of image capture device acquired image is lower The placement location of acquisition equipment makes regulation.As shown in figure 15, a mounting height is recommended in figure, i.e. recommendation Image Acquisition is set The position of standby installation should be in the region of the 10cm apart from top.During practice, it can suitably increase article and place sky Between height, reduce article placement space width, after highly reaching, can guarantee installation image capture device can Capture complete article region, it can acquire the Item Information of placement complete.Length and numerical value that is wide and recommending in figure, It is recommended that the length and width of each article placement space, avoids article placement space excessive wide, cause image capture device without Method all collects the article of placement.

After getting target item image and comparison images of items, object can be identified by same identification method Location information of the type of goods and each article for including in product image and comparison images of items in article placement space.It can be with Target item image is input in image detection model, by the feature extraction network in image detection model to target item Image carries out feature extraction, obtains the Q-character confidence of corresponding characteristics of image and each characteristics of image in target item image Breath, and each characteristics of image is detected by the detection network of image detection model, it is wrapped in the target item image of output The Q-character confidence of the corresponding images of items feature of the target item contained and each images of items feature in target item image Breath.

For example, including 2 articles A, 2 articles B, 3 article C, then in target item in article placement space It also can include 2 articles A, 2 articles B, 3 article C in image.Target item image is input to image detection model In, image detection model can carry out feature extraction to target item image, wherein including the corresponding images of items of each article A Feature A1, A2, A3 ..., An etc., similarly, also can include the corresponding images of items feature B1, B2 of each article B, B3 ..., Bn etc. the and corresponding images of items feature C1, C2 of each article B, C3 ..., Cn etc., meanwhile, can also extract not including has Corresponding feature in the region of non-putting on article etc. in the characteristics of image of article, such as article placement space.Carrying out feature extraction When, while will record down location information of the feature extracted in target item image, such as record images of items spy The coordinate for levying A1 is (x1, y1, x1 ', y1 '), therefore image detection model is in the corresponding images of items feature of output target item When, feature locations information of each images of items feature in target item image can be exported simultaneously.In this way, with Standard article feature in item database is matched, to determine each target item included in article placement space When type of goods, each target item position corresponding in article placement space exact can be also known.

It is specific as schemed when the feature extraction network by image detection model carries out feature extraction to target item image Shown in 16A, target item image is input to the first convolution module of feature extraction network, by the first convolution module to object Product image carries out convolution algorithm.The fisrt feature of first convolution module output is normalized, can also normalized Fisrt feature is corrected after processing linear.To the input pond layer progress of the fisrt feature after linear process pond will be corrected Effect obtains pond as a result, the second convolution module pond result being input in feature extraction network carries out cyclic convolution fortune It calculates.The second feature for obtaining the output of the second convolution module, using second feature as the corresponding characteristics of image of target item image.Its In, it include that multilayer repeats convolution block and the non-duplicate convolution block of single layer in the second convolution module.As shown in fig 16b, due to multilayer It is the same for repeating the size for the vector matrix that convolution block is output and input, therefore it is multiple to connect, and the non-duplicate volume of single layer The matrix size that block is output and input is different, therefore can not connect between the non-duplicate convolution block of single layer.In order to true Protecting feature can repeat to carry out convolution operation between convolution block and the non-duplicate convolution block of single layer in multilayer, can carry out edge filling Padding operation, to ensure that feature can be repeated operation.

Step 1414, corresponding order is completed according to the article changed in article placement space to pay.

After the article changed in article placement space has been determined, it can be got according to the article acquisition request that terminal is sent User identifier, so as to generate corresponding pay invoice according to the article changed in user identifier and article placement space, eventually End can complete corresponding delivery operation according to pay invoice.

In traditional image recognition technology, when determining the type of goods in image, what is taken is image classification method. And when using image classification, it needs to train image classification model for a long time, and work as newly-increased there are article or occur When change, image classification model modification is complex, and the article for change is needed to be trained again.And in the present embodiment In, what is taken is that the feature in article characteristics and item database is carried out to retrieval matching, under this mode, without to article number It is trained according to library, and when increasing newly or changing there are article, is trained again without to item database, only The feature by change or newly-increased article is needed to be updated to item database.It can reach second grade to update, to realize higher The identification method of effect.It is this that be extracted and distinguished to target item image by image detection model wherein include article Target image characteristics, then by item database to the target image characteristics comparison match for including article, so that it is determined that target Type of goods and specific location information in the presence of images of items, can not only improve recognition efficiency, additionally it is possible to be promoted and be known Other accuracy rate.

