CN111461694B - Dish identification and pricing system and method based on multi-level learning model - Google Patents

Dish identification and pricing system and method based on multi-level learning model Download PDF

Info

Publication number
CN111461694B
CN111461694B CN202010189968.9A CN202010189968A CN111461694B CN 111461694 B CN111461694 B CN 111461694B CN 202010189968 A CN202010189968 A CN 202010189968A CN 111461694 B CN111461694 B CN 111461694B
Authority
CN
China
Prior art keywords
model
restaurant
layer
identification
dish
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010189968.9A
Other languages
Chinese (zh)
Other versions
CN111461694A (en
Inventor
杨傲雷
邱育
徐昱琳
费敏锐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Shanghai for Science and Technology
Original Assignee
University of Shanghai for Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Shanghai for Science and Technology filed Critical University of Shanghai for Science and Technology
Priority to CN202010189968.9A priority Critical patent/CN111461694B/en
Publication of CN111461694A publication Critical patent/CN111461694A/en
Application granted granted Critical
Publication of CN111461694B publication Critical patent/CN111461694B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q20/00—Payment architectures, schemes or protocols
    • G06Q20/08—Payment architectures
    • G06Q20/14—Payment architectures specially adapted for billing systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/12—Hotels or restaurants
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07G—REGISTERING THE RECEIPT OF CASH, VALUABLES, OR TOKENS
    • G07G1/00—Cash registers
    • G07G1/12—Cash registers electronically operated
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Abstract

The invention discloses a dish identification and pricing system and method based on a multi-level learning model. The system is logically composed of a multi-layer structure from bottom to top, and comprises a terminal interaction layer, a restaurant identification layer and a cloud optimization layer, wherein an information flow closed loop iteration update mechanism exists between the layers, and the upper layer model optimization fusion and the lower layer model iteration update and identification are cooperatively completed. The system takes a restaurant identification layer as a functional core, interacts with the lower layer of the menu image information and iteratively trains an updated closed-loop path of the menu identification model; and a closed loop path which is optimized and fused with the upper layer composition global dish identification model. The system hierarchical structure design decouples the function of a complex dish identification system, and improves the richness and accuracy of dish identification models in a mode of iterative updating and optimizing of different hierarchical model distribution. Over time, various indexes identified by the system can be gradually improved, and the longer the system is used, the higher the accuracy of identification is.

Description

Dish identification and pricing system and method based on multi-level learning model
Technical Field
The invention belongs to the field of artificial intelligent recognition and computer vision detection, and particularly relates to a dish recognition and pricing system and method based on a multi-level learning model.
Background
In the rush hour of dining, the scene that the dining queue of densely populated college dining halls, self-selected dining halls generally remains for tens of meters, and the staff in the dining room window is very busy and needs to hold dishes and hold meals for customers and calculate prices fast, so that the problems of low price calculating speed, easy price calculating mistakes and the like are unavoidable, and the economic losses of students, customers or dining halls and dining halls are directly caused. In response to these problems, the market places new demands on restaurant services, such as: the price is rapid and accurate; manual intervention is reduced as much as possible; the dish identification has an intelligent identification mechanism which can be memorized continuously and can deal with continuous change of canteen dishes, etc. Related solutions can be divided into three categories, via retrieval of review material:
the first is in the manner of adding auxiliary devices or machine codes to the dishes. The method mainly utilizes chips embedded in dishes or machine code marks printed on the surfaces to assist in identification, different dishes correspond to different dishes, when the method is used, dishes with certain dishes are placed in dishes corresponding to the dishes, and when the method is used, the dishes corresponding to the dishes can be obtained through identifying the chips or the machine code marks. Such solutions include chinese patent number CN201520183235, publication date: 2015, 07 month 22 days, patent name: an automatic billing system with a mobile settlement terminal obtains a corresponding serving price of the food by reading information of an RFID chip of the food container. Chinese patent number CN107798521, publication date: 2018, 03, 13, patent name: a dish identification and cashing system and method based on machine code image identification, the system identifies the machine code on the outer package through a visual camera, so as to obtain the dish price. The method is an indirect method, has poor flexibility and high cost of dishes, and can generate loss along with the increase of the service time of chips and machine code information on the dishes. In addition, the machine code is also easy to be covered by the dish soup and cannot be identified.
The second type is a manner in which dishes and cutlery used for serving dishes are themselves identified as features. The method mainly trains the shape and color characteristics of the tableware by a machine learning method so as to classify the dishes. Dishes with different types and different prices can be placed in dishes with specific shapes and colors corresponding to the different prices. Chinese patent publication No. CN103971471, publication date: 2014, 8 months and 6 days, patent name: an automatic dinner plate pricing method based on color recognition and a system thereof, wherein the system utilizes a color sensor to distinguish prices through the recognition of the colors at the bottom of the dinner plate. The method indirectly realizes the price of dishes by identifying the characteristics of dishes, and the dishes are required to be placed in dishes with specific shapes and colors defined in advance, so that the situations of more dishes and more price standards cannot be dealt with because the shapes and the colors of the dishes are limited after all.
And the third category, in which the dishes themselves are used as the identification features. The method mainly utilizes a machine learning method to train different dish characteristics to identify dishes, and directly obtains dish price information corresponding to the dishes. Chinese patent publication No. CN201410311841, publication date: 10.01.2014, patent name: the self-service payment device of the self-service restaurant based on the image recognition and the application method thereof, the method carries out offline training and recognition on the dishes by applying a method based on a convolutional neural network, but because offline training is adopted, new dishes are added and need to be subjected to model training again by developers, so that a dish model cannot be timely adjusted according to the characteristics of each canteen dish, and the practicality and the efficiency are lower. In addition, chinese patent publication No. CN201610070288, publication date: 2016, 07, 06, patent name: the automatic charging method for the self-selected restaurant and the implementation device thereof are characterized in that dishes to be identified are put into a dish identification model obtained by training all dish data of the self-selected area in the last period of time for identification. According to the method, the new dish identification model is trained only through a small number of images in a short time, the previous dish information is abandoned, the dish identification model obtained through the previous training is not fully utilized, and the identification accuracy and the robustness of the dish identification model have a certain problem.
Disclosure of Invention
In order to solve the problems in the scheme, the invention provides a dish identification and pricing system and method based on a multi-level learning model, and provides a system and method based on a multi-level learning framework, which are combined with universality and specificity of the dish identification model, so that the identification system has stronger reliability.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the system mainly comprises three layers from bottom to top, namely a terminal interaction layer, a restaurant identification layer and a cloud optimization layer, wherein an information flow transmission mechanism exists among the three layers, and the three layers are iteratively updated and interdependence.
The terminal interaction layer is at the lowest layer and is mainly responsible for dish image acquisition and customer man-machine interaction and pricing, and the terminal interaction layer consists of hardware terminals for dish image acquisition and man-machine interaction in a restaurant, and each terminal mainly comprises a vision acquisition camera, an interaction display, an IC card swiping machine, a terminal controller and a dish placement area.
The restaurant identification layer is positioned in the middle layer and consists of all restaurant identification systems in the system, each restaurant in the layer is responsible for maintaining respective dish identification models, the restaurant identification layer periodically obtains a global fusion-optimized dish model-cloud optimization model-from the upper layer-cloud optimization layer, periodically obtains a dish sample image from the lower layer-cloud optimization model-and further carries out incremental learning training to generate a dish identification model-restaurant characteristic model-matched with the dish characteristics of the restaurant, responds to the identification request of the terminal interaction layer, and periodically sends the model to the cloud optimization layer.
The cloud optimization layer is located at the uppermost layer and is mainly responsible for establishing a cloud initial model to generate a cloud basic model, and collecting and model fusion of different dish identification models of a plurality of restaurants in the subsequent restaurant identification layer to generate a cloud optimization model. The cloud optimizing layer maintains a unified cloud dish identifying model, and periodically issues a cloud optimizing model to the restaurant identifying layer for supporting the cloud optimizing model to construct a restaurant characteristic model adapting to a restaurant.
Among three levels of dish identification and pricing systems, the restaurant identification layer is the functional core, and has dynamic bidirectional information paths with both the lower layer (terminal interaction layer) and the upper layer (cloud optimization layer). Bi-directional path to the lower layer: and acquiring the dish image and the interaction information from the lower layer, training, identifying and pricing the dish identification model, and feeding back the identification and pricing result to the lower layer. Bi-directional path to upper layer: and periodically sending the model trained by the layer to the upper layer, receiving the system dish identification model optimized by the upper layer, and then carrying out the next incremental learning training on the model. In a word, the cloud optimization layer is used for coordinating model data of different restaurants in the restaurant identification layer, and dish identification knowledge learned by other restaurants can be 'migrated' to the identification model of the current restaurant through model optimization, fusion and release of the cloud level, so that the cloud optimization layer has the function of sharing the data and the model knowledge.
The running process of the multi-level learning model in the invention mainly comprises 3 parts in the implementation logic: initializing, constructing and acquiring a dish identification model in the whole system; iterative training updating and optimizing of dish identification models; the dish identification and pricing system operates.
The dish identification and pricing method based on the multi-level learning model comprises the following operation procedures:
1. initializing, constructing and acquiring a cloud base model and a restaurant base model for system dish identification.
Step 1, a cloud base model in a cloud optimization layer "Is a construction of (3).
Step 1-1, further collecting dish images on the basis of an ImageNet dish data set, manufacturing a large dish sample data set, and adding dish labels.
Step 1-2, inputting the collected dish sample data set into a convolutional neural network for training a dish identification model based on the collected dish sample data set to obtain an initial dish identification cloud base model
Step 1-3, the cloud optimization layer models a cloud base modelRelease to the restaurant where the "join applications" are presented.
Step 2, newly opened restaurant K dish identification 'restaurant basic model'Is performed in the first step.
Step 2-1, a newly opened restaurant K puts forward a 'joining application' to the cloud optimization layer, and then waits for the response of the cloud optimization layer.
Step 2-2, receiving a model release starting response released by the cloud optimization layer by the restaurant K, and simultaneously receiving a cloud base model "
Step 2-3. Through each terminal T in the restaurant K terminal interaction layer K_i (i=1, 2.,. N represents terminal identification), collect all dish image samples of the restaurant, and make labels to construct a restaurant K dish training sample dataset
Step 2-4. Fusion of the received dish identification cloud modelSample training dataset constructed with restaurant K>Training a dish identification model by adopting an incremental learning method, and constructing a restaurant basic model for identifying dishes in the restaurant "
Step 3, a restaurant is newly opened later, such as restaurant K+1, and the dish identification 'restaurant basic model' is acquired, and the step 2 is returned.
2. Iterative training updating and optimizing of the system dish identification model.
The method mainly relates to multi-model fusion training updating of a dish identification 'cloud optimization model' in a system cloud optimization layer and iterative training updating of a dish identification 'restaurant characteristic model' in a restaurant identification layer. Dish identification cloud optimization model and restaurant characteristic model are used as requiredDifferent periodic updates, for example: the cloud optimization model is updated every week, and the restaurant feature model is updated every day. The dining hall recognition layer of the dish recognition and pricing system comprises K restaurants, wherein the dish characteristic model of the restaurant K on the N-1 day is that The dish optimization model of the cloud optimization layer at week W is +.>Restaurant K nth day dish identification "restaurant feature model" and cloud W+1th week dish identification "cloud optimization model">The updating and optimizing steps of (a) are as follows:
step 1, iterative training update of restaurant K nth day dish identification 'restaurant characteristic model'.
Step 1-1, if the restaurant K currently receives a cloud W-th week dish identification model issued by a cloud optimization layerThen preference will->Performing network pruning to obtain a model->Otherwise, selecting the identification model of the restaurant K on the N-1 th dayPerforming network pruning to obtain a model->
Step 1-2. Through each terminal T in the restaurant K terminal interaction layer K_i (i=1, 2.,. N represents terminal identification), and original menu map collected on day N of labeling is collectedConstructing a sample dataset from a sampleMeanwhile, collecting newly added dish image samples collected on the N-th day of labeling, and constructing a sample data set +.>
Step 1-3. Adding a sample set of dish imagesAnd performing incremental learning training to obtain a model +.>Add dish image sample set +.>And performing incremental learning training to obtain a model +.>
Step 1-4. ModelingOr->And model->Model fusion is carried out by adopting an incremental learning method, and a dish identification restaurant characteristic model is obtained >
Step 1-5. The latest model obtained at present is used forFor use inThe original model is updated to be used as a dish identification model of the restaurant K on the N day, and then the dish identification model is used for subsequent dish identification.
Step 1-6. Identifying the N-th dish of restaurant K as the characteristic model of restaurantReporting to a cloud optimization layer.
And 2, updating fusion training of the cloud optimization model for the W+1th week dish identification.
Step 2-1, cloud optimization model of the W-th week of the cloud optimization layerPerforming network pruning to obtain a model->
Step 2-2, the cloud optimization layer receives a characteristic model set reported by K restaurants in the restaurant identification layerWherein->A feature model set representing restaurant K, W Zhou Qitian.
Step 2-3, adopting an incremental learning method to modelAnd->Fusing to obtain a dish identification 'cloud optimization model' of the W+1th week of the cloud optimization layer>
Step 2-4. Cloud optimization layer updates the model at week W+1Issue to each restaurant in the restaurant identification layer.
3. Dish identification and pricing system operation
Step 1, acquiring and obtaining dish image information in a current scene when each terminal in a restaurant K terminal interaction layer interacts with a customer.
And 2, uploading the acquired information such as the dish images to a server for running a dish identification model in the restaurant K of the restaurant identification layer by the interactive terminal, and providing a dish identification service request to the dish identification server.
And 3, receiving a terminal service request by a dish identification server of the restaurant K, and identifying the uploaded dish image and pricing the dish price by adopting a latest dish identification model.
And 4, the dish identification server responds to the identification and pricing results and feeds back the identification and pricing results to the interactive terminal which makes a service request.
And step 5, after receiving the response of the identification and pricing results, the interactive terminal displays the name and price of the dishes on a display screen for customers to refer to, and simultaneously sends price information to a payment module for subsequent payment.
In the above steps, the incremental learning training method mainly means that a learning system can continuously learn new knowledge from new samples and can save most of the previously learned knowledge. The incremental learning method adopted here is to fuse the newly added dish identification model with the original dish identification model on the basis of a current dish identification model, and the generated new model inherits most of the weights of the two models. The implementation method is described as follows: (1) pruning the current neural network model: the weight values in each convolution layer and the complete connection layer in the neural network are sequenced according to absolute values, the lowest weight value sequenced according to fixed percentage is removed by utilizing the neural network weight pruning technology, and redundant parameters are eliminated to release a model parameter space; (2) retraining the pruned model: and (3) adjusting parameter disorder caused by network pruning, keeping parameters obtained after training adjustment fixed, and repeatedly executing the process to add a plurality of model training tasks by using the released parameter space for learning new tasks.
For a single restaurant, the restaurant dish identification system can continuously expand the model richness through the dish image information acquired by the dish identification terminal in the restaurant, so that the identification accuracy of the dish identification model is improved. For a restaurant system composed of a plurality of restaurants, each restaurant dish identification system can acquire dish identification model characteristics of other restaurants through the cloud layer, further integrate the dish characteristics of other restaurants and improve the model richness. Meanwhile, when a restaurant is newly built, the dish identification cloud model can be directly transferred to a new restaurant dish identification system, all information of the dish identification cloud model is inherited, so that the newly built restaurant can directly use the built dish identification model, and the restaurant is also iteratively updated based on the model.
Compared with the prior art, the invention obviously has the outstanding substantive characteristics and obvious technical progress that:
(1) Architecture based on a multi-level learning model is presented. The system adopts a hierarchical structure design, decouples the function of a complex dish identification system, and improves the richness and accuracy of the dish identification model in a mode of iterative updating and optimizing of different hierarchical models. Over time, various indexes identified by the system can be gradually improved, and the longer the system is used, the higher the accuracy of identification is.
