WO2018077121A1

WO2018077121A1 - Method for recognizing target object in image, method for recognizing food article in refrigerator and system

Info

Publication number: WO2018077121A1
Application number: PCT/CN2017/107099
Authority: WO
Inventors: 徐达; 唐军
Original assignee: 合肥美的智能科技有限公司; 合肥华凌股份有限公司; 合肥美的电冰箱有限公司
Priority date: 2016-10-24
Filing date: 2017-10-20
Publication date: 2018-05-03
Also published as: CN106529427A

Abstract

A method for recognizing a target object in an image, a method for recognizing a food article in a refrigerator, and a system. The method for recognizing a target object in an image comprises: acquiring a training image and using the training image to establish an image recognition model; acquiring a test image, matching the test image with the image recognition model so as to recognize a target object in the test image, wherein the acquired training image and the acquired test image undergo normalization processing, and the training image and the test image are mapped to a uniform space; and/or carrying out regularization processing when establishing the image recognition model, so that the image recognition model expresses full sample distribution. The method allows for the performance of the image recognition model on the test data be as close as possible to the performance on the training data, thereby achieving consistency of image recognition and improving accuracy of image recognition.

Description

Target object recognition method in image, food identification method and system in refrigerator

Technical field

The present invention relates to the field of image recognition technologies, and in particular, to a target object recognition method in an image, a food identification method and system in a refrigerator.

Background technique

In the existing image processing technology, the image acquisition and the acquisition of the image cannot guarantee the consistency of the image. This difference causes the information of the learning to intelligently reflect some characteristics of the data, so the performance on the test data is not satisfactory, and the recognition rate of the image recognition is better. low.

Especially in the field of image recognition technology in refrigerators, there are many types of refrigerators. If each refrigerator is photographed separately and image recognition is performed, the recognition rate will be affected by the difference in picture quality. The images taken by different models of refrigerators are inconsistent, and the images taken by foods placed in different positions inside the refrigerator may also differ. The difference in images is expressed in brightness and color. Due to the different training and adopted images, this difference leads to the learning information only reflecting part of the characteristics of the data, so the performance on the test data is not ideal, and the recognition rate of image recognition is not high.

Summary of the invention

The technical problem to be solved by the present invention is to provide a target object recognition method in an image, a food identification method and system in a refrigerator, in view of the deficiencies of the prior art.

The technical solution of the present invention to solve the above technical problem is as follows: a method for identifying a target object in an image, comprising the following steps:

Obtaining a training image, and establishing an image recognition model using the training image;

Obtaining a test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

Wherein, the acquired training image and the acquired test image are normalized, and the training image is to be And the test image is mapped to a unified space; and/or a regularization process is performed when the image recognition model is established, so that the image recognition model expresses the full sample distribution.

In order to achieve the above object, the present invention also provides a food identification method in a refrigerator, comprising collecting image data in a refrigerator as a training image and a test image, and obtaining image data to be identified according to the target object recognition method in the image according to the above technical solution. The property information of the food.

To achieve the above object, the present invention also provides an object recognition system in an image, comprising:

a training module, configured to acquire a training image, and use the training image to establish an image recognition model;

a test module for acquiring a test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

a normalization module, configured to normalize the acquired training image and the acquired test image, and map the training image and the test image to a unified space;

A regularization module is used to perform regularization processing when the image recognition model is established, so that the image recognition model expresses a full sample distribution.

In order to achieve the above object, the present invention further provides a food identification system in a refrigerator, comprising: an image collection device and a server installed inside the refrigerator, wherein the image collection device collects image data in the refrigerator and uploads the image to a server, wherein the server adopts The target object recognition system in the image described in the above technical solution obtains attribute information of the food in the image data to be identified.

The invention has the beneficial effects that the invention normalizes the training image and the test image, maps the training image and the test image to a unified space, and makes the training sample and the test sample share the same, and the image recognition model obtained by using the training data Good test results can be obtained on the test data; by regularizing the image recognition model, the image recognition model expresses the full sample distribution, ignoring the components that only describe the training samples, and making the image recognition model express the full sample. Distribution, reduce the parameter space; through the above processing, the performance of the image recognition model on the test data is as close as possible to the performance on the training data, achieving consistency of image recognition and improving image recognition accuracy.

DRAWINGS

1 is a flowchart of a method for identifying a target object in an image according to an embodiment of the present invention;

2 is a flowchart of a method for identifying a target object in an image according to an embodiment of the present invention;

FIG. 3 is a flowchart of a method for identifying a target object in an image according to an embodiment of the present invention;

4 is a block diagram of a target object recognition system in an image according to an embodiment of the present invention;

FIG. 5 is a block diagram of a method for identifying a target object in an image according to an embodiment of the present invention;

FIG. 6 is a block diagram of a method for identifying a target object in an image according to an embodiment of the present invention;

FIG. 7 is a block diagram of a food identification system in a refrigerator according to an embodiment of the present invention.

detailed description

The principles and features of the present invention are described in the following with reference to the accompanying drawings.

As shown in FIG. 1 , an embodiment of the present invention provides a method for identifying a target object in an image, including the following steps:

S110: acquiring a training image in a training phase, normalizing the training image, and establishing an image recognition model by using the training image;

S120: obtaining a test image in a test phase, normalizing the test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

The training image acquired during the training phase and the test image acquired during the test phase are normalized, and the training image and the test image are mapped to a unified space.

In the above embodiment, the training image and the test image are normalized, and the training image and the test image are mapped to a unified space, so that the training sample and the test sample are distributed in the same manner, and the image recognition model obtained by using the training data is on the test data. Good test results can be obtained; regularization processing is performed when the image recognition model is established during the training phase, so that the image recognition model expresses the full sample distribution. Ignore the components that only describe the training samples, so that the image recognition model expresses the full sample distribution and reduces the parameter space; through the above processing, the performance of the model on the test data is as close as possible to the performance on the training data.

Optionally, as an embodiment of the present invention, the training phase acquiring the training image and the testing phase acquiring the test image have the same shooting environment.

In this embodiment, it is ensured that the image acquired in the training phase and the image acquired in the test phase have the same shooting environment, and the consistency of the training image and the test image is ensured from the hardware.

Optionally, as an embodiment of the present invention, the step of establishing an image recognition model by using the training image in the testing phase specifically includes:

a, obtaining a predetermined number of image data including the target object;

b, calibrating the attribute parameter of the target object in the image data, and using the calibrated image data as a training image;

c, setting an image recognition model, an initial model parameter, and a loss function, and obtaining a risk function by calculating a desired loss function;

d. Updating the image recognition model by using the risk function and the training data, so that the image recognition model satisfies the requirement of the loss function, and obtain an image recognition model corresponding to the target object.

Specifically, the attribute parameter of the target object includes the position and type of the target object in the calibration image data.

Specifically, in this embodiment, the image recognition model identifies a mapping relationship between the output and the input, the loss function representing the difference between the actual output and the model output; the risk function is the expectation of the loss function.

Specifically, in this embodiment, the attribute parameter of the target object in the image data may be calibrated by using a point labeling method: the food is in the picture by the coordinates (x, y), width w and h of the upper left corner of the food. Location information. At the same time, the type information of the food material is given. No matter how you mark it, you can determine the position and type of the object of interest in the picture.

In the embodiment of the present invention, the image recognition model is

Where x is image data,

For model output, θ is the model parameter;

The loss function is

L=(y,f(x,θ))

The loss function L represents the difference between the actual output and the model output, y is the actual output; f(x, θ) is the image recognition model;

The risk function is

Where R(θ) is a risk function, N represents the total number of training images, and i represents the i-th training image.

Image recognition includes a training phase and a testing phase.

The training phase includes: 1. obtaining training data by shooting and marking; 2. setting a decision function (model) to establish a mapping relationship between input and output; 3. setting an evaluation function to measure the quality of the decision function; 4. learning the algorithm and The training data updates the decision function so that the decision function satisfies the requirements of the evaluation function. Among them, the risk function loses the expectation of the function, and the risk function is minimized by a learning algorithm such as backpropagation.

Specifically, for the obtained image data (x), the target object position and kind (y) are calibrated as training data. Set the image recognition model and initial parameters, express the relationship between output and input, expressed by the decision function f(.), for the parameter x that has been calibrated, the output is

θ is the model parameter. Define the loss function L = (y, f (x, θ)), which represents the difference between the actual output and the model output. Then evaluate the risk of the decision function on all samples

The risk function is calculated on a known sample (x, y) and is called empirical risk. The training problem is transformed into solving θ ^* such that the empirical risk is minimal θ ^* = argminR(θ).

The test phase, specifically, includes: 1. taking a picture of the target object, obtaining a test sample, that is, inputting the parameter x; 2 obtaining the output result through the pre-trained image recognition model

The recognition result is the position and type of the target object.

In the embodiment of the present invention, the risk function is calculated on the known sample (x, y), which is called empirical risk, and the training problem is transformed into solving θ ^* , so that the empirical risk is minimum θ ^* = argminR (θ).

In the embodiment of the present invention, the risk function is an empirical risk, and the empirical risk is obtained by calculating a desired loss function according to all training data; and obtaining a model parameter that minimizes the empirical risk of the image recognition model as an image recognition model by calculation The final model parameter obtains an image recognition model corresponding to the target object.

It should be noted that, in the embodiment of the present invention, a plurality of algorithms may be used to train the model parameters. In the embodiment of the present invention, only the foregoing implementation method is introduced, and other implementation methods are also within the protection scope of the present invention. .

Image recognition model training is performed by machine learning algorithms, but this practice usually has training overfitting and the resulting generalization errors. Specifically, the model parameter θ ^{* is} determined according to the training sample (x, y), and the training sample usually does not respond well to the true distribution. The model with the least risk criterion is good for the training sample, but But can not adapt to the test data outside the training set. Therefore, the embodiment of the present invention solves the above technical problem by a normalization and regularization processing manner.

The normalization technique maps training samples and test samples to a unified space by transforming methods.

Regularization techniques reduce the parameter space and avoid overfitting by constraining empirical risks.

Optionally, as an embodiment of the present invention, the normalization process specifically includes: performing normalization processing on the image size, performing normalization processing on the image data feature vector, and performing at least one of moving and scaling the image data. Kind.

In this embodiment, the image size is normalized by the above various normalization methods, or the image data feature vector is normalized, or the image data is moved and scaled. Or, the above technical means can be used in any combination to achieve normalization of image data, ensure the consistency of training data and test data, and improve the recognition accuracy of the image recognition model.

Optionally, as an embodiment of the present invention, the image size normalization process includes: collecting an image exceeding a target area in the process of acquiring image data, and retaining the region of interest in an intermediate position of the image, according to a known sense The size of the area of interest, correcting the image size, removing the redundant area, and retaining the entire content of the area of interest, so that all image data have the same size;

The image data feature vector normalization process includes: performing normalization processing using mean values and covariances of the training data;

Specifically, the formula of the normalization process is as follows.

Where x ^(k) is the eigenvector of a set of training data, E[x ^(k) ] is the mean of all training data, and Var[x ^(k) ] is an unbiased estimate of the variance of all training data,

Is a normalized feature vector;

The moving and scaling processing of the image data includes: scaling the image data according to a scaling factor, and moving the image data by a translation constant;

Specifically, the moving and scaling processing formula is as follows.

among them,

Is the output of the shift and scale process, λ is the scaling factor, the scaling of the sample, multiplying each sample point by the same coefficient, and β is the same factor for each sample point.

As shown in FIG. 2, an embodiment of the present invention provides a method for identifying a target object in an image, including the following steps:

S210: acquiring training images in the training phase, establishing an image recognition model by using the training images, and performing regularization processing when establishing the image recognition model, so that the image recognition model expresses the full sample distribution;

S220: The test image is acquired in the test phase, and the test image is matched with the image recognition model to realize the recognition of the target object in the test image.

Specifically, the regularization process includes converting an empirical risk into a structural risk, specifically adding a regularization term to the structural risk to obtain a structural risk, and obtaining a model parameter that minimizes the structural risk of the image recognition model as an image recognition model. The final model parameter obtains an image recognition model corresponding to the target object.

In the above embodiment, the regularization process is performed when the image recognition model is established in the training phase, the components that can describe the full sample are retained, the components that describe only the training samples are ignored, and the image recognition model expresses the full sample distribution and reduces the parameter space. . Through the above processing, the performance of the model on the test data is as close as possible to the performance on the training data.

Optionally, as an embodiment of the present invention, the empirical risk is converted into a structural risk during the regularization process, and the structural risk is:

θ ^* =arg minR(θ)+λ|θ| (3)

Or, θ ^* = arg minR(θ) + λ(θ) ² (4)

Where equation (3) is L1 regularization, λ|θ| is a regularization term, |θ| is the L1 norm; equation (4) is L2 regularization, λ(θ) ² is a regularization term, (θ) ² Is the L2 norm and λ is a constant.

After the data is normalized, it can compensate for the generalization problem caused by the different distribution of training and test samples to some extent. However, the normalized mean and variance are estimated using training samples and cannot express the full sample distribution. If the model is described as too fine, the overfitting phenomenon still exists.

The idea of regularization is to reduce the parameter space, that is, to preserve the components of the full sample during the training process, ignoring the components that are only describing the training samples. The approach is to modify the risk function to translate empirical risk into structural risk.

θ ^* =arg minR(θ)+λ|θ| (3)

Or, θ ^* = arg minR(θ) + λ(θ) ² (4)

Where (3) is L1 regularization, |θ| is the L1 norm, (4) is L2 regularization, and (θ) ² is the L2 norm. In the embodiment of the present invention, by adding λ|θ|, the uncorrelated input (noise) is suppressed because the weight is not obtained. Adding λ(θ) ² can reduce the weight of each feature vector and make the parameters sparse.

As shown in FIG. 3, an embodiment of the present invention provides a method for identifying a target object in an image, including the following steps:

S310: acquiring training images in the training phase, normalizing the training images, establishing an image recognition model by using the training images, and performing regularization processing when establishing the image recognition model, so that the image recognition model expresses the whole sample distribution;

S320, obtaining a test image in a test phase, normalizing the test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

In the above embodiment, the training image and the test image are normalized, and the training image and the test image are mapped to a unified space, so that the training sample and the test sample are distributed in the same manner, and the image recognition model obtained by using the training data is on the test data. Good test results can be obtained; the regularization process is performed when the image recognition model is established in the training phase, so that the image recognition model expresses the full sample distribution, ignoring the components that only describe the training samples, and the image recognition model expresses the full sample distribution. The parameter space is reduced; the performance of the image recognition model on the test data is made as close as possible to the performance on the training data by the above processing.

A method for identifying a target object in an image according to an embodiment of the present invention is described in detail above with reference to FIG. 1 to FIG. 3, and an image object recognition system in an image according to an embodiment of the present invention is described below with reference to FIG. 4 to FIG. Carry out a detailed description.

As shown in FIG. 4, an image object recognition system in an image provided by an embodiment of the present invention includes

a training module, configured to acquire a training image in a training phase, and use the training image to establish an image recognition model;

a test module for obtaining a test image in a test phase, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

a normalization module for normalizing the training image acquired during the training phase and the test image obtained during the testing phase, and mapping the training image and the test image to a unified space

As shown in FIG. 5, an image object recognition system in an image provided by an embodiment of the present invention includes

The regularization module is used to perform regularization processing when the image recognition model is established in the training phase, so that the image recognition model expresses the full sample distribution.

As shown in FIG. 6, an image object recognition system in an image provided by an embodiment of the present invention includes

a normalization module for normalizing the training image acquired during the training phase and the test image obtained during the testing phase, and mapping the training image and the test image to a unified space;

Regularization module for regularization processing when establishing an image recognition model during the training phase The image recognition model expresses a full sample distribution.

It should be understood that in this embodiment, the identification system may correspond to an execution body of the identification method according to an embodiment of the present invention, and the above-described and other operations and/or functions of the respective modules in the identification system are respectively implemented in order to implement FIGS. 1 to 3 The corresponding processes of each method in the following are not repeated here for brevity.

An embodiment of the present invention provides a food identification method in a refrigerator, which includes collecting image data in a refrigerator as a training image and a test image, and obtaining target attribute information of the food in the image data to be identified according to the target object recognition method in the image according to the above embodiment. .

It should be noted that in this embodiment, in order to obtain a good effect of the food picture recognition in the refrigerator, the pictures in the test stage and the picture quality in the learning stage are similar. In order to achieve image quality consistency, the present invention firstly requires similar conditions on the hardware to ensure that the quality of the photographs taken is as uniform as possible, and the model is tested on the test data by normalization and regularization. Performance as close as possible to the performance of the training data, improve the accuracy of food identification in the refrigerator.

Specifically, the training and testing process is relatively independent, and the entire training process is performed offline on the server side. First, you need to take a picture of each food. The pictures you take need to include as many different scenes as possible, such as background, lighting, size, foreground occlusion, etc.; to ensure that the training image is acquired during the training phase and the test image is taken to have the same shooting environment. Take a good picture of the food and first mark its position in the picture and its type. Finally, for each type of food, there is a large amount of calibrated image data, and a corresponding model is established for each food by a machine learning algorithm.

The purpose of the recognition process is to determine the type of food in the refrigerator by means of image recognition. First, the food picture in the refrigerator is photographed by the camera in the refrigerator, and then the picture is passed to the identified server, and the model is matched with the established model on the server side to obtain the information of the position and type of the food in the picture.

Optionally, as an embodiment of the present invention, the training stage uses the training image to obtain image recognition. The model is performed offline on the server side, and the identification of the attribute information of the food in the test image during the test phase further includes turning on the lighting device in the refrigerator when the door closing signal of the refrigerator door is detected, and adjusting the light intensity of the lighting device to a uniform light intensity; / or defogging the camera before shooting; after the above processing, the shooting conditions are stabilized for a preset time and then captured to obtain image data.

In the above embodiment, in order to ensure the consistency of the test picture, the corresponding processing is first performed from the hardware and the photographing process.

On the hardware, the illumination light is installed at the position of each camera at the same time to ensure that the camera cooperates with other lighting devices in the refrigerator, and secondly, dustproof and fogging measures are added to each camera to prevent the camera from being contaminated. The fog measure can be achieved by adding a dust-proof anti-fog cover. Finally, the same pixel of the photographic device is used to ensure that the captured picture contains the same amount of information.

During the photo taking process, each time the user closes the refrigerator door, a door closing signal will be generated. At this time, the food in the refrigerator may change, and the shooting in the refrigerator is selected at this time. In order to ensure the food lighting conditions in the refrigerator are the same, open the lighting equipment in the refrigerator before shooting, adjust the light attached to the camera to adjust the uniform light intensity by adjusting the duty ratio, and adjust the original lighting equipment in the refrigerator to the approximate light as much as possible. Strong. In addition, in order to prevent condensation of the camera itself due to the temperature difference between the inside and outside of the refrigerator, the anti-fog treatment is performed by heating or the like. After this part of the processing is completed, it is necessary to stabilize the shooting conditions for a certain period of time, to ensure that the delay caused by the shooting process (software logic control to control the shooting delay, shutter speed, etc.) will not affect, and prevent the user from switching multiple times in a short time. Refrigerator door (no photo taken at this time).

An embodiment of the present invention provides a food identification system in a refrigerator, comprising an image collecting device and a server installed in the refrigerator, wherein the image collecting device collects image data in the refrigerator and uploads the image data to the server, wherein the server adopts the foregoing embodiment. The target object recognition system in the image obtains attribute information of the food in the image data to be identified.

It should be noted that the food identification system in the refrigerator guarantees the consistency of the training data and the test data from hardware and/or software, so that the image recognition model performs as much as possible on the test data. The performance on the near training data improves the accuracy of food identification in the refrigerator.

Optionally, as an embodiment of the present invention, as shown in FIG. 7, a food identification system in a refrigerator further includes a lighting device, a refrigerator door detecting device and a control device installed at a position of each camera in the refrigerator, and the refrigerator door detecting When detecting the door closing signal of the refrigerator door, the device sends a door closing signal to the control device, and the control device controls to turn on the lighting device in the refrigerator according to the door closing signal, and adjusts the light intensity of the lighting device to a uniform light intensity.

It should be noted that, in this embodiment, by adjusting the light intensity of the illumination device to a uniform light intensity before photographing, it is ensured that the food illumination conditions in the refrigerator are consistent, and the image consistency is ensured from the photographing environment.

Optionally, as an embodiment of the present invention, a heating device installed at a position of each camera in the refrigerator is configured to perform a defogging process for the camera before the camera performs the shooting, and the control module is further configured to control the shooting condition. Stabilize the preset time and then shoot to obtain image data.

It should be noted that, in this embodiment, a dust-proof and fog-removing device and a heating device are added to each camera to prevent the camera from being contaminated, and the photosensitive device using the same pixel ensures that the captured image contains the same amount of information. In order to prevent the camera itself from causing condensation or the like due to temperature difference between the inside and outside of the refrigerator, the anti-fog treatment is performed by heating or the like. Stabilize the shooting conditions for a period of time, to ensure that the delay caused by the shooting process will not be affected, and to prevent the user from switching the refrigerator door multiple times in a short time.

The term "and/or" in this context is merely an association describing the associated object, indicating that there may be three relationships, for example, A and / or B, which may indicate that A exists separately, and both A and B exist, respectively. B these three situations. In addition, the character "/" in this article generally indicates that the contextual object is an "or" relationship.

Those of ordinary skill in the art will appreciate that the elements and algorithm steps of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both, for clarity of hardware and software. Interchangeability, in accordance with function one in the above description The composition and steps of the examples are generally described. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the solution. A person skilled in the art can use different methods for implementing the described functions for each particular application, but such implementation should not be considered to be beyond the scope of the present invention.

A person skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system, the device and the unit described above can refer to the corresponding process in the foregoing method embodiment, and details are not described herein again.

In the several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods may be implemented in other manners. For example, the device embodiments described above are merely illustrative. For example, the division of cells is only a logical function division. In actual implementation, there may be another division manner. For example, multiple units or components may be combined or integrated. Go to another system, or some features can be ignored or not executed.

The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the embodiments of the present invention.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

An integrated unit, if implemented in the form of a software functional unit and sold or used as a standalone product, can be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention contributes in essence or to the prior art, or all or part of the technical solution may be embodied in the form of a software product stored in a storage medium. Including a number of instructions to make a computer device (which can be a personal computer, The server, or network device, etc.) performs all or part of the steps of the various embodiments of the present invention. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and the like. .

In the description of the present specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like means a specific feature described in connection with the embodiment or example. A structure, material or feature is included in at least one embodiment or example of the invention. In the present specification, the schematic representation of the above terms is not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, various embodiments or examples described in the specification, as well as features of various embodiments or examples, may be combined and combined.

The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalents, improvements, etc., which are within the spirit and scope of the present invention, should be included in the protection of the present invention. Within the scope.

Claims

A method for identifying a target object in an image, comprising the steps of:

Obtaining a training image, and establishing an image recognition model using the training image;

Obtaining a test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

Wherein, the acquired training image and the acquired test image are normalized to map the training image and the test image to a unified space; and/or a regularization process is performed when the image recognition model is established, so that the image recognition model expresses the full sample distribution .
The target object recognition method in the image according to claim 1, wherein the normalization processing comprises: normalizing the image size, normalizing the image data feature vector, and performing image data on the image data. Perform at least one of moving and scaling processing.
The target object recognition method in an image according to claim 2, wherein

The image size normalization process includes: collecting an image exceeding the target area in the process of acquiring the image data, leaving the region of interest in the middle of the image, and correcting the image size according to the size of the known region of interest, and removing the redundancy The remaining area, retaining the entire content of the region of interest, so that all image data has the same size;

The image data feature vector normalization process includes: performing normalization processing using mean values and covariances of the training data;

The moving and scaling processing of the image data includes: scaling the image data according to a scaling factor, and moving the image data by a translation constant.
The target object recognition method in an image according to any one of claims 1 to 3, wherein the image recognition model is established by using the training image, comprising:

Obtaining a predetermined number of image data including the target object;

Calibrating the attribute parameters of the target object in the image data, and calibrating the map Like data as a training image;

Setting an image recognition model, an initial model parameter, and a loss function, and obtaining a risk function by calculating a desired loss function;

The image recognition model is updated by using the risk function and the training data, so that the image recognition model satisfies the requirement of the loss function, and an image recognition model corresponding to the target object is obtained.
The target object recognition method in an image according to claim 4, wherein the risk function is an empirical risk, and the empirical risk is obtained by calculating a desired loss function based on all training data; obtaining an image recognition model by calculation The model parameter with the least risk of experience is used as the final model parameter of the image recognition model, and the image recognition model corresponding to the target object is obtained.
The target object recognition method in an image according to claim 5, further comprising performing a regularization process including converting the empirical risk into a structural risk, specifically adding a regularization term to the empirical risk The structural risk is obtained, and the model parameter that minimizes the structural risk of the image recognition model is obtained as the final model parameter of the image recognition model, and the image recognition model corresponding to the target object is obtained.
The target object recognition method in an image according to claim 4, wherein the attribute parameter of the target object includes a position and a type of the target object in the calibration image data.
The target object recognition method in an image according to any one of claims 1 to 3, wherein the acquisition of the training image and the acquisition of the test image have the same shooting environment.
A food identification method in a refrigerator, characterized in that image data in a refrigerator is collected as a training image and a test image, and the target object recognition method in the image according to any one of claims 1-8 obtains food in the image data to be identified Attribute information.
The food identification method in the refrigerator according to claim 9, wherein the image recognition model obtained by using the training image is performed offline on the server side, and identifying the attribute information of the food in the test image further comprises: opening the refrigerator when detecting the door closing signal of the refrigerator door Lighting equipment inside, and adjust the light intensity of the lighting equipment to a uniform light intensity; and/or defogging the camera before shooting; After the above processing, the shooting conditions are stabilized for a preset time and then captured to acquire image data.
An image recognition system for an object in an image, comprising:

a training module, configured to acquire a training image, and use the training image to establish an image recognition model;

a test module for acquiring a test image, matching the test image with the image recognition model, and realizing recognition of the target object in the test image;

a normalization module, configured to normalize the acquired training image and the acquired test image, and map the training image and the test image to a unified space;

A regularization module is used to perform regularization processing when the image recognition model is established, so that the image recognition model expresses a full sample distribution.
The image object recognition system according to claim 11, wherein the normalization module comprises a size normalization unit, a feature normalization unit, and a shift scaling unit; and the size normalization unit is used for image size. In the normalization process, the feature normalization unit is used for normalizing the image data feature vector, and the shift scaling unit is configured to perform image shifting and scaling processing.
The target object recognition system in an image according to claim 11, wherein the training module comprises:

a training image acquiring unit, configured to acquire a predetermined number of image data including the target object;

An image calibration unit, configured to calibrate an attribute parameter of the target object in the image data, and use the image data of the calibration process as a training image;

a model setting module for setting an image recognition model, an initial model parameter, and a loss function, and obtaining a risk function by calculating a desired loss function;

The model training unit updates the image recognition model with the risk function and the training data, so that the image recognition model satisfies the requirement of the loss function, and obtains an image recognition model corresponding to the target object.
A food identification system in a refrigerator, comprising: a figure installed inside a refrigerator For example, the image capturing device collects the image data in the refrigerator and uploads it to the server, and the server obtains the food in the image data to be identified by using the target object recognition system in the image according to any one of claims 11-13. Attribute information.
The food identification system in an refrigerator according to claim 14, further comprising a lighting device, a refrigerator door detecting device and a control device installed at a position of each camera in the refrigerator, wherein the refrigerator door detecting device detects that the refrigerator door is closed When the signal is sent, the door closing signal is sent to the control device, and the control device controls the lighting device in the refrigerator according to the door closing signal, and adjusts the light intensity of the lighting device to a uniform light intensity.
The food identification system in a refrigerator according to claim 15, further comprising heating means installed at a position of each camera in the refrigerator for performing a defogging process for the camera before the camera performs the shooting, the control module It is also used to control the shooting conditions to stabilize the preset time and then perform shooting to acquire image data.