In one embodiment, it is necessary to which the user for obtaining article can scan article placement space corresponding two by terminal Code or other identifier are tieed up, article acquisition request is triggered and is sent to server, carries user identifier in article acquisition request Space identification corresponding with article placement space.Server is transmittable to beat after the article acquisition request for receiving terminal transmission The instruction extremely article placement space corresponding with space identification in space is opened, so that the goalkeeper of article placement space can open, user The right to use to article placement space can be obtained.For example user then can freely obtain the article in article placement space.Mesh Mark images of items is needs to obtain the user of the article collected article figure before using terminal triggers article acquisition request Picture.After user has taken article in article placement space, the door of article placement space can be closed manually, thus user couple Terminate in the right to use of article placement space.After the door of article placement space is closed, article placement space will send correspondence Notice to server, server can get object after receiving the notice of " door is turned off " of article placement space transmission Image capture device in product placement space carries out the comparison images of items that Image Acquisition obtains again.Server can be according to mesh Mark images of items and comparison images of items determine the article changed in article placement space, that is, can determine that the user places in article The type of goods and number of articles taken away in space, so as to carry out phase to the corresponding account of user identifier according to the article of change The operation of withholing answered.In the two dimensional code mark of scanning input article placement space it may require that close payment is exempted from user's agreement, therefore After the article changed in article placement space has been determined, corresponding numerical value can be directly deducted in the account of user.

Fig. 2-Fig. 5, Figure 10-Figure 14 are respectively the flow diagram of item identification method in each embodiment.It should be understood that Although each step in each flow chart successively shows that these steps are not certainty according to the instruction of arrow It is successively executed according to the sequence that arrow indicates.Unless expressly stating otherwise herein, there is no stringent for the execution of these steps Sequence limits, these steps can execute in other order.Moreover, at least part step in each figure may include more Perhaps these sub-steps of multiple stages or stage are not necessarily to execute completion in synchronization to sub-steps, but can be with It executes at different times, the execution sequence in these sub-steps or stage, which is also not necessarily, successively to be carried out, but can be with it At least part of the sub-step or stage of its step or other steps executes in turn or alternately.

In one embodiment, as shown in figure 17, a kind of item identification devices are provided, comprising:

Image collection module 1702, for obtaining at least one target item in the target item figure of article placement space Picture.

Detection module 1704, for target item image to be input to image detection model；It is defeated to obtain image detection model Characteristics of image distributed intelligence out, characteristics of image distributed intelligence include the corresponding images of items feature of each target item and every A images of items feature corresponding feature locations information in target item image.

A matching module 1706, for carrying out the standard article feature in each images of items feature and item database Match, obtains corresponding matching degree；The corresponding target item type of each images of items feature is confirmed according to matching degree.

Type of goods determining module 1708, it is every for identifying according to characteristics of image distributed intelligence and target item type Article position and corresponding type of goods of a target item in article placement space.

In one embodiment, image collection module 1702 is also used to through the image capture device in article placement space It carries out Image Acquisition and obtains target item image；Standard article in item database is characterized in basis and image capture device The consistent image capture device of parameter is set and carries out what the image progress feature extraction that Image Acquisition obtains obtained.

In one embodiment, detection module 1704 is also used to the feature extraction network by image detection model to target Images of items carries out feature extraction, obtains corresponding images of items feature and each images of items feature in target item image Feature locations information；The detection network for obtaining image detection model detects images of items feature, each mesh of output Mark the corresponding images of items feature of article and each images of items feature corresponding feature locations information in target item image.

In one embodiment, detection module 1704 is also used to the first convolution module by feature extraction network to target Images of items carries out convolution algorithm；The fisrt feature of first convolution module output is normalized；Normalizing will be carried out Change the input pond layer progress pond effect of treated fisrt feature, obtains pond result；Pond result is input to feature to mention The second convolution module in network is taken to carry out cyclic convolution operation；The second feature for obtaining the output of the second convolution module, by second Feature is as the corresponding characteristics of image of target item image.

In one embodiment, it is corresponding more to be also used to obtain each type of goods in item database for matching module 1706 A standard article feature；It obtains European between images of items feature multiple standard article features corresponding with each type of goods Distance value；The matching between the standard article feature in images of items feature and item database is determined according to Euclidean distance value Degree.

In one embodiment, matching module 1706 is also used to images of items feature is corresponding with each type of goods respectively The Euclidean distance value of standard article feature be ranked up to obtain ranking results according to sequence from small to large；According to from small to large Direction the standard article feature of preset quantity is obtained from ranking results；It counts and belongs in the standard article feature of preset quantity The quantity of the standard article feature of the same type of goods, the quantity according to the corresponding standard article feature of each type of goods are big The corresponding target item type of small determining images of items feature.