(2) The universality and the special characteristics of the dish identification model are fused, and the different chefs of each restaurant can make different dishes. The specific fusion process is summarized as follows: firstly, establishing a universal dish identification 'cloud base model' by a cloud optimization layer; secondly, a newly built restaurant in the restaurant identification layer acquires a current dish identification 'cloud model' from the cloud as a restaurant dish identification 'basic model'; then, based on the model, incremental learning training is performed again according to actual image samples of dishes made in the restaurant. Along with the time, the restaurant continuously iterates the restaurant model, generates a dish identification 'restaurant characteristic model' of the restaurant, and reports the restaurant characteristic model to the cloud optimization layer regularly; and the cloud optimization layer receives all reported restaurant feature models, and performs optimization fusion with the previous cloud model to obtain a 'cloud optimization model'. And repeating the steps, namely performing loop iteration, optimization and fusion.
(3) The identification system has strong expandability. The dining room of the dining room recognition layer in the system can be dynamically added and deleted, so that the normal operation of the system is not affected, the number of the dining room characteristic models can be increased by adding the dining room, and the accuracy of cloud model recognition is improved.
Drawings
FIG. 1 is a three-layer architecture and information flow diagram of a dish identification and pricing system;
FIG. 2 is a schematic diagram of a three-layer architecture implementation of a dish identification and pricing system;
FIG. 3 is a schematic diagram of a terminal structure of a dish identification and pricing system;
FIG. 4 is a schematic diagram of iterative update of restaurant 1, day 71 dish identification "restaurant feature model";
FIG. 5 is a schematic diagram of a cloud optimization model update flow at week 11 of the cloud optimization layer;
FIG. 6 is a logic flow diagram of the operation of the dish identification and pricing system;
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and preferred embodiments, but the scope of the present invention is not limited to the following specific embodiments. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, shall fall within the scope of the present application. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs, the terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the present invention.
Example 1
Referring to fig. 1-6, the dish identification and pricing system based on the multi-level learning model mainly comprises a plurality of layers from bottom to top, wherein the system comprises a terminal interaction layer, a restaurant identification layer and a cloud optimization layer, an information flow closed loop iteration update mechanism exists among the layers, and the upper layer model optimization fusion and the lower layer identification model iteration update are cooperatively completed.
Example two
This embodiment is substantially the same as the first embodiment, and is specifically as follows:
the restaurant identification layer is a functional core of the information flow transmission mechanism among multiple layers, and the restaurant identification layer, the lower terminal interaction layer and the upper cloud end optimization layer are provided with dynamic bidirectional closed-loop information paths. Wherein, with the two-way passageway of lower floor: acquiring a dish image and interaction information from the lower layer, performing iterative training, recognition and pricing of a dish recognition model, and feeding back recognition and pricing results to the lower layer; bi-directional path to upper layer: and periodically sending the latest model trained by the layer to the upper layer, receiving the system dish identification model optimized by the upper layer, and then carrying out the next incremental learning training on the model.
The terminal interaction layer is positioned at the lowest layer and is responsible for dish image acquisition, customer man-machine interaction and pricing; the restaurant identification layer is positioned in the middle layer and consists of all restaurant identification systems in the system, the restaurant identification layer regularly obtains a global fused optimized dish model from the cloud optimization layer to form a cloud optimization model, regularly obtains a dish sample image from the terminal interaction layer to generate a restaurant characteristic model, and regularly sends the model to the cloud optimization layer; the cloud optimization layer is located at the uppermost layer and is mainly responsible for collecting and fusing different dish identification models of a plurality of restaurants in the subsequent restaurant identification layer, so that a cloud optimization model is generated; the cloud optimizing layer maintains a unified cloud dish identifying model, and periodically issues the cloud optimizing model to the restaurant identifying layer for supporting the restaurant identifying layer to construct a restaurant characteristic model suitable for the restaurant.
Example III
The dish identification and pricing method based on the multi-level learning model adopts the system to operate, and the operation program comprises the following steps:
1) In the multi-level iterative updating mechanism, in the operation process of a multi-level learning model, the initialization construction and acquisition of a cloud base model and a restaurant base model are carried out on dish identification in the whole system.
2) The dish identification 'cloud optimization model' and 'restaurant characteristic model' are updated according to different periods as required. The dining hall recognition layer of the dish recognition and pricing system comprises K restaurants, wherein the dish characteristic model of the restaurant K on the N-1 day is thatThe dish optimization model of the cloud optimization layer at week W is +.>Updating and optimizing a restaurant characteristic model for the N-th dish identification of a restaurant K and a cloud optimizing model for the W+1th week dish identification of a cloud.
3) The dish identification and pricing system runs a logic flow.
4) Dynamic joining and exiting procedure of dining hall in dining hall identification layer.
Example IV
This embodiment is substantially the same as embodiment three, with the following specific points:
the specific steps of the operation procedure 1) are as follows:
step 1, a cloud base model in a cloud optimization layer "Is constructed according to the following steps;
step 1-1, further collecting dish images on the basis of an ImageNet dish data set;
Step 1-2, training to obtain an initial dish identification cloud base model based on the collected dish sample data set
Step 1-3, the cloud optimization layer models a cloud base modelRelease to restaurant where "join applications" are presented;
step 2, newly opened restaurant K dish identification 'restaurant basic model'Is obtained;
step 2-1, a newly opened restaurant K puts forward a 'joining application' to a cloud optimization layer, and then waits for the response of the cloud optimization layer;
step 2-2, receiving a model release starting response released by the cloud optimization layer by the restaurant K, and simultaneously receiving a cloud base model "
Step 2-3. Through each terminal T in the restaurant K terminal interaction layer K_i (i=1, 2,., N represents terminal identification), collect all dish image samples of the restaurant, construct a restaurant K dish training sample dataset
Step 2-4. Fusion of the receivedAnd->Training a model by adopting an incremental learning method, and constructing a restaurant basic model for identifying dishes in the restaurant>
Step 3, a restaurant is newly opened later, such as restaurant K+1, and the dish identification 'restaurant basic model' is acquired, and the step 2 is returned.
In the iterative update mechanism between the multiple layers, the specific steps of the operation program 2) are as follows:
Step 1, iterative training update of restaurant K nth day dish identification 'restaurant characteristic model'.
Step 1-1, if the restaurant K currently receives a cloud W-th week dish identification model issued by a cloud optimization layerThen preference will->Performing network pruning to obtain a model->Otherwise, selecting the identification model of the restaurant K on the N-1 th dayPerforming network pruning to obtain a model->
Step 1-2. Through each terminal T in the restaurant K terminal interaction layer K_i (i=1, 2,., N represents terminal identification), original menu image samples collected on day N of labeling are collected, and a sample data set is constructedMeanwhile, collecting newly added dish image samples collected on the N-th day of labeling, and constructing a sample data set +.>
Step 1-3. Adding a sample set of dish imagesAnd performing incremental learning training to obtain a model +.>Add dish image sample set +.>And performing incremental learning training to obtain a model +.>
Step 1-4. ModelingOr->And model->Model fusion is carried out by adopting an incremental learning method, and a dish identification restaurant characteristic model is obtained>
Step 1-5. The latest model obtained at present is used forThe method is used for updating the original model to be used as a dish identification model of the restaurant K on the N day, and further used for subsequent dish identification.
Step 1-6. Identifying the N-th dish of restaurant K as the characteristic model of restaurantReporting to a cloud optimization layer.
And 2, updating fusion training of the cloud optimization model for the W+1th week dish identification.
Step 2-1, cloud optimization model of the W-th week of the cloud optimization layerPerforming network pruning to obtain a model->
Step 2-2, the cloud optimization layer receives a characteristic model set reported by K restaurants in the restaurant identification layerWherein->A feature model set representing restaurant K, W Zhou Qitian.
Step 2-3, adopting an incremental learning method to modelAnd->Fusing to obtain a dish identification 'cloud optimization model' of the W+1th week of the cloud optimization layer>
Step 2-4. Cloud optimization layer updates the model at week W+1Issue to each restaurant in the restaurant identification layer.
The specific operation steps of the operation program 3) are as follows:
step 1, initializing a dish identification and pricing system;
step 2, judging whether the dishes need to be trained, if not, entering step 3, and if so, entering step 4;
step 3, identifying dishes and paying price in the terminal interaction layer controller;
step 3-1, placing a dish tray to be identified;
step 3-2, collecting a dish image, sending an identification service request to a restaurant identification layer server, and waiting for a response;
Step 3-3, receiving a 'recognition result response' from the restaurant recognition layer server to obtain a dish recognition result;
step 3-4, displaying the names and the prices of the dishes in the terminal interactive interface;
step 3-5, judging whether to pay, if not, circularly waiting, and if so, returning to the step 3-1;
step 4, constructing and updating a restaurant special model of a restaurant identification layer server;
step 4-1, acquiring and labeling a dish image sample, and simultaneously receiving a cloud optimization model issued by a cloud optimization layer and a restaurant characteristic model issued by a restaurant;
step 4-2, performing incremental learning training based on the collected image samples and various models to obtain a restaurant characteristic model at the current moment;
step 4-3, saving a current restaurant characteristic model for supporting an identification service request of a terminal interaction layer, and simultaneously finishing iterative updating of a restaurant model;
and 4-4, judging whether to report the restaurant characteristic model to a cloud optimization layer. If not, returning to the step 2, and if uploading is needed, adding the step 5;
step 5, constructing and updating a cloud optimization model of a cloud optimization layer server;
Step 5-1, periodically collecting 'restaurant characteristic models' reported by all restaurants by a cloud optimization layer;
step 5-2, model increment learning training and fusion are carried out on the cloud end, and a 'cloud end optimization model' is constructed;
and 5-3, judging whether to update the model to the restaurant optimization layer, returning to the step 5-1 if the model does not need to be updated, and returning to the step 4-1 if the model does not need to be updated.
The specific operation steps of the operation program 4) are as follows: the restaurant in the restaurant identification layer can be dynamically added and withdrawn, the newly-built restaurant N provides a 'adding application' for the cloud optimization layer, the cloud optimization layer feeds back a 'model release start' to realize the dynamic adding of the restaurant, the to-be-withdrawn restaurant M provides a 'withdrawing application' for the cloud optimization layer, and the cloud optimization layer feeds back a 'model release stop' to realize the dynamic withdrawing of the restaurant.
Example five
The embodiment is realized by a hierarchical structure of a dish identification and pricing system and a model identification updating loop.
In this embodiment, the dish identification and pricing system consists of three layers: the terminal interaction layer, the restaurant identification layer and the cloud optimization layer have information flow transmission mechanisms among the three layers, are interdependence, and cooperatively complete an upper layer model updating mechanism and a lower layer identification updating mechanism, and one embodiment of the information flow transmission mechanisms is shown in fig. 1. The cloud optimizing layer comprises a dish identification cloud model, the restaurant identification layer comprises a dish identification restaurant model, and the terminal interaction layer comprises a dish identification terminal in the restaurant. Two information updating loops exist between the three layers, namely: an upper model updating loop between the cloud optimization layer and the restaurant identification layer and a lower identification updating loop between the restaurant identification layer and the terminal interaction layer. The cloud optimization layer is responsible for fusion updating of cloud models and regularly issuing the latest optimized recognition model to each restaurant of the restaurant recognition layer. And each restaurant further performs incremental learning and iterative updating of the identification model of the restaurant according to the characteristics of the dishes of the restaurant on the basis of the acquired cloud model, and periodically reports the model obtained by restaurant training to the cloud optimization layer. And each terminal of the terminal interaction layer is responsible for information interaction between terminal equipment and a customer, acquires and uploads a menu image selected by the customer and man-machine interaction information to a restaurant server in the identification layer, and then the server identifies the menu image based on the latest menu identification model and feeds back identification and pricing results to the interaction terminal for display.
Example six
The embodiment is three-layer architecture hardware implementation of a dish identification and pricing system.
In this embodiment, a system three-tier architecture hardware implementation is shown in fig. 2. The hardware carrier of the cloud end optimization layer in the system is a cloud server, and establishment of an initial dish identification model and optimization updating of a subsequent identification model are completed. The restaurant recognition layer consists of a series of restaurant servers which are independent of each other, no independent communication channel exists between the restaurant servers, and the knowledge interaction of the recognition model is realized through the model fusion and release of the cloud optimization layer. The restaurant in the restaurant identification layer can be dynamically added and withdrawn, the newly-built restaurant N provides a 'adding application' for the cloud optimization layer, the cloud optimization layer feeds back a 'model release start' to realize the dynamic adding of the restaurant, the to-be-withdrawn restaurant M provides a 'withdrawing application' for the cloud optimization layer, and the cloud optimization layer feeds back a 'model release stop' to realize the dynamic withdrawing of the restaurant.
The terminal interaction layer of each restaurant comprises different numbers of interaction terminals, each terminal structure schematic diagram is shown in fig. 3, and the marks in the drawings are respectively: 1. a vision acquisition camera; 2. an interactive display; 3. an IC card reader; 4. a terminal controller; 5. a dish placement area; 6. a bracket; 7. an auxiliary light source. Installation layout of terminal: the auxiliary light source 7 is arranged above the vision acquisition camera 1; the high-definition camera is arranged right above the dish placing area 5, and the visual field requirement can cover the coverage range specified by the dish placing area 5; the terminal controller 4 is arranged under the desk top of the cash register; the IC card swiping machine 3, the dish placing area 5 and the interactive display 2 are all arranged on the cash register surface; the camera 1 is connected with the terminal controller 4 through a network port; the IC card reader 3 is connected with the terminal controller 4 through a serial port.
Example seven
The embodiment is an example construction of a cloud base model and a restaurant base model for system dish identification.
In this embodiment, the construction of the system dish identification model begins with the construction of the dish identification "cloud base model" in the cloud optimization layer. Now, a description will be given by taking a first dish identification cloud model of a cloud optimization layer and a first dish identification 'restaurant basic model' of a restaurant identification layer as examples.
Step 1, identifying a cloud base model by a first dish in a cloud optimization layer "Is a construction of (3).
And step 1-1, further collecting dish images on the basis of the ImageNet dish data set, manufacturing a large dish data set, and adding dish labels.
Step 1-2, based on the collected sample data set, inputting the sample data set into a convolutional neural network to train a dish identification model, and obtaining a first dish identification cloud base model
Step 1-3. The cloud optimization layer obtains a cloud modelTo the first restaurant 1 where the "join application" was filed.
Step 2, a cloud model is obtained by a first restaurant 1 in a restaurant identification layerOn the basis of the model, a dish identification 'restaurant basic model' is constructed.
Step 2-1, the restaurant 1 puts forward a 'joining application' to a cloud optimization layer server, and waits for the response of the cloud optimization layer.
Step 2-2, restaurant 1 receives the "model release start" response issued by the cloud optimization layer, and receives the "cloud base model" issued by the cloud optimization layer server at the same time;
step 2-3. Each terminal T in the terminal interaction layer is interacted with through restaurant 1 1_i (i=1, 2.,. N represents terminal identification), collect all dish image samples of the restaurant, and make labels to construct a restaurant 1 dish training sample dataset
Step 2-4. Fusion of the received dish identification cloud modelSample training dataset constructed with restaurant 1 +.>Training a dish identification model by adopting an incremental learning method, and constructing a restaurant 1 dish identification restaurant basic model>
Example eight
The embodiment builds a model of restaurant features for example of system dish identification.
The present embodiment is described by taking the system including 3 restaurants and 5 terminals in each restaurant as an example. Setting: the 11 th monday of system operation is currently; the dish identification 'restaurant characteristic model' of 3 restaurants in the restaurant identification layer is iteratively trained and updated once a day. The construction process of the restaurant 1 dish identification 'restaurant characteristic model' is shown in fig. 4, the construction process of the restaurant 2 and restaurant 3 models, and so on.
Step 1, receiving cloud 11 th week dish identification 'cloud optimization model' issued by a cloud optimization layer by restaurants 1-3 " And performing network pruning on the plant to obtain a model +.>(in the case of Tuesday weekly, "restaurant feature model" for the day 1-3 of the previous day of restaurants was selected for network pruning).
Step 2 restaurant 1 through its 5 terminals T 1_i (i=1, 2,3,4,5 represents terminal identification), original menu image samples collected on the 71 st day of labeling are collected, and a sample data set is constructedMeanwhile, collecting a newly added dish image sample collected on the 71 st day of labeling, and constructing a sample data set +.>Restaurant 2, restaurant 3 and the like。
Step 3, adding a dish image sample set in the restaurant 1And performing incremental learning training to obtain a modelAdd dish image sample set +.>And performing incremental learning training to obtain a model +.>Restaurant 2, restaurant 3, and so on.
Step 4, restaurant 1 modelAnd model->Model fusion is carried out by adopting an incremental learning method, and a dish identification restaurant characteristic model is obtained>Restaurant 2, restaurant 3 and the like are respectively given +.>
Step 5, restaurant 1 uses the current latest modelThe method is used for updating the original model to be used as a dish identification 'restaurant characteristic model' of the restaurant 1 on the 72 th day, and then is used for subsequent dish identification. Restaurant 2, restaurant 3, and so on.
Step 6, identifying the dishes on the 71 st day by the restaurants 1-3 into a restaurant characteristic model And respectively reporting to a cloud optimization layer.
Example nine
The embodiment is an example construction of a cloud optimization model for system dish identification.
The present embodiment is described by taking the system including 3 restaurants and 5 terminals in each restaurant as an example. Setting: the 11 th week of system operation is currently; and optimizing updating and releasing once every week of a dish identification 'cloud optimizing model' in the cloud optimizing layer server. The construction process of the cloud optimization model of the system is shown in fig. 5.
Step 1, a cloud optimization model of the 11 th week of a cloud optimization layerPerforming network pruning to obtain a model->
Step 2, the cloud optimization layer receives restaurant characteristic models reported by 3 restaurants of the restaurant identification layer in total in 11 th week and 7 days, and builds a characteristic model setWherein->A feature model set constructed in restaurant 1 week 11 and 7 is shown. Similarly, the case of->Representing a feature model set constructed in 7 days of week 11 for restaurant 2 and restaurant 3.
Step 3, the cloud optimization layer adopts an incremental learning method to modelAnd->Fusing to obtain a dish identification 'cloud optimization model' of the 12 th week of the cloud optimization layer>
Step 4, the cloud optimization layer optimizes the optimized 'cloud optimization model'Release to each restaurant in the restaurant-identifying layer on day 78 (day 1 of week 12) for supporting each restaurant for dish identification construction of the "restaurant-feature model".
Examples ten
The embodiment is an embodiment of newly-built restaurant joining and original restaurant exiting of a dish identification and pricing system.
The embodiment is described taking the system as an example, which comprises 3 restaurants, 4 th restaurant (with 5 terminals) and 2 nd restaurant. Setting: currently, the 11 th week of system operation.
(1) New restaurant 4 joins
Step 1, a server of the newly opened restaurant 4 puts forward a 'joining application' to the cloud optimization layer, and waits for the response of the cloud optimization layer.
And 2, after receiving the 'joining request' of the restaurant 4, the cloud optimization layer sends a 'model issuing start' instruction to the restaurant 4 if the joining condition is met.
Step 3, receiving a model release starting instruction released by a cloud optimization layer by the restaurant 4, and simultaneously receiving a cloud optimization model "
Step 4, restaurant 4 is connected with 5 terminals T in the terminal interaction layer 4_i (i=1, 2.,. 5 denotes terminal identification), collect all dish image samples of the restaurant, and make labels to construct a restaurant 4 dish training sample dataset
Step 5, fusing the received dish identification cloud optimization model "Sample training dataset constructed with restaurant 4 at current week 11 +.>Training a dish identification model by adopting an incremental learning method, and constructing a restaurant 4 dish identification restaurant characteristic model >
(2) Original restaurant 2 exits
Step 1, a server of the original restaurant 2 puts forward an exit application to a cloud optimization layer, and waits for response of the cloud optimization layer.
And 2, after receiving the ' exit application ' of the restaurant 2, if the ' exit condition is met, the cloud optimization layer sends a ' model release termination ' instruction to the restaurant 2, and simultaneously interrupts the subsequent release of the ' cloud optimization model ' to the restaurant 2.
And 3, after receiving a model release termination instruction released by the cloud optimization layer, the restaurant 2 server automatically deletes all menu identification models of the restaurant 2 and backups of the cloud model.
Example eleven
The embodiment is an embodiment of a logic process for operating a dish identification and pricing system.
In this embodiment, a logic flow chart of the dish identification and pricing system operation is shown in fig. 6, and the specific steps are as follows:
and step 1, initializing a dish identification and pricing system.
And step 2, judging whether the dishes need to be trained, if not, entering the step 3, and if so, entering the step 4.
And 3, identifying dishes and paying in a pricing manner in the terminal interaction layer controller.
And 3-1, placing a dish tray to be identified.
And 3-2, collecting a dish image, and sending an identification service request to a restaurant identification layer server to wait for a response.
And 3-3, receiving a 'recognition result response' from the restaurant recognition layer server, and obtaining a dish recognition result.
And 3-4, displaying the names and the prices of the dishes in the terminal interactive interface.
And 3-5, judging whether payment is carried out, if not, carrying out cyclic waiting, and if so, returning to the step 3-1.
And 4, constructing and updating a restaurant characteristic model of the restaurant identification layer server.
And 4-1, collecting and labeling a dish image sample, and simultaneously receiving a cloud optimization model issued by a cloud optimization layer and a restaurant characteristic model issued by a restaurant of the user.
And 4-2, performing incremental learning training based on the collected image samples and various models to obtain a restaurant characteristic model at the current moment.
And 4-3, saving the current restaurant characteristic model for supporting the identification service request of the terminal interaction layer, and simultaneously finishing the iterative updating of the restaurant model.
And 4-4, judging whether to report the restaurant characteristic model to a cloud optimization layer. If not, returning to the step 2, and if uploading is needed, adding the step 5.
And 5, constructing and updating a cloud optimization model of the cloud optimization layer server.
And 5-1, periodically collecting 'restaurant characteristic models' reported by all restaurants by a cloud optimization layer.
And 5-2, performing model increment learning training and fusion on the cloud end to construct a 'cloud end optimization model'.
And 5-3, judging whether to update the model to the restaurant optimization layer, returning to the step 5-1 if the model does not need to be updated, and returning to the step 4-1 if the model does not need to be updated.
The above embodiments are only for illustrating the technical solution of the present invention, and the present invention is not limited to the above specific embodiments, but can be modified in many ways. It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention.

Claims (6)

1. A dish identification and pricing method based on a multi-level learning model is operated by adopting a dish identification and pricing system based on the multi-level learning model, wherein the dish identification and pricing system based on the multi-level learning model consists of a plurality of layers from bottom to top, and comprises a terminal interaction layer, a restaurant identification layer and a cloud optimization layer, an information flow closed-loop iteration update mechanism exists among the layers, and the upper model optimization fusion and the lower model iteration update are cooperatively completed.
1) In the multi-level iterative updating mechanism, in the running process of a multi-level learning model, the initialization construction and acquisition of a cloud base model and a restaurant base model are carried out on the dish identification in the whole system;
2) The dish identification cloud optimization model and the restaurant characteristic model are updated according to different periods as required; the dining hall recognition layer of the dish recognition and pricing system comprises K restaurants, wherein the dish characteristic model of the restaurant K on the N-1 day is thatThe dish optimization model of the cloud optimization layer at week W is +.>Updating and optimizing a restaurant K nth day dish identification restaurant characteristic model and a cloud end W+1th week dish identification cloud end optimizing model;
3) The dish identification and pricing system operates a logic flow;
4) Dynamic joining and exiting program of dining hall in dining hall identification layer;
in the iterative update mechanism among the multiple layers, the specific steps of the operation program 2) are as follows:
step 2-1, iterative training update of restaurant characteristic models for restaurant K nth day dishes identification;
step 2-1-1. If the restaurant K currently receives the cloud W-th week dish identification model issued by the cloud optimization layerThen preference will->Performing network pruning to obtain a model->Otherwise, select the identification model for the N-1 th day of restaurant K- >Performing network pruning to obtain a model->
Step 2-1-2. Through restaurant K terminal each terminal T in the interaction layer K_i Collecting original dish image samples collected on the N th day of labeling, and constructing a sample data setWhere i=1, 2, &.. i represents a terminal identifier; meanwhile, collecting newly added dish image samples collected on the N-th day of labeling, and constructing a sample data set +.>
Step 2-1-3. Adding a sample set of dish imagesAnd performing incremental learning training to obtain a model +.>Add dish image sample set +.>And performing incremental learning training to obtain a model +.>
Step 2-1-4. ModelingOr->And model->Model fusion is carried out by adopting an incremental learning method to obtain a dish identification restaurant characteristic model ++>
Step 2-1-5. The latest model obtained at present is used for the modelThe method is used for updating the original model to be used as a dish identification model of the restaurant K on the N day, and further used for subsequent dish identification;
step 2-1-6. Identifying the N-th dish of restaurant K as the characteristic model of restaurantReporting to a cloud optimization layer;
step 2-2, fusion training update of cloud optimization models is performed on the W+1th week dishes in the cloud;
step 2-2-1, a cloud optimization model of the W-th week of the cloud optimization layer is obtainedPerforming network pruning to obtain a model- >
Step 2-2-2. The cloud optimization layer receives the characteristic model set reported by K restaurants in the restaurant identification layerWherein->A feature model set representing restaurant K, W Zhou Qitian;
step 2-2-3. Model is obtained by adopting the incremental learning methodAnd->Fusion is carried out, and a dish identification cloud optimization model of the W+1th week of the cloud optimization layer is obtained>
Step 2-2-4. Cloud optimization layer updates model at week W+1Issue to each restaurant in the restaurant identification layer.
2. The dish identification and pricing method based on the multi-level learning model according to claim 1, wherein the information flow transmission mechanism among the multiple levels is that the restaurant identification layer is a functional core, and the restaurant identification layer, the lower terminal interaction layer and the upper cloud end optimization layer are provided with dynamic bidirectional closed-loop information paths; the bidirectional passage is connected with the lower layer, acquires dish images and interaction information from the lower layer, performs iterative training, identification and pricing of a dish identification model, and feeds back identification and pricing results to the lower layer; bi-directional path to upper layer: and periodically sending the latest model trained by the layer to the upper layer, receiving the system dish identification model optimized by the upper layer, and then performing the next round of incremental learning training on the model.
3. The dish identification and pricing method based on the multi-level learning model according to claim 1, wherein the terminal interaction layer is at the lowest layer and is responsible for dish image acquisition and customer man-machine interaction and pricing; the restaurant identification layer is positioned in the middle layer and consists of all restaurant identification systems in the system, the restaurant identification layer regularly obtains a global fused optimized dish model from the cloud optimization layer to form a cloud optimization model, regularly obtains a dish sample image from the terminal interaction layer to generate a restaurant characteristic model, and regularly sends the model to the cloud optimization layer; the cloud optimization layer is positioned at the uppermost layer and is responsible for collecting and fusing different dish identification models of a plurality of restaurants in the subsequent restaurant identification layer, so as to generate a cloud optimization model; the cloud optimizing layer maintains a unified cloud dish identifying model, and periodically issues the cloud optimizing model to the restaurant identifying layer for supporting the restaurant identifying layer to construct a restaurant characteristic model suitable for the restaurant.
4. The dish identification and pricing method based on multi-level learning model as set forth in claim 1, wherein the specific steps of the operation procedure 1) are as follows:
Step 1-1. Cloud end base model in cloud optimization layerIs constructed according to the following steps;
step 1-1-1, further collecting dish images on the basis of an ImageNet dish data set;
step 1-1-2, training to obtain an initial dish identification cloud base model based on the collected dish sample data set
Step 1-1-3. The cloud optimization layer models the cloud base modelIssuing to a restaurant in which the joining application is presented;
step 1-2. New open restaurant K dish identification restaurant basic modelIs obtained;
step 1-2-1, a newly opened restaurant K puts forward a joining application to a cloud optimization layer, and then waits for the response of the cloud optimization layer;
step 1-2-2. The restaurant K receives a model release starting response released by the cloud optimization layer and receives a cloud base model at the same time
Step 1-2-3. Through restaurant K terminal each terminal T in the interaction layer K_i Collecting all dish image samples of a restaurant, and constructing a K dish training sample data set of the restaurantWhere i=1, 2, &.. i represents a terminal identifier;
step 1-2-4. Fusion of the receivedAnd->Training a model by adopting an incremental learning method, and constructing a restaurant dish identification restaurant basic model +.>
Step 1-3, a restaurant is newly opened later, such as restaurant K+1, and the acquisition of the dish identification restaurant basic model returns to step 1-2.
5. The dish identification and pricing method based on multi-level learning model according to claim 1, wherein the specific operation steps of the operation procedure 3) are as follows:
step 3-1, initializing a dish identification and pricing system;
step 3-2, judging whether the dishes need to be trained, if not, entering step 3-3, and if so, entering step 3-4;
step 3-3, identifying dishes and paying price in the terminal interaction layer controller;
step 3-3-1, placing a dish tray to be identified;
step 3-3-2, collecting a dish image, sending an identification service request to a restaurant identification layer server, and waiting for a response;
step 3-3-3, receiving a recognition result response from the restaurant recognition layer server to obtain a dish recognition result;
step 3-3-4, displaying the names and prices of the dishes in the terminal interactive interface;
step 3-3-5, judging whether to pay, if not, waiting circularly, and if so, returning to the step 3-3-1;
step 3-4, constructing and updating a restaurant characteristic model of the restaurant identification layer server;
step 3-4-1, acquiring and labeling a dish image sample, and simultaneously receiving a cloud optimization model issued by a cloud optimization layer and a restaurant characteristic model issued by a restaurant of the user;
Step 3-4-2, performing incremental learning training based on the collected image samples and various models to obtain a restaurant characteristic model of the restaurant at the current moment;
step 3-4-3, saving the current restaurant characteristic model for supporting the identification service request of the terminal interaction layer, and simultaneously finishing the iterative updating of the restaurant model;
step 3-4-4, judging whether to report the restaurant characteristic model to a cloud optimization layer; if not, returning to the step 3-2, and if so, entering the step 3-5;
step 3-5, constructing and updating a cloud optimization model of the cloud optimization layer server;
step 3-5-1, periodically collecting restaurant characteristic models reported by all restaurants by a cloud optimization layer;
step 3-5-2, model increment learning training and fusion are carried out on the cloud end, and a cloud end optimization model is constructed;
and step 3-5-3, judging whether to update the model to the restaurant identification layer, returning to step 3-5-1 if updating is not needed, and returning to step 3-4-1 if updating is needed.
6. The dish identification and pricing method based on multi-level learning model according to claim 1, wherein the specific operation steps of the operation program 4) are: the restaurants in the restaurant identification layer can be dynamically added and withdrawn, the newly built restaurants k+1 provide an adding application to the cloud optimization layer, the cloud optimization layer feedback model is issued and started to realize the dynamic adding of the restaurants, the to-be-withdrawn restaurants M provide a withdrawing application to the cloud optimization layer, and the cloud optimization layer feedback model is issued and stopped to realize the dynamic withdrawing of the restaurants.
CN202010189968.9A 2020-03-17 2020-03-17 Dish identification and pricing system and method based on multi-level learning model Active CN111461694B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010189968.9A CN111461694B (en) 2020-03-17 2020-03-17 Dish identification and pricing system and method based on multi-level learning model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010189968.9A CN111461694B (en) 2020-03-17 2020-03-17 Dish identification and pricing system and method based on multi-level learning model

Publications (2)

Publication Number Publication Date
CN111461694A CN111461694A (en) 2020-07-28
CN111461694B true CN111461694B (en) 2023-07-18

Family

ID=71683217

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010189968.9A Active CN111461694B (en) 2020-03-17 2020-03-17 Dish identification and pricing system and method based on multi-level learning model

Country Status (1)

Country Link
CN (1) CN111461694B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107766889A (en) * 2017-10-26 2018-03-06 济南浪潮高新科技投资发展有限公司 A kind of the deep learning computing system and method for the fusion of high in the clouds edge calculations
CN109118200A (en) * 2018-07-26 2019-01-01 上海凯景信息技术有限公司 A kind of commodity identification and cash register system based on image recognition
CN109147170A (en) * 2018-07-26 2019-01-04 上海凯景信息技术有限公司 A kind of self-service cabinet based on image recognition
CN110197128A (en) * 2019-05-08 2019-09-03 华南理工大学 The recognition of face architecture design method planned as a whole based on edge calculations and cloud
CN110414373A (en) * 2019-07-08 2019-11-05 武汉大学 A kind of deep learning palm vein identification system and method based on cloud side end cooperated computing
CN110705684A (en) * 2019-08-22 2020-01-17 中国科学院计算技术研究所 Environment self-adaptive learning method and system based on end cloud cooperation
CN110866588A (en) * 2019-11-08 2020-03-06 中国科学院软件研究所 Training learning method and system for realizing individuation of learnable ability model of intelligent virtual digital animal

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190236360A1 (en) * 2018-01-30 2019-08-01 Mashgin Inc. Feedback loop for image-based recognition
US11257004B2 (en) * 2018-07-31 2022-02-22 Ncr Corporation Reinforcement machine learning for item detection

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107766889A (en) * 2017-10-26 2018-03-06 济南浪潮高新科技投资发展有限公司 A kind of the deep learning computing system and method for the fusion of high in the clouds edge calculations
CN109118200A (en) * 2018-07-26 2019-01-01 上海凯景信息技术有限公司 A kind of commodity identification and cash register system based on image recognition
CN109147170A (en) * 2018-07-26 2019-01-04 上海凯景信息技术有限公司 A kind of self-service cabinet based on image recognition
CN110197128A (en) * 2019-05-08 2019-09-03 华南理工大学 The recognition of face architecture design method planned as a whole based on edge calculations and cloud
CN110414373A (en) * 2019-07-08 2019-11-05 武汉大学 A kind of deep learning palm vein identification system and method based on cloud side end cooperated computing
CN110705684A (en) * 2019-08-22 2020-01-17 中国科学院计算技术研究所 Environment self-adaptive learning method and system based on end cloud cooperation
CN110866588A (en) * 2019-11-08 2020-03-06 中国科学院软件研究所 Training learning method and system for realizing individuation of learnable ability model of intelligent virtual digital animal

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Shota Horiguchi et al..Personalized Classifier for Food Image Recognition.《IEEE Transactions on Multimedia》.2018,第20卷(第10期),第2836-2848页. *
丁春辉.基于深度学习的暴力检测及人脸识别方法研究.《中国博士学位论文全文数据库 信息科技辑》.2017,(第09期),论文第69-71页. *

Also Published As

Publication number Publication date
CN111461694A (en) 2020-07-28

Similar Documents

Publication Publication Date Title
US20210035069A1 (en) Data structures, graphical user interfaces, and computer-implemented processes for automation of project management
CN107392405B (en) Data processing method, device and equipment
CN107122730A (en) Free dining room automatic price method
CN107423923A (en) Material flow tracking and dispatching method using dispatching person as core
CN109615184A (en) The method and system of shops, retailer automatic cargo allocation, the goods that replenishes, adjusts
CN106126587A (en) Multinode data based on flow process display recording mechanism
CN108364146A (en) Logistics distribution emulation mode and device
CN109784531A (en) A kind of production plan generates processing method, system, platform and storage medium
CN106815699A (en) Distributed cloud framework Intelligent logistics delivery system
CN108427772A (en) Online report form generation method, system, equipment and storage medium
CN107968819A (en) A kind of new task management system and its operating procedure based on APP
Kung Object-oriented Software Engineering
CN109389245A (en) A kind of multifactor fusion school district school age population prediction technique based on deep neural network
CN111695747A (en) Intelligent dispatching method and device and computer readable storage medium
CN110322337A (en) A kind of inquiry business looks into multiple method and device automatically
CN111461694B (en) Dish identification and pricing system and method based on multi-level learning model
CN102968918A (en) Civil aviation passenger ticket reservation training system
CN106372980A (en) Cross-e-commerce channel order information acquisition and management method
CN104899688A (en) Unified asset management and planning method for center and distribution substations
CN104616073A (en) Precise management system based on digitized two-dimensional animation production
CN106004716B (en) A kind of intelligent logistic car networked system
CN106096917A (en) Application rule editor administrative mechanism
CN107341632A (en) A kind of data capture method and device based on user's request
AU2013201074A1 (en) System and method for management of event attendance packages and event attendance inventory
US20070118416A1 (en) Method and system for planning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant