CN110211065B - Color correction method and device for food material image - Google Patents

Abstract

The embodiment of the present invention discloses a method and device for color correction of food material images. The method includes: obtaining a video stream about food materials and determining key frames and non-key frames in the video stream; inputting the key frames into a preset In the assumed color correction model, the key frames are color corrected through the color correction model; the non-key frames are input into the preset atmospheric scattering model, and the non-key frames are corrected through the atmospheric scattering model. Color distortion is restored. Through this embodiment solution, the problem of color cast and highlight in the pictures taken by the camera is solved, and this solution can avoid the interference of high temperature inside the oven cavity and is robust to photographing different types of food ingredients.

Description

Color correction method and device for food image

Technical field

Embodiments of the present invention relate to image processing technology, and in particular, to a color correction method and device for food images.

Background technique

As artificial intelligence and home appliances become more and more closely integrated, ovens, one of the small home appliances, are becoming more and more intelligent and automated. The premise of this intelligence is good data acquisition channels and closed-loop data processing. Taking pictures and capturing video information with a camera is the most direct way to obtain visual data, and it is also one of the commonly used ways for ovens to obtain data. However, the oven plus camera solution still has many challenges on the data acquisition side. One of them is that when collecting image data in the internal environment of the oven cavity, serious interference will be caused by internal light sources and reflection and scattering of different materials. The video data collected by the camera requires high-quality pictures at the source, which can also reduce the burden of post-processing and increase the robustness of the entire post-processing system.

The more common methods currently used include the following:

1. Adaptive adjustment scheme based on white balance. Because most lights in the oven have color casts, this scheme can suppress the color cast problem to a certain extent. However, in actual use, it is often found that different ingredients have different colors due to white balance. This results in very different adjustment effects, and some may even reversely aggravate the color cast.

2. Avoid color casts and highlights through physical methods. The most common method is to use filters and camera masks. Filters can suppress the light component of a certain band to a certain extent, thereby correcting the color deviation problem. However, The filter cannot be adaptive, and the camera is often calibrated on a white background, and the image in the actual oven is very poor; the hood has a relatively large suppression effect on highlights, and is also a relatively direct solution.

3. By adding an RGB (red, green, and blue) color sensor to coordinate with the camera for correction, this solution can theoretically completely avoid the color cast problem, because the current three primary color components can be directly obtained through the RGB color sensor, and can be restored through a certain strategy. Color, eliminate color distortion, however, the following problems were found in actual use. First, the accuracy of correction completely depends on the RGB color sensor, which will bring a great burden to the installation and correction of the sensor; second, the entire The closed-loop color adjustment system is not stable. Especially in the high-temperature oven internal environment, it will cause serious interference and even lead to sudden changes in color adjustment.

Contents of the invention

Embodiments of the present invention provide a color correction method and device for food images, which can solve the problem of color casts and highlights in pictures taken by cameras.

In order to achieve the purpose of the embodiments of the present invention, the embodiments of the present invention provide a color correction method for food material images. The method may include:

Obtaining a video stream of ingredients and determining key frames and non-key frames in the video stream;

Input the key frames into a preset color correction model, and perform color correction on the key frames through the color correction model;

The non-key frames are input into a preset atmospheric scattering model, and the color distortion in the non-key frames is restored through the atmospheric scattering model.

In an exemplary embodiment of the present invention, the color correction model may be obtained by training a convolutional neural network model using pictures of food sample samples in the simulated oven cavity as training data and test data.

In an exemplary embodiment of the present invention, using food sample pictures in the simulated oven cavity as training data and test data, training the convolutional neural network model may include:

Obtain a picture set of food ingredients in the simulated oven cavity;

Extract the food sample pictures from the food material picture set, and divide the food sample pictures into two parts; one part is used as the training data, and the other part is used as the test data;

The convolutional neural network model is trained using the training data and the preset deep learning algorithm, and the training results are verified using the test data.