In one embodiment, above-mentioned apparatus further includes line of demarcation determining module (not shown), for obtaining image Acquire the corresponding reference point of article placement space where equipment；The article placement space for obtaining image capture device acquisition is corresponding Spatial image；Spatial image is pre-processed；The boundary point of pretreated spatial image is obtained according to reference point；According to Boundary point determines the line of demarcation of article placement space.Detection module 1704 is also used to be determined in target item image according to line of demarcation Belong to the target image in article placement space, target image is input to image detection model.,

In one embodiment, above-mentioned apparatus further includes that item database establishes module (not shown), for obtaining Standard article image；Image procossing is carried out to standard article image, the quantity of images of items is extended to obtain standard article Image collection；Feature extraction is carried out to each image in standard article image collection；Duplicate removal is carried out to the feature that extraction obtains Screening；Feature after screening is stored according to the corresponding type of goods of standard item each in standard article image to product data In library.

In one embodiment, above-mentioned item database establishes module and is also used to obtain type of goods according to standard article Position and angle to standard article carry out the standard article image that Image Acquisition obtains after being adjusted.

In one embodiment, standard article is placed on the first cargo path region of article placement space.Above-mentioned product data Library is established module and is also used to when the height of standard article is lower than article height threshold value, in the second cargo path area of article placement space Domain placing height is higher than the article of article height threshold value, is adjusted according to the first preset rules to standard article, and carry out figure As collecting corresponding first standard article image collection；It is pre- according to second when standard article corresponds to multiple type of goods If rule is adjusted standard article, and carries out Image Acquisition and obtain corresponding second standard article image collection；Work as standard When article is the article of default type of goods, standard article is adjusted according to third preset rules, and carry out Image Acquisition Obtain corresponding third standard article image collection.

In one embodiment, above-mentioned apparatus further includes training module (not shown), for carrying out to article sample Image Acquisition obtains article sample image；Feature extraction is carried out to article sample image, obtains multiple sample image features；It obtains Take the image-region and corresponding markup information divided to article sample image；According to image-region and corresponding mark Note information determines the sample image feature for belonging to article；Label information is added to sample image feature according to the information of article；It is right The weight of image detection model is adjusted, and belongs to the sample image feature of article and corresponding according to weight adjusted output Location information；The predictablity rate of image detection model is obtained according to the label information of sample image feature；Work as predictablity rate When reaching accuracy rate threshold value, trained image detection model is obtained.

In one embodiment, above-mentioned training module is also used to obtain the degree of overlapping of sample image feature and image-region； When degree of overlapping reaches anti-eclipse threshold, then determine that the sample image for belonging to article is special according to the corresponding markup information of image-region Sign；When the corresponding markup information of image-region shows to belong to article, then sample image feature belongs to the sample image spy of article Sign；When the corresponding markup information of image-region shows to be not belonging to article, then sample image feature is not belonging to the sample graph of article As feature.

In one embodiment, above-mentioned apparatus further includes database update module (not shown), for when to target When type of goods in images of items identifies mistake, the wrong type of goods of identification target image in target item image is obtained Region and the corresponding correct type of goods markup information of object region；Object region is input to image detection model In feature extraction network carry out feature extraction；Obtain the target that feature extraction network extracts object region Article characteristics；Target item feature is updated to corresponding article kind in item database according to correct type of goods markup information In class.

In one embodiment, target item image is to carry out image by the image capture device in article placement space It collects.Above-mentioned apparatus further includes comparison module (not shown), for obtaining comparison images of items, compares article figure It seem to carry out Image Acquisition by image capture device after changing the article in article placement space to obtain；It will comparison Images of items is input to image detection model, and the comparison article until obtaining including in comparison images of items and comparison article are in object Article position in product placement space；According to target item and target item article placement space article position and comparison The article position of product and comparison article in article placement space, determines the article changed in article placement space.

In one embodiment, above-mentioned apparatus further includes payment module (not shown), for obtaining terminal transmission Article acquisition request carries user identifier in article acquisition request；According to the article changed in user identifier, article placement space Generate corresponding pay invoice；Pay invoice is sent to terminal, completes corresponding delivery operation for terminal.

Figure 18 shows the internal structure chart of computer equipment in one embodiment.The computer equipment specifically can be figure Image detection model 120 or item database 130 in 1.As shown in figure 18, which includes the computer equipment packet Include processor, memory and the network interface connected by system bus.Wherein, memory include non-volatile memory medium and Built-in storage.The non-volatile memory medium of the computer equipment is stored with operating system, can also be stored with computer program, should When computer program is executed by processor, processor may make to realize item identification method.It can also be stored in the built-in storage Computer program when the computer program is executed by processor, may make processor to execute item identification method.

It will be understood by those skilled in the art that structure shown in Figure 18, only part relevant to application scheme The block diagram of structure, does not constitute the restriction for the computer equipment being applied thereon to application scheme, and specific computer is set Standby may include perhaps combining certain components or with different component layouts than more or fewer components as shown in the figure.

In one embodiment, item identification devices provided by the present application can be implemented as a kind of shape of computer program Formula, computer program can be run in computer equipment as shown in figure 18.Composition can be stored in the memory of computer equipment Each program module of the item identification devices, for example, image collection module, detection module shown in Figure 17, matching module and Type of goods determining module.The computer program that each program module is constituted makes processor execute described in this specification Apply for the step in the item identification method of each embodiment.

In one embodiment, a kind of computer equipment, including memory and processor are provided, is stored in memory Computer program, the processor realize the article identification side provided in any one embodiment of the application when executing computer program The step of method.

In one embodiment, a kind of computer readable storage medium is provided, computer program is stored thereon with, is calculated Machine program realizes the step of item identification method provided in any one embodiment of the application when being executed by processor.

Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the program can be stored in a non-volatile computer and can be read In storage medium, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, provided herein Each embodiment used in any reference to memory, storage, database or other media, may each comprise non-volatile And/or volatile memory.Nonvolatile memory may include that read-only memory (ROM), programming ROM (PROM), electricity can be compiled Journey ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory may include random access memory (RAM) or external cache.By way of illustration and not limitation, RAM is available in many forms, such as static state RAM (SRAM), dynamic ram (DRAM), synchronous dram (SDRAM), double data rate sdram (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronization link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) directly RAM (RDRAM), straight Connect memory bus dynamic ram (DRDRAM) and memory bus dynamic ram (RDRAM) etc..

Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance Shield all should be considered as described in this specification.

The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously The limitation to the application the scope of the patents therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art For, without departing from the concept of this application, various modifications and improvements can be made, these belong to the guarantor of the application Protect range.Therefore, the scope of protection shall be subject to the appended claims for the application patent.

Claims

1. a kind of item identification method, comprising:

The target item image is input to image detection model；

The characteristics of image distributed intelligence of described image detection model output is obtained, described image feature distribution information includes each institute It states the corresponding images of items feature of target item and each images of items feature is corresponding in the target item image Feature locations information；

Each images of items feature is matched with the standard article feature in item database, obtains corresponding matching Degree；

According to described image feature distribution information and the target item type, identify each target item in the object Article position and corresponding type of goods in product placement space.

2. the method according to claim 1, wherein the acquisition target item image includes:

Image Acquisition, which is carried out, by the image capture device in article placement space obtains the target item image；

Standard article in the item database is characterized according to the consistent figure of setting parameter with described image acquisition equipment Carry out what feature extraction obtained as acquisition equipment carries out the image that Image Acquisition obtains.

3. the method according to claim 1, wherein the image for obtaining the output of described image detection model is special Levy distributed intelligence, comprising:

Feature extraction is carried out to the target item image by the feature extraction network of described image detection model, is corresponded to Feature locations information in the target item image of images of items feature and each images of items feature；

The detection network for obtaining described image detection model detects the images of items feature, each of output mesh Mark the corresponding images of items feature of article and each images of items feature corresponding feature in the target item image Location information.

4. the method according to claim 1, wherein described by each images of items feature and product data Standard article feature in library is matched, and corresponding matching degree is obtained, comprising:

Obtain the corresponding multiple standard article features of each type of goods in the item database；

Obtain the Euclidean distance value between the images of items feature multiple standard article features corresponding with each type of goods；

It is determined according to the Euclidean distance value between the standard article feature in the images of items feature and item database Matching degree.

5. the method according to claim 1, wherein being placed in described at least one target item of acquisition in article Before the target item image in space, further includes:

Obtain the corresponding reference point of article placement space where image capture device；

Obtain the corresponding spatial image of the article placement space of described image acquisition equipment acquisition；

The spatial image is pre-processed；

The boundary point of pretreated spatial image is obtained according to the reference point；

The line of demarcation of the article placement space is determined according to the boundary point；

The target item image is input to image detection model, comprising:

The target image belonged in the article placement space in the target item image is determined according to the line of demarcation, by institute It states target image and is input to image detection model.

6. the method according to claim 1, wherein the establishment step of the item database, comprising:

Obtain standard article image；

Image procossing is carried out to the standard article image, the quantity of the images of items is extended to obtain standard article figure Image set closes；

Feature extraction is carried out to each image in standard article image collection；

Duplicate removal screening is carried out to the feature that extraction obtains；

Feature after screening is stored according to the corresponding type of goods of standard item each in the standard article image to described In item database.

7. the method according to claim 1, wherein the standard article is placed on the article placement space First cargo path region；The acquisition standard article image, comprising:

When the height of the standard article is lower than article height threshold value, put in the second cargo path region of the article placement space The article that height is higher than article height threshold value is set, the standard article is adjusted according to the first preset rules, and carry out figure As collecting corresponding first standard article image collection；When the standard article corresponds to multiple type of goods, according to Two preset rules are adjusted the standard article, and carry out Image Acquisition and obtain corresponding second standard article image set It closes；

When the standard article is the article of default type of goods, the standard article is adjusted according to third preset rules It is whole, and carry out Image Acquisition and obtain corresponding third standard article image collection.

8. the method according to claim 1, wherein the training step of described image detection model, comprising:

Image Acquisition is carried out to article sample, obtains article sample image；

Feature extraction is carried out to the article sample image, obtains multiple sample image features；

Obtain the image-region and corresponding markup information divided to the article sample image；

The sample image feature for belonging to article is determined according to image-region and corresponding markup information；

Label information is added to the sample image feature according to the information of the article；

The weight of described image detection model is adjusted, the sample image for belonging to article according to weight adjusted output is special Sign and corresponding location information；

The predictablity rate of described image detection model is obtained according to the label information of the sample image feature；

When the predictablity rate reaches accuracy rate threshold value, trained image detection model is obtained.

9. according to the method described in claim 8, it is characterized in that, described determine according to image-region and corresponding markup information Belong to the sample image feature of article, comprising:

Obtain the degree of overlapping of the sample image feature and described image region；

When the degree of overlapping reaches anti-eclipse threshold, then article is belonged to according to the corresponding markup information determination in described image region Sample image feature；

When described image region, corresponding markup information shows to belong to article, then the sample image feature belongs to the sample of article This characteristics of image；

When described image region, corresponding markup information shows to be not belonging to article, then the sample image feature is not belonging to article Sample image feature.

10. the method according to claim 1, wherein obtaining wrapping in the target item image in the identification After the location information of the type of the target item contained and the target item, further includes:

When identifying mistake to the type of goods in the target item image, the type of goods of identification mistake is obtained in the mesh Mark object region and the corresponding correct type of goods markup information of the object region in images of items；

The object region is input to the feature extraction network in described image detection model and carries out feature extraction；

Obtain the target item feature that the feature extraction network extracts the object region；

The target item feature is updated in the item database according to the correct type of goods markup information and is corresponded to Type of goods in.

11. the method according to claim 1, wherein the target item image is by article placement space Interior image capture device carries out what Image Acquisition obtained, the method also includes:

Comparison images of items is obtained, the comparison images of items is led to after changing to the article in the article placement space It crosses described image acquisition equipment and carries out what Image Acquisition obtained；

The comparison images of items is input to described image detection model, until obtaining including in the comparison images of items Compare the article position of article and the comparison article in the article placement space；

According to the target item and the target item the article placement space article position and the comparison article And article position of the comparison article in the article placement space, determine the object changed in the article placement space Product.

12. according to the method for claim 11, which is characterized in that the method also includes:

The article acquisition request that terminal is sent is obtained, carries user identifier in the article acquisition request；

Corresponding pay invoice is generated according to the article changed in the user identifier, the article placement space；

The pay invoice is sent to the terminal, completes corresponding delivery operation for the terminal.

13. a kind of item identification devices, which is characterized in that described device includes:

Detection module, for the target item image to be input to image detection model；It is defeated to obtain described image detection model Characteristics of image distributed intelligence out, described image feature distribution information include that the corresponding images of items of each target item is special Sign and each images of items feature corresponding feature locations information in the target item image；

Matching module, for each images of items feature to be matched with the standard article feature in item database, Obtain corresponding matching degree；The corresponding target item type of each images of items feature is confirmed according to matching degree；

Type of goods determining module, it is every for identifying according to described image feature distribution information and the target item type Article position and corresponding type of goods of a target item in the article placement space.

14. a kind of computer readable storage medium is stored with computer program, when the computer program is executed by processor, So that the processor is executed such as the step of any one of claims 1 to 12 the method.

15. a kind of computer equipment, including memory and processor, the memory is stored with computer program, the calculating When machine program is executed by the processor, so that the processor is executed such as any one of claims 1 to 12 the method Step.