CN111523483B - Chinese meal dish image recognition method and device - Google Patents

Abstract

Embodiments of the present invention provide a Chinese food dish image recognition method and device. The method includes: acquiring a target Chinese food dish image, performing a preprocessing operation on the target Chinese food dish image; and inputting the preprocessed target Chinese food dish image. to the Chinese food dish image recognition model to obtain Chinese food dish recognition results; wherein, the Chinese food dish image recognition model is obtained based on preprocessed Chinese food dish image samples and corresponding Chinese food dish category label training, and the Chinese food dish image recognition model The model is built based on the DenseNet model. The network structure of the Chinese food dish image recognition model includes: N dense connection blocks for feature reuse and N-1 transition layers for compressing the number of parameters; N is a natural number greater than 1 . Embodiments of the present invention can accurately detect and identify a variety of Chinese food, with a wide range of identification types and high identification accuracy.

Description

Chinese food dish image recognition method and device

Technical field

The present invention relates to the field of computer technology, and more specifically, to a Chinese food dish image recognition method and device.

Background technique

With the rapid development of deep learning algorithms, computer vision has become the fastest growing and most widely implemented field of artificial intelligence, and has been widely used in all aspects of people's lives. Among them, food recognition is an emerging field that has attracted much attention in the field of computer vision. topic.

There are currently many studies on recognition algorithms for Western and Japanese dishes, but there are not many studies on more mature methods for image recognition of Chinese dishes. This is not only because there are very few public large-scale Chinese dish classification data sets, but also because Chinese dishes are compared with Western or Japanese dishes. Chinese dishes are more difficult to identify because the same category of Chinese dishes may appear in a variety of different forms. At the same time, the images of Chinese food dishes are also affected by background noise such as the color of the plate and the brightness of the light; in addition, different Chinese food dishes may look very similar.

For these reasons, the existing technologies that can accurately identify Chinese food dishes are currently very limited. These situations have increased the difficulty of accurately identifying Chinese food dish images. Therefore, a method that can accurately detect and identify Chinese food dishes is needed.

Contents of the invention

In order to solve or at least partially solve the above problems, embodiments of the present invention provide a Chinese food dish image recognition method and device.

In a first aspect, an embodiment of the present invention provides a Chinese food dish image recognition method, including:

Obtain the target Chinese food dish image, and perform a preprocessing operation on the target Chinese food dish image;

Input the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the Chinese food dish recognition result;

Wherein, the Chinese food dish image recognition model is obtained by training based on preprocessed Chinese food dish image samples and corresponding Chinese food category labels. The Chinese food dish image recognition model is constructed based on the DenseNet model. The Chinese food dish image recognition model is The network structure includes: N densely connected blocks for feature reuse and N-1 transition layers for compressing the number of parameters; N is a natural number greater than 1.

Optionally, the step of inputting the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the recognition result specifically includes:

The preprocessed target Chinese food dish image is input into the Chinese food dish image recognition model, and through the operations of the first convolution layer, the first batch normalization layer and the excitation layer of the Chinese food dish image recognition model, the first feature map;

Input the first feature map into the maximum pooling layer of the Chinese food dish image recognition model to obtain a second feature map;

Input the second feature map into the first dense connection block of the Chinese food dish image recognition model, and then obtain a third feature map through the operation of the first transition layer;

Input the third feature map into the second dense connection block of the Chinese food dish image recognition model, and then obtain a fourth feature map through the operation of the second transition layer;

The fourth feature map is input into the third dense connection block of the Chinese food dish image recognition model, and then through the operation of the third transition layer, a fifth feature map is obtained;

Input the fifth feature map into the fourth dense connection block of the Chinese food dish image recognition model, and then obtain the sixth feature map through the operation of the fourth transition layer;

The sixth feature map is input into the second batch normalization layer of the Chinese food image recognition model, and then through the operation of the fully connected layer and the classifier, the Chinese food recognition result is obtained.

Optionally, the first densely connected block, the second densely connected block, the third densely connected block and the fourth densely connected block each include a plurality of densely connected bottleneck layers, and each bottleneck layer has a layer containing multiple densely connected blocks. A composite function of multiple operations including: batch normalized BN, ReLU activation function and 3×3 convolution.

Optionally, the various operations also include: 1×1 convolution.

Optionally, the first transition layer, the second transition layer, the third transition layer and the fourth transition layer all perform the following operations: batch normalized BN, ReLU activation function, 1×1 convolution and 2×2 averaging. Pooling, step size is 2.

Optionally, before the steps of obtaining the target Chinese food dish image and performing a preprocessing operation on the target Chinese food dish image, the method further includes:

Construct a DenseNet model, which includes a first convolution layer, a first batch normalization layer, an excitation layer, a maximum pooling layer, a first dense connection block, a first transition layer, and a second dense connection block connected in sequence , the second transition layer, the third dense connection block, the third transition layer, the fourth dense connection block, the second batch normalization layer, the fully connected layer and the classifier;

Obtain Chinese food dish image samples, and preprocess the Chinese food dish image samples;

Input the preprocessed Chinese food dish image samples to the DenseNet model to obtain the output results;

Based on the output result and the Chinese food category label corresponding to the Chinese food image sample, use the cross entropy loss function to calculate the loss function value;

Based on the Adam optimization algorithm, adjust various parameters of the densely connected convolutional neural network starting from the output layer of the DenseNet model so that the loss function value moves in the direction of minimization;

Determine whether the training end condition is reached. If so, save the parameters of the DenseNet model in the current iteration to obtain the trained Chinese food dish image recognition model.

Optionally, perform a preprocessing operation on the target Chinese food dish image, specifically:

Perform random center rotation on the target Chinese food dish image according to a preset angle;

Randomly crop the image of the target Chinese food dish after random center rotation according to a preset aspect ratio;

Horizontally flipping the randomly cropped image of the target Chinese food dish according to a preset probability;

The horizontally flipped image of the target Chinese food dish is normalized.

In a second aspect, an embodiment of the present invention provides a Chinese food dish image recognition device, including:

A preprocessing module, used to obtain a target Chinese food dish image and perform a preprocessing operation on the target Chinese food dish image;

A recognition module, configured to input the preprocessed image of the target Chinese food dish into the Chinese food dish image recognition model to obtain a Chinese food dish recognition result;

Wherein, the Chinese food dish image recognition model is obtained based on training of preprocessed Chinese food dish image samples. The Chinese food dish image recognition model is constructed based on the DenseNet model. The network structure of the Chinese food dish image recognition model includes: N users Dense connection blocks for feature reuse and N-1 transition layers for compressing the number of parameters; N is a natural number greater than 1.

In a third aspect, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps described in the first aspect are implemented. The steps of the Chinese food dish image recognition method are provided.

In a fourth aspect, embodiments of the present invention provide a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the Chinese food dish image recognition method provided in the first aspect are implemented. .

The Chinese food dish image recognition method and device provided by embodiments of the present invention can accurately detect and identify a variety of Chinese food, with a wide range of recognition types and high recognition accuracy.

Description of the drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic flow chart of a Chinese food dish image recognition method provided by an embodiment of the present invention;

Figure 2 is a schematic network structure diagram of the Chinese food dish image recognition model provided by an embodiment of the present invention;

Figure 3 is a schematic structural diagram of the dense block dense block;

Figure 4 is a schematic structural diagram of the bottleneck layer;

Figure 5 is a schematic structural diagram of a Chinese food dish image recognition device provided by an embodiment of the present invention.

FIG. 6 is a schematic diagram of the physical structure of an electronic device according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

Figure 1 is a schematic flowchart of a Chinese food dish image recognition method provided by an embodiment of the present invention, including:

Step 100: Obtain the target Chinese food dish image, and perform a preprocessing operation on the target Chinese food dish image;

Specifically, in the embodiment of the present invention, a fixed-position camera is used to collect a single target Chinese food dish image, and then a preprocessing operation is performed on the target Chinese food dish image, and the preprocessing operation includes a data enhancement operation. Commonly used basic data enhancement operations include the following methods: rotation, translation, scaling, random occlusion, horizontal flipping, color aberration and noise disturbance, etc. Some of these data enhancement methods can be selected to perform preprocessing operations on the target Chinese food dish image.

Step 101: Input the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the Chinese food dish recognition result;

Specifically, in the embodiment of the present invention, the target Chinese food dish image obtained through the above preprocessing operation is input into the pre-trained Chinese food dish image recognition model, and the Chinese food dish recognition result can be obtained.

Wherein, the Chinese food dish image recognition model is obtained by training based on preprocessed Chinese food dish image samples and corresponding Chinese food dish category labels.

Compared with general food images, Chinese food dish images generally do not show unique spatial layout and obvious semantic features like most Western food, making it more difficult to extract the semantic information of Chinese food dish images. Therefore, in the embodiment of the present invention, the Chinese food dish image recognition model is built based on the DenseNet model, because the DenseNet model does not simply obtain representation capabilities through a very deep or very wide network, but through the repeated use of low-level features to high-level features , combining and connecting the features of different layers, increasing the diversity of subsequent layer inputs and achieving the ultimate utilization of image features. And compared with other networks, the DenseNet model has fewer parameters, which prevents gradient disappearance and reduces overfitting on small sample data sets, making it simpler and more efficient.

Furthermore, based on the DenseNet network model, the network structure of the Chinese food dish image recognition model includes: N dense connection blocks for feature reuse and N-1 transition layers for compressing the number of parameters.

Different from other convolutional neural networks, the present invention uses dense connection methods to realize feature reuse and utilize image features to the extreme, which can better extract semantic information of image pairs and achieve accurate recognition with greater probability. The dense connection block is used to alleviate gradient disappearance, reduce training parameters, resist overfitting and implement feature reuse. The transition layer is used to compress the number of parameters and reduce model complexity caused by the introduction of the dense connection block. question.

The Chinese food dish image recognition method provided by the embodiment of the present invention can accurately detect and identify a variety of Chinese food, with a wide range of recognition types and high recognition accuracy.

Based on the content of the above embodiment, the step of inputting the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the recognition result specifically includes:

The preprocessed target Chinese food dish image is input into the Chinese food dish image recognition model, and through the operations of the first convolution layer, the first batch normalization layer and the excitation layer of the Chinese food dish image recognition model, the first feature map;

Input the first feature map into the maximum pooling layer of the Chinese food dish image recognition model to obtain a second feature map;

Input the second feature map into the first dense connection block of the Chinese food dish image recognition model, and then obtain a third feature map through the operation of the first transition layer;

Input the third feature map into the second dense connection block of the Chinese food dish image recognition model, and then obtain a fourth feature map through the operation of the second transition layer;

The fourth feature map is input into the third dense connection block of the Chinese food dish image recognition model, and then through the operation of the third transition layer, a fifth feature map is obtained;

Input the fifth feature map into the fourth dense connection block of the Chinese food dish image recognition model, and then obtain the sixth feature map through the operation of the fourth transition layer;

The sixth feature map is input into the second batch normalization layer of the Chinese food image recognition model, and then through the operation of the fully connected layer and the classifier, the Chinese food recognition result is obtained.

Figure 2 is a schematic network structure diagram of a Chinese food dish image recognition model provided by an embodiment of the present invention. The Chinese food dish image recognition model includes a first convolution layer, a first batch normalization layer, an excitation layer, and a maximum pooling layer connected in sequence. layer, the first densely connected block, the first transition layer, the second densely connected block, the second transitional layer, the third densely connected block, the third transitional layer, the fourth densely connected block, the second batch normalization layer, the full Connect layers and classifiers.

Specifically, after preprocessing, the target Chinese food dish image is input into the Chinese food dish image recognition model, and undergoes the convolution operation of the first convolution layer, the BN operation of the first batch normalization layer, and the RELU activation function of the excitation layer. After the operation, the dimensionality is reduced to obtain the first feature map, and then the first feature map is input into the maximum pooling layer. The maximum pooling layer is to downsample the feature map and remove unnecessary redundant information in the map, and obtain The second feature map passes through four dense blocks in sequence, and there is a transition layer between each dense block layer.

In a specific embodiment, the target Chinese food dish image with pixels of 224×224 is first subjected to the convolution, BN and ReLU operations in Figure 2 in order to achieve dimensionality reduction, and a first feature map with pixels of 112×112 is obtained. picture. Then the first feature map is input into the maximum pooling layer, which uses 3×3 convolution with a stride of 2. The second feature map with a pixel size of 56×56 is obtained as the input of the first dense block dense block.

Figure 3 is a schematic structural diagram of the dense block dense block. One layer in the dense block is called the bottleneck layer. What makes DenseNet superior to other convolutional neural networks is the dense block. With dense block, DenseNet has the advantages of mitigating gradient disappearance, parameter reduction, anti-overfitting and feature reuse.

Assume that a dense block has l layers, and x ₀ is the input of the dense block. Each layer has a composite function H _l (·) that contains three operations: BN, ReLU and 3×3 convolution. In order to better improve the information transfer between denseblocks, DenseNet proposes a unique connection method: dense connection. Dense connection connects each layer in a dense block to all subsequent layers to achieve feature reuse, as shown in Figure 3. Therefore, the lth layer takes the feature maps x ₀ ,...,x _l-1 of all previous layers as input:

x _l =H _l ([x ₀ ,x ₁ ,...,x _l-1 ])

Among them, [x ₀ , x ₁ ,..., x _l-1 ] indicates that the feature map output by the 0,..., l-1 layer will be used as the input of the l-th layer after combined connection.

Optionally, each bottleneck layer has a composite function including multiple operations, including: batch normalized BN, ReLU activation function and 3×3 convolution.

Figure 4 is a schematic structural diagram of the bottleneck layer. Considering that the number of feature maps will be large after using dense connections, in order to reduce the number of feature maps and reduce the dimension of each feature map, a 1×1 volume was added before the 3×3 convolution of the bottleneck layer. Product can reduce the amount of calculation.

Further, the first transition layer, the second transition layer, the third transition layer and the fourth transition layer all perform the following operations: batch normalized BN, ReLU activation function, 1×1 convolution and 2×2 average pooling , the step size is 2. Its function is to further compress the number of parameters. The dimensions and number of channels of the output feature map of each dense block will increase dramatically. The convolution operation of the transition layer can reduce the dimensionality of the feature map and average pooling can solve the feature map. The problem of too many channels in the graph prevents the model from becoming complicated after too many dense blocks.

If m feature maps are generated through a dense block, θm feature maps are generated after a transition layer, where θ is the compression coefficient, and 0<θ≤1. When θ=1, the number of feature maps passing through the transition layer remains unchanged. In the embodiment of the present invention, θ=0.5 is set, and the number of feature maps is reduced by half after passing through the transition layer.

In a specific embodiment, the pixels of the feature map after four dense blocks are 56×56, 28×28, 14×14, and 7×7 respectively. After the last dense block, BN and softmax classifiers are used, and the output of the fully connected layer is set to the total number of types of Chinese food dishes.

Before using the trained Chinese food image recognition model to recognize the target Chinese food image, the Chinese food image recognition model also needs to be trained.

Based on the contents of the above embodiments, before the steps of obtaining the target Chinese food dish image and performing the preprocessing operation on the target Chinese food dish image, it also includes:

Step 200: Construct a DenseNet model. The DenseNet model includes a first convolution layer, a first batch normalization layer, an excitation layer, a maximum pooling layer, a first dense connection block, a first transition layer, a second Densely connected block, second transition layer, third densely connected block, third transition layer, fourth densely connected block, second batch normalization layer, fully connected layer and classifier;

Specifically, the DenseNet model in this example is an improved DenseNet 169 model, with a network structure as shown in Figure 3.

Step 201: Obtain Chinese food dish image samples, and preprocess the Chinese food dish image samples;

The purpose of preprocessing is to achieve image enhancement.

Step 202: Input the preprocessed Chinese food dish image samples to the DenseNet model to obtain an output result;

Step 203: Based on the output result and the Chinese food category label corresponding to the Chinese food image sample, use the cross entropy loss function to calculate the loss function value;

The loss function uses a cross-entropy model to speed up the convergence speed and the update speed of the weight matrix.

Step 204: Based on the Adam optimization algorithm, adjust various parameters of the densely connected convolutional neural network starting from the output layer of the DenseNet model, so that the loss function value moves in the direction of minimization;

The optimizer in the training model uses the Adam algorithm to achieve adaptive learning rate, speed up training, and enhance the robustness of the network.

Step 205: Determine whether the training end condition is met. If so, save the parameters of the DenseNet model in the current iteration to obtain the trained Chinese food dish image recognition model.

Specifically, a fixed-position camera is used to collect multiple single-dish images and save them in the database, and add a category label to each image. If there is no image of this category in the database, add a category label to create a new category, and divide the database into proportions. are the training set and test set. During training, in order to make the model have better classification performance on the data set, the network parameters are adjusted as follows: epoch is set to 150; the batch size is 64; the optimizer selects Adam, which can provide an adaptive learning rate. The initial The learning rate is 1e-4, which greatly improves the training speed and enhances the robustness of the network; because this invention is aimed at classification problems, the loss function adopts the cross-entropy model, which can realize that the learning rate will be accelerated when the model convergence effect is poor. When the effect is good, the learning rate slows down. After 150 epochs of DenseNet169, the optimal model was selected as the final trained Chinese food dish image recognition model. After training, testing can also be performed. During testing, the test set is input into the optimal model for testing, and the test results can be obtained.

The Chinese food dish image recognition method provided by the embodiment of the present invention makes full use of the advantages of DenseNet network dense connection mode to achieve feature reuse. The adjustment of paired network super parameters not only greatly reduces the number of training parameters and the redundancy of the training network, but also makes the dishes The image features are fully utilized, which is beneficial to capturing the semantic information of the dish images. After multiple iterative trainings, a training model with high recognition accuracy and excellent performance can be obtained. Because DenseNet has strong generalization ability, the present invention is not only suitable for identifying difficult Chinese food. In principle, it can be applied to the identification of more types of food as long as it is trained on other categories of food data sets.

Based on the content of the above embodiment, a preprocessing operation is performed on the target Chinese food dish image, specifically as follows:

Perform random center rotation on the target Chinese food dish image according to a preset angle;

Randomly crop the image of the target Chinese food dish after random center rotation according to a preset aspect ratio;

Horizontally flipping the randomly cropped image of the target Chinese food dish according to a preset probability;

The horizontally flipped image of the target Chinese food dish is normalized.

Specifically, perform random center rotation on the target Chinese food dish image according to a preset angle, for example, between -10 degrees and 10 degrees;

Randomly crop the image of the target Chinese food dish after random center rotation according to a preset aspect ratio, such as an aspect ratio of 224×224;

According to a preset probability, such as a probability of 0.5, horizontally flip the randomly cropped image of the target Chinese food dish;

The horizontally flipped image of the target Chinese food dish is normalized to eliminate the dimensional influence between data features.

The preprocessing steps provided by the embodiments of the present invention are conducive to obtaining accurate training models and Chinese food dish recognition results.

Figure 5 is a schematic structural diagram of a Chinese food dish image recognition device provided by an embodiment of the present invention, including: a preprocessing module 510 and a recognition module 520, wherein,

The preprocessing module 510 is used to obtain the target Chinese food dish image and perform preprocessing operations on the target Chinese food dish image;

The recognition module 520 is used to input the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the Chinese food dish recognition result;

Wherein, the Chinese food dish image recognition model is obtained by training based on preprocessed Chinese food dish image samples and corresponding Chinese food category labels. The Chinese food dish image recognition model is constructed based on the DenseNet model. The Chinese food dish image recognition model is The network structure includes: N densely connected blocks for feature reuse and N-1 transition layers for compressing the number of parameters; N is a natural number greater than 1.

The Chinese food dish image recognition device provided by the embodiment of the present invention is used to implement the foregoing Chinese food dish image recognition method embodiment. Therefore, for understanding of each functional module in the embodiment of the present invention, refer to the foregoing method embodiment and will not be described again here.

The Chinese food dish image recognition device provided by the embodiment of the present invention can accurately detect and identify a variety of Chinese food, with a wide range of recognition types and high recognition accuracy.

Figure 6 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. As shown in Figure 6, the electronic device may include: a processor (processor) 610, a communications interface (Communications Interface) 620, a memory (memory) 630 and a communication interface. Bus 640, in which the processor 610, the communication interface 620, and the memory 630 complete communication with each other through the communication bus 640. The processor 610 can call a computer program stored on the memory 630 and executable on the processor 610 to execute the Chinese food dish image recognition method provided by each of the above method embodiments, for example, including: acquiring the target Chinese food dish image, and The target Chinese food dish image performs a preprocessing operation; the preprocessed target Chinese food dish image is input into the Chinese food dish image recognition model to obtain the Chinese food dish recognition result; wherein the Chinese food dish image recognition model is based on the preprocessed Chinese food dish image samples and corresponding Chinese food dish category labels are obtained through training. The Chinese food dish image recognition model is built based on the DenseNet model. The network structure of the Chinese food dish image recognition model includes: N dense connections for feature reuse. block and N-1 transition layers used to compress the number of parameters; N is a natural number greater than 1.

In addition, the above-mentioned logical instructions in the memory 630 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the embodiment of the present invention is essentially, or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, It includes several instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

Embodiments of the present invention also provide a non-transitory computer-readable storage medium on which a computer program is stored. When executed by a processor, the computer program implements the Chinese food dish image recognition method provided by the above method embodiments. For example, it includes: obtaining target Chinese food dish image, perform a preprocessing operation on the target Chinese food dish image; input the preprocessed target Chinese food dish image into the Chinese food dish image recognition model to obtain the Chinese food dish recognition result; wherein, the Chinese food dish image The recognition model is obtained based on preprocessed Chinese food image samples and corresponding Chinese food category labels. The Chinese food image recognition model is built based on the DenseNet model. The network structure of the Chinese food image recognition model includes: N users Dense connection blocks for feature reuse and N-1 transition layers for compressing the number of parameters; N is a natural number greater than 1.

The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., including a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. The Chinese meal image recognition method is characterized by comprising the following steps of:

acquiring a target Chinese food image, and performing preprocessing operation on the target Chinese food image;

inputting the preprocessed target Chinese food image into a Chinese food image recognition model to obtain a Chinese food recognition result;

the Chinese meal image recognition model is obtained based on a preprocessed Chinese meal image sample and a corresponding Chinese meal category label, and is constructed based on a DenseNet model, and the network structure of the Chinese meal image recognition model comprises: n dense connecting blocks for realizing feature multiplexing and N-1 transition layers for compressing parameter quantity; n is a natural number greater than 1;

the step of inputting the preprocessed target Chinese food image into a Chinese food image recognition model to obtain a recognition result specifically comprises the following steps:

inputting the preprocessed target Chinese food image into a Chinese food image recognition model, and obtaining a first characteristic map through the operation of a first convolution layer, a first batch normalization layer and an excitation layer of the Chinese food image recognition model;

inputting the first characteristic map to a maximum pooling layer of the Chinese dish image recognition model to obtain a second characteristic map;

inputting the second characteristic map to a first intensive connection block of the Chinese dish image recognition model, and then obtaining a third characteristic map through the operation of a first transition layer;

inputting the third characteristic map to a second intensive connection block of the Chinese dish image recognition model, and then obtaining a fourth characteristic map through the operation of a second transition layer;

inputting the fourth characteristic map to a third intensive connection block of the Chinese meal dish image recognition model, and then obtaining a fifth characteristic map through the operation of a third transition layer;

inputting the fifth characteristic map to a fourth intensive connection block of the Chinese dish image recognition model to obtain a sixth characteristic map;

inputting the sixth characteristic map to a second batch normalization layer of the Chinese meal dish image recognition model, and then obtaining a Chinese meal dish recognition result through the operation of a full connection layer and a classifier;

the preprocessing operation is executed on the target Chinese food image, specifically:

randomly and centrally rotating the target Chinese food image according to a preset angle;

randomly cutting the target Chinese food image after random center rotation according to a preset length-width ratio;

performing horizontal overturning on the target Chinese food image subjected to random clipping according to preset probability;

and normalizing the target Chinese food image subjected to horizontal overturning.

2. The method of claim 1, wherein the first, second, third, and fourth densely connected blocks each comprise a plurality of densely connected bottleneck layers, each bottleneck layer having a composite function comprising a plurality of operations, the plurality of operations comprising: batch normalization BN, reLU activation function and 3 x 3 convolution.

3. The method of claim 2, wherein the plurality of operations further comprises: 1 x 1 convolution.

4. The method of claim 1, wherein the first transition layer, the second transition layer, and the third transition layer each perform the following operations: batch normalization BN, reLU activation function, 1×1 convolution and 2×2 average pooling, step size 2.

5. The method of claim 1, further comprising, prior to the step of obtaining the target chinese meal image and performing a preprocessing operation on the target chinese meal image:

constructing a DenseNet model, wherein the DenseNet model comprises a first convolution layer, a first batch normalization layer, an excitation layer, a maximum pooling layer, a first intensive connection block, a first transition layer, a second intensive connection block, a second transition layer, a third intensive connection block, a third transition layer, a fourth intensive connection block, a second batch normalization layer, a full connection layer and a classifier which are sequentially connected;

acquiring a Chinese meal dish image sample, and preprocessing the Chinese meal dish image sample;

inputting the preprocessed Chinese meal image sample into the DenseNet model to obtain an output result;

calculating a loss function value by using a cross entropy loss function based on the output result and a Chinese food category label corresponding to the Chinese food image sample;

based on an Adam optimization algorithm, starting from an output layer of the DenseNet model, adjusting various parameters of the densely connected convolutional neural network so as to move the loss function value towards a minimizing direction;

judging whether the training ending condition is reached, if so, saving the parameters of the DenseNet model of the current iteration to obtain a training-completed Chinese dish image recognition model.

6. A Chinese meal image recognition device, comprising:

the preprocessing module is used for acquiring a target Chinese food image and executing preprocessing operation on the target Chinese food image;

the identification module is used for inputting the preprocessed target Chinese food image into a Chinese food image identification model to obtain a Chinese food identification result;

the Chinese meal image recognition model is obtained based on a preprocessed Chinese meal image sample and a corresponding Chinese meal category label, and is constructed based on a DenseNet model, and the network structure of the Chinese meal image recognition model comprises: n dense connecting blocks for realizing feature multiplexing and N-1 transition layers for compressing parameter quantity; n is a natural number greater than 1;

the step of inputting the preprocessed target Chinese food image into a Chinese food image recognition model to obtain a recognition result specifically comprises the following steps:

inputting the preprocessed target Chinese food image into a Chinese food image recognition model, and obtaining a first characteristic map through the operation of a first convolution layer, a first batch normalization layer and an excitation layer of the Chinese food image recognition model;

inputting the first characteristic map to a maximum pooling layer of the Chinese dish image recognition model to obtain a second characteristic map;

inputting the second characteristic map to a first intensive connection block of the Chinese dish image recognition model, and then obtaining a third characteristic map through the operation of a first transition layer;

inputting the third characteristic map to a second intensive connection block of the Chinese dish image recognition model, and then obtaining a fourth characteristic map through the operation of a second transition layer;

inputting the fourth characteristic map to a third intensive connection block of the Chinese meal dish image recognition model, and then obtaining a fifth characteristic map through the operation of a third transition layer;

inputting the fifth characteristic map to a fourth intensive connection block of the Chinese dish image recognition model to obtain a sixth characteristic map;

inputting the sixth characteristic map to a second batch normalization layer of the Chinese meal dish image recognition model, and then obtaining a Chinese meal dish recognition result through the operation of a full connection layer and a classifier;

the preprocessing operation is executed on the target Chinese food image, specifically:

randomly and centrally rotating the target Chinese food image according to a preset angle;

randomly cutting the target Chinese food image after random center rotation according to a preset length-width ratio;

performing horizontal overturning on the target Chinese food image subjected to random clipping according to preset probability;

and normalizing the target Chinese food image subjected to horizontal overturning.

7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor performs the steps of the method for identifying a chinese meal image according to any one of claims 1 to 5 when the program is executed.

8. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the chinese meal order image recognition method of any one of claims 1 to 5.