In an exemplary embodiment of the present invention, the method may further include: when training the convolutional neural network model, the intermediate layer convolution in the convolutional network adopts Deformable Convolution as the basic layer structure.

In an exemplary embodiment of the present invention, the food ingredient picture set may include one or more of the following picture data:

In the sample collection environment of the simulated oven cavity, image data of various food ingredients collected under a standard light source;

In the sample collection environment of the simulated oven cavity, image data of various food ingredients collected under the oven light; and,

In the sample collection environment of the simulated oven cavity, picture data of various food materials were collected at different collection temperatures and during uniform changes in the collection temperature.

In an exemplary embodiment of the present invention, the method may further include: separating the food material image from the background image in the extracted food material sample picture, and establishing an image binary segmentation sample set regarding the food material image. , and divide the image binary segmentation sample set into the training data and the test data.

In an exemplary embodiment of the present invention, the method may further include: inputting the non-key frame into a preset atmospheric scattering model, and performing color distortion in the non-key frame through the atmospheric scattering model. During the restoration process, the parameters of the current atmospheric scattering model are corrected based on the parameters of the atmospheric scattering model used in historical non-key frames before the current non-key frame.

In an exemplary embodiment of the present invention, modifying the parameters of the current atmospheric scattering model based on the parameters of the atmospheric scattering model used in historical non-key frames before the current non-key frame may include:

The parameters of the atmospheric scattering model used in the historical non-key frames and the parameters of the current atmospheric scattering model are weighted and accumulated to obtain corrected atmospheric scattering model parameters.

In an exemplary embodiment of the present invention, determining key frames and non-key frames in the video stream may include:

The video stream is input into a preset key frame judgment module to determine key frames through the judgment strategy in the key frame judgment module, and frames other than the key frames in the video stream are regarded as the non-key frames. Keyframe.

In an exemplary embodiment of the present invention, the determination strategy may include:

Where thr is the preset threshold, the initial value is 0, calculated through the first two frames, g is the gradient map of the image, n represents the nth frame, n is a positive integer, S is the saturation of the image, and num is the overall pixels of the image number, α is the equilibrium parameter.

An embodiment of the present invention also provides a color correction device for food material images, including a processor and a computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed by the processor, Implement the color correction method of food image as described in any one of the above.

The beneficial effects of embodiments of the present invention may include:

1. The color correction method of food material images according to the embodiment of the present invention may include: obtaining a video stream about the food material, and determining key frames and non-key frames in the video stream; inputting the key frames into a preset color correction model , perform color correction on the key frame through the color correction model; input the non-key frame into a preset atmospheric scattering model, and restore the color distortion in the non-key frame through the atmospheric scattering model . Through this embodiment solution, the problem of color cast and highlight in the pictures taken by the camera is solved, and this solution can avoid the interference of high temperature inside the oven cavity and is robust to photographing different types of food ingredients.

2. The color correction model of the embodiment of the present invention can be obtained by training the convolutional neural network model using pictures of food sample samples in the simulated oven cavity as training data and test data. Through this embodiment, the convolutional neural network model can be trained through pictures of various types of food materials, so that it is not affected by the type of food materials, and the image color can be adaptively corrected according to the types of food materials in the scene, which increases the advantages of the embodiments of the present invention. The color correction scheme is robust to different types of food being photographed.

3. According to the embodiment of the present invention, using the sample pictures of food materials in the simulated oven cavity as training data and test data, training the convolutional neural network model may include: obtaining a set of pictures of food materials in the simulated oven cavity; The food material sample pictures are extracted from the food material pictures, and the food material sample pictures are divided into two parts; one part is used as the training data, and the other part is used as the test data; through the training data and the preset deep learning algorithm The convolutional neural network model is trained, and the test data is used to verify the training results. Through this embodiment solution, the effectiveness of the color correction model can be ensured, and the reliability of the embodiment solution of the present invention can be ensured.

4. The method of the embodiment of the present invention may also include: when training the convolutional neural network model, the intermediate layer convolution in the convolutional network adopts Deformable Convolution as the basic layer structure. Through this embodiment, the convolutional neural network improves the extraction of color features of the food ingredients themselves, while suppressing the impact of irrelevant backgrounds (such as oven trays) on color correction.

5. The food material picture set in the embodiment of the present invention may include one or more of the following picture data: picture data of a variety of food materials collected under a standard light source in the sample collection environment of the simulated oven cavity; In the sample collection environment of the simulated oven cavity, the picture data of various food materials collected under the oven light; and, in the sample collection environment of the simulated oven cavity, under different collection temperatures and collection Picture data of various food ingredients collected during uniform changes in temperature. Through this embodiment solution, the comprehensiveness of the sample pictures is guaranteed, thereby further ensuring the reliability of the color correction model.

6. The method of the embodiment of the present invention may also include: inputting the non-key frame into a preset atmospheric scattering model, and restoring the color distortion in the non-key frame through the atmospheric scattering model. , correct the parameters of the current atmospheric scattering model based on the parameters of the atmospheric scattering model used by historical non-key frames before the current non-key frame. This embodiment solution increases the robustness to noise of the embodiment solution of the present invention through closed-loop adjustment.

Additional features and advantages of embodiments of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the embodiments of the invention may be realized and obtained by the structure particularly pointed out in the specification, claims and appended drawings.

Description of the drawings

The drawings are used to provide a further understanding of the technical solutions of the embodiments of the present invention, and constitute a part of the description. Together with the embodiments of the present application, they are used to explain the technical solutions of the embodiments of the present invention, and do not constitute a limitation to the technical solutions of the embodiments of the present invention. .

Figure 1 is a flow chart of a color correction method for food material images according to an embodiment of the present invention;

Figure 2 is a schematic diagram of a color correction method for food material images according to an embodiment of the present invention;

Figure 3 is a flow chart of a method for training a convolutional neural network model using pictures of food samples in the simulated oven cavity as training data and test data according to an embodiment of the present invention;

Figure 4 is a schematic diagram of a method for training a convolutional neural network model using pictures of food samples in a simulated oven cavity as training data and test data according to an embodiment of the present invention;

Figure 5 is a schematic diagram of the basic structure of an FCN according to an embodiment of the present invention;

Figure 6 is a schematic diagram of one-dimensional convolution according to an embodiment of the present invention;

Figure 7 is a schematic diagram of one-dimensional deconvolution according to an embodiment of the present invention;

Figure 8 is a basic flow chart of deformable convolution according to an embodiment of the present invention;

FIG. 9 is a block diagram and structural diagram of a color correction device for food image images according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the embodiments of the present invention more clear, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that, as long as there is no conflict, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

The steps illustrated in the flowcharts of the figures may be performed in a computer system, such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart diagrams, in some cases the steps shown or described may be performed in a different order than herein.

An embodiment of the present invention provides a color correction method for food material images, as shown in Figures 1 and 2. The method may include S101-S103:

S101. Obtain a video stream about food ingredients, and determine key frames and non-key frames in the video stream.

In the exemplary embodiment of the present invention, the solution of this embodiment can be applied to different types of scenarios such as steam ovens, built-in ovens, and desktop ovens in the current industry, and can be used for ingredients that often appear in recipes, such as meat and aquatic products. , vegetables, etc. can be applied to achieve good color adjustment effect.

In an exemplary embodiment of the present invention, determining key frames and non-key frames in the video stream may include:

The video stream is input into a preset key frame judgment module to determine key frames through the judgment strategy in the key frame judgment module, and frames other than the key frames in the video stream are regarded as the non-key frames. Keyframe.

In an exemplary embodiment of the present invention, after the video stream is input to the pre-processing system, it can directly enter the key frame judgment module, and the key frame is judged through the judgment strategy in the key frame judgment module.

In an exemplary embodiment of the present invention, the determination strategy may include:

Where thr is the preset threshold, the initial value is 0, calculated through the first two frames, g is the gradient map of the image, n represents the nth frame, n is a positive integer, S is the saturation of the image, and num is the overall pixels of the image number, α is the equilibrium parameter.

In an exemplary embodiment of the present invention, the value range of α can be between 0 and 1. According to experiments, this parameter can be updated using the following calculation formula:

Among them, λ can be set as the frame frequency. The larger the λ is, the slower the rate at which α tends to 1. At the same time, the larger the α is, the higher the confidence of using the picture saturation information. x is the sequence number of the frame.

S102. Input the key frame into a preset color correction model, and perform color correction on the key frame through the color correction model.

In an exemplary embodiment of the present invention, after the key frame is determined, the key frame can be sent to the color correction model for color correction. The obtained correction result is saved and can be recorded as F _d , and the original frame picture can be recorded as F _s .

In an exemplary embodiment of the present invention, the color correction model may be obtained by training a convolutional neural network model using pictures of food sample samples in the simulated oven cavity as training data and test data.

In an exemplary embodiment of the present invention, a neural network is used for color correction on key frames, which is a method of 'fitting' the color differences produced by different ingredients inside the oven based on statistical rules, and finally produces a relatively more suitable Models are used to express the colors of different ingredients, and finally the color correction is performed on key frame images.

In an exemplary embodiment of the present invention, the convolutional neural network model may be a fully convolutional neural network FCN structure.

In an exemplary embodiment of the present invention, as shown in Figures 3 and 4, using the food sample pictures in the simulated oven cavity as training data and test data, training the convolutional neural network model may include S201-S203 :

S201. Obtain a picture set of food ingredients in the simulated oven cavity.

In an exemplary embodiment of the present invention, the food ingredient picture set may include one or more of the following picture data:

In the sample collection environment of the simulated oven cavity, image data of various food ingredients collected under a standard light source;

In the sample collection environment of the simulated oven cavity, image data of various food ingredients collected under the oven light; and,

In the sample collection environment of the simulated oven cavity, picture data of various food materials were collected at different collection temperatures and during uniform changes in the collection temperature.

S202. Extract the food material sample picture from the food material picture set, and divide the food material sample picture into two parts; one part is used as the training data, and the other part is used as the test data.

S203. Train the convolutional neural network model using the training data and the preset deep learning algorithm, and use the test data to verify the training results.

In the exemplary embodiment of the present invention, any existing deep learning tool can be used to train the FCN (full convolutional network) network.

In the exemplary embodiment of the present invention, during actual operation, in order to increase the generalization ability of the model and improve the accuracy of color correction, some special methods are used during the training of the convolutional neural network model.

In the exemplary embodiment of the present invention, the FCN network structure is adopted. FCN is a typical fully convolutional neural network structure. Its core structure is the convolution operation layer, in order to be able to extract images under different visual fields. The characteristics of the target image and the basic structure of FCN are shown in Figure 5.

In an exemplary embodiment of the present invention, a convolution layer is used in the downsample part, and the sampling interval is controlled by the stride in the convolution window to gradually reduce the size of the feature map output by each layer.

In an exemplary embodiment of the present invention, the basic convolution structure is shown in Figure 6, which is a schematic diagram of one-dimensional convolution.

In an exemplary embodiment of the present invention, taking one-dimensional convolution as an example, the input a data in Figure 6 is a one-dimensional vector with a length of 5, and the convolution kernel b is a one-dimensional vector with a length of 3. In The given conditions are: padding=1, stride=2, the output of the convolution c=ab, where the symbol · represents the convolution operation, and the length of the output c is outSize=(inSize-kerSize+2*padding)/2 +1=(5-3+2*1)/2+1=3. Similarly, in two-dimensional convolution, the size of the output feature map can also be calculated in this way, and the so-called downsampling can be achieved by adjusting stride to control the size.

In the exemplary embodiment of the present invention, deconvolution is used for upsampling. Deconvolution is also a convolution method, but from the results, it is often used to increase the size of the feature map and is also used. It is commonly used in the upsampling layer. One of the deconvolution methods is briefly described below. The schematic diagram is shown in Figure 7.

In an exemplary embodiment of the present invention, as shown in Figure 7, the input is a one-dimensional vector a with a length of 3, the convolution kernel b is a one-dimensional vector with a length of 3, and the dimension of the output c is calculated as: outSize =(inSize-1)*stride+kerSize=(3-1)*1+3=5. It can be seen from the results that deconvolution can achieve the effect of upsampling.

In an exemplary embodiment of the present invention, the method may further include: when training the convolutional neural network model, the intermediate layer convolution in the convolutional network adopts Deformable Convolution as the basic layer structure.

In exemplary embodiments of the present invention, deformable convolution (deformable convolution) can also be used as a link in the upsampling process. Deformable convolution is based on the conventional convolution operation by modifying each windowing process. The input position corresponding to the convolution kernel can achieve the function of focusing on key features or object positions. The purpose of using deformable convolution in the plan is to improve the FCN network's extraction of the color features of the ingredients themselves, while suppressing irrelevant backgrounds (such as the oven's pallet) on color correction.

In an exemplary embodiment of the present invention, a basic block diagram of deformable convolution may be shown in Figure 8.

In an exemplary embodiment of the present invention, what is represented in parentheses in Figure 8 is the size of the feature map of each step. The value of the position offset can be obtained through the convolution operation of the first step, and then bilinear interpolation is used to obtain the value of the position offset from The original image is mapped to a new feature map, and finally an ordinary convolution operation is used on the new feature map to obtain the output.

In an exemplary embodiment of the present invention, the method may further include: separating the food material image from the background image in the extracted food material sample picture, and establishing an image binary segmentation sample set regarding the food material image. , and divide the image binary segmentation sample set into the training data and the test data.

In the exemplary embodiments of the present invention, during the implementation process, in addition to the basic methods used above, additional methods can also be adopted to enhance the effect.

In the exemplary embodiment of the present invention, the solution is to collect pictures in a standard simulation environment, so the background of the picture is highly relevant. In order to take advantage of this feature, the background can be filtered out and the focus can be focused on As for the ingredients themselves, in the actual implementation process, some samples can be used to separate the ingredients from the background (tray, inner wall of the oven cavity, etc.), and a sample set C for binary image segmentation can be established.

In an exemplary embodiment of the present invention, before training the FCN network, FCN can be used to complete an image segmentation task training for sample C. For this step, the average IU (Intersection and Union Ratio) of the test data set can be set. Above 70%, the IU can be calculated as follows:

In an exemplary embodiment of the present invention, after the segmentation task is completed, the parameters obtained in the previous step can be used as a benchmark to alternately train the up-sampling layer and the down-sampling layer. Because of the limitations of the sample set collection conditions, in order to make full use of the sample Effective information can be used multiple times during training, such as 5 times of folding training, and finally an FCN network model with relatively high accuracy is obtained.

S103. Input the non-key frame into a preset atmospheric scattering model, and restore the color distortion in the non-key frame through the atmospheric scattering model.

In the exemplary embodiment of the present invention, because the redundancy of data information is relatively serious for video streams, especially in the internal environment of the oven, within a period of time, it is assumed that color changes (mainly temperature) for the camera without losing the general The influence of the color of the collected pictures) on the color distortion is smooth, so there is no need to use neural network correction processing for non-key frame pictures. The atmospheric scattering model is used here. The atmospheric scattering model is generally used for illumination intensity. Simulation is of great significance to the dehazing and enhancement of images, because the presentation of color is also based on the intensity contrast of the three primary colors. Therefore, the use of atmospheric models can effectively restore color distortion.

In an exemplary embodiment of the present invention, after obtaining the key frame correction result, the atmospheric scattