s3, carrying out first iterative training on the coding-decoding model by taking the illegal picture of the public map set A and the labeled file in the public illegal region set a as training samples, and then carrying out second iterative training on the coding-decoding model by taking the labeled file in the real map set B and the labeled file in the real illegal region set B as training samples to obtain a kitchen illegal behavior detection model;

it should be emphasized that, in order to further ensure the privacy and security of the detection model data information, the detection model data information may also be stored in a node of a block chain.

In one embodiment, the step S3, before the first iterative training and the second iterative training, further includes:

After the first iterative training and the second iterative training in step S3, the method further includes:

and decoding the m-dimensional vector by using the decoder to obtain the ID of the picture.

In this embodiment, conversion between input and output of the encoding-decoding model is realized.

The cross entropy can be used as a loss function in a neural network (machine learning), p represents the distribution of real marks, q is the distribution of predicted marks of the trained model, and the cross entropy loss function can measure the similarity between p and q. The cross entropy as the loss function has the advantage that the problem of the learning rate reduction of the mean square error loss function can be avoided when the gradient is reduced by using the sigmoid function, because the learning rate can be controlled by the output error. In feature engineering, it can be used to measure the similarity between two random variables.

In the training stage, a training picture and a corresponding marking file are input into a model, the model extracts features through a multilayer convolution network, the position of a target is located according to feature matching, and then the violation type is obtained through a classification network. The training process of the model is driven by data, and final model parameters are obtained by minimizing a cross entropic loss function without manually adjusting the parameters. The number of layers of the model exceeds 50 layers, deep information in the image can be extracted, the generalization capability of the model in detection is stronger, and parameters of the model can be continuously optimized along with the continuous increase of training samples, so that the detection accuracy is improved.

In one embodiment, the first iterative training in step S3 includes the following steps:

The second iterative training in step S3 includes the same steps as described above, and the training samples used mainly differ between the first iterative training and the second iterative training, where the former uses public data and the latter uses actual scene data.

The method comprises the following steps that a cross-over ratio iterative lifting method is adopted in feature matching training, and the accuracy of the position of a prediction frame is improved;

the intersection ratio (IOU) is the coincidence degree of the prediction frame and the target object, and the specific calculation formula of the IOU is as follows:

wherein, S1 is the area of the intersection of the prediction box boundary and the actual boundary, and S2 is the entire area of the prediction box.

The current method for determining the violation is that the prediction is correct when the cross-over ratio is larger than a certain set threshold, and generally, the detection result is correct when the cross-over ratio is larger than 0.5. In the step, a method for improving the prediction accuracy by iteratively improving the cross-over ratio is adopted, wherein a threshold set R = { R1, R2, …, Rn } is preset in the method, wherein n represents the number of thresholds, R1< R2< … < Rn, and the size of the thresholds in the threshold set R and the number n of the thresholds are determined by effects required by experiments. In one embodiment, the threshold set R = {0.3, 0.4, 0.5, 0.6}, and the specific feature matching training steps are as follows:

when the intersection ratio IOU1 is smaller than a first threshold value 0.3 in a preset threshold value set R, judging that the prediction is wrong, and finishing training;

when the intersection ratio IOU1 is greater than the first threshold value 0.3, comparing the intersection ratio IOU1 with a second threshold value 0.4 in the preset threshold value set R;

when the intersection ratio IOU1 is less than the second threshold value 0.4, judging that the prediction is wrong;

and when the cross-over ratio IOU1 is greater than the second threshold value 0.4, taking the result of the original prediction interval [ R1, R2] as a negative sample, balancing the positive and negative samples for retraining, calculating the cross-over ratio IOU2 again, continuously comparing with the next threshold value 0.5 in the preset threshold value set R, iteratively improving the cross-over ratio, and finishing the feature matching training if the prediction is wrong or all the threshold values in the preset threshold value set R are compared.

In the embodiment, a position of a relatively higher intersection is obtained than a position of a corresponding prediction frame, so that the prediction frame can be more accurately matched with the position area corresponding to the feature.

And the accuracy of model classification is measured by using cross entropy in violation classification training, the positioning error of the model is measured by using L1 norm, the result output by each model is compared with the result manually marked to obtain an error, parameters are corrected according to the error, and model training is completed after multiple iterations until the error is less than a set threshold value, and the final model parameters are stored.

Taking the illegal picture and the annotation file as samples, training the coding-decoding model by using a random gradient descent method until a cross entropy loss function between input data and output data of the coding-decoding model converges to a first threshold, wherein the first threshold is preferably 0.001. Wherein the cross entropy loss function is specifically as follows:

wherein,

as weights, dependent on the pixel point

If the pixel point is

Is located in the violation area of the corresponding picture, then

＝

，0.6≤

1 or less, otherwise

＝1-

；

For pixel points in illegal pictures

Image ofThe prime value;

for pixel points in output result of coding-decoding model with the illegal picture as input

The pixel value of (2).

The violation classification capability of the detection model is trained, so that the judgment of the violation type of the image by the model is more accurate.

S4, acquiring an image shot by the kitchen camera in real time, and inputting the image into a kitchen violation detection model for violation detection;

in the using stage, as long as the detection model of the kitchen violation behavior is loaded into the network framework, and the picture shot by the kitchen camera is input, the model can perform feature extraction, positioning and classification on the picture, and the violation behavior is detected.

According to the embodiment, the behavior action of the worker can be conveniently and accurately captured in the kitchen scene, and whether violation is caused or not is judged without manual supervision.

Fig. 3 is a structural diagram of a kitchen violation detection device in an embodiment of the present application, and as shown in fig. 3, a kitchen violation detection device includes the following modules:

the samplelibrary construction module 10 is used for acquiring a kitchen violation picture disclosed on a network, constructing a public atlas A, acquiring a kitchen violation picture captured by a camera in an actual kitchen scene, constructing a real atlas B, constructing a public violation area set a by using the violation picture in the public atlas A, and constructing a real violation area set B by using the violation picture in the real atlas B;

amodel initialization module 20, configured to construct a convolutional neural network structure-based encoding-decoding detection model;

and themodel training module 30 is configured to perform first iterative training on the coding-decoding model by using the illegal picture in the public map set a and the labeled file in the public illegal region set a as training samples, and perform second iterative training on the coding-decoding model by using the labeled file in the real map set B and the labeled file in the real illegal region set B as training samples, so as to obtain a kitchen illegal behavior detection model.

And theviolation detection module 40 is used for acquiring an image shot by the kitchen camera in real time and inputting the image into the kitchen violation detection model for violation detection.

Wherein the memory has stored therein computer readable instructions that, when executed by the processor, cause the processor to perform the steps of the above-described method of kitchen violation detection.

In one embodiment, a storage medium storing computer-readable instructions is provided, the computer-usable storage medium may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function, and the like; the storage data area may store data created according to the use of the blockchain node, and the like.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by hardware instructions related to a program, and the program may be stored in a computer readable storage medium, which includes: a Read Only Memory (ROM), a Random Access Memory (RAM), a magnetic or optical disk, or the like.

The technical features of the embodiments described above can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-described embodiments are merely illustrative of some embodiments of the present application, which are described in more detail and detail, but are not to be construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent should be subject to the appended claims.

The block chain is a novel application mode of computer technologies such as distributed data storage, point-to-point transmission, a consensus mechanism, an encryption algorithm and the like. A block chain (Blockchain), which is essentially a decentralized database, is a series of data blocks associated by using a cryptographic method, and each data block contains information of a batch of network transactions, so as to verify the validity (anti-counterfeiting) of the information and generate a next block. The blockchain may include a blockchain underlying platform, a platform product service layer, an application service layer, and the like.

Claims

1. A kitchen violation detection method, comprising:

and acquiring an image shot by the kitchen camera in real time, and inputting the image into a kitchen violation detection model for violation detection.

2. The method of kitchen violation detection according to claim 1, wherein after constructing public atlas a, the method further comprises:

3. The method for detecting kitchen violation according to claim 1, wherein the applying the violation pictures in the public atlas a to construct a public violation area set a comprises:

4. The method of kitchen violation detection according to claim 1, wherein said first iterative training of said codec model comprises:

5. The method of kitchen violation detection according to claim 4, wherein said feature matching training the code-decode model comprises:

and when the cross-over ratio IOU1 is larger than the second threshold value R2, taking the result of the original prediction interval [ R1, R2] as a negative sample, balancing the positive and negative samples for retraining, calculating again to obtain the cross-over ratio IOU2, continuously comparing with the next threshold value in the preset threshold value set R, iteratively improving the cross-over ratio, and finishing the feature matching training if the prediction is wrong or all the threshold values in the preset threshold value set R are compared.

6. The method for kitchen violation detection according to claim 4, wherein the performing violation classification training on the detection model with accurate feature matching comprises:

7. The method of kitchen violation detection according to claim 6, wherein said training a first iteration of an encoding-decoding model further comprises:

8. A kitchen violation detection device is characterized by comprising the following modules:

a sample library construction module: acquiring a kitchen violation picture disclosed on a network, constructing a public atlas A, acquiring a kitchen violation picture captured by a camera in an actual kitchen scene, constructing a real atlas B, constructing a public violation area set a by applying the violation picture in the public atlas A, and constructing a real violation area set B by applying the violation picture in the real atlas B;

a model initialization module: constructing a coding-decoding detection model based on a convolutional neural network structure;

a model training module: carrying out first iterative training on the coding-decoding model by taking the illegal picture of the public map set A and the labeled file in the public illegal region set a as training samples, and then carrying out second iterative training on the coding-decoding model by taking the labeled file in the real map set B and the labeled file in the real illegal region set B as training samples to obtain a kitchen illegal behavior detection model;

9. A kitchen violation detection device comprising a memory and a processor, the memory having stored therein computer-readable instructions that, when executed by the processor, cause the processor to perform the kitchen violation detection method of any of claims 1-7.

10. A storage medium having stored thereon computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the method of galley violation detection of any of claims 1-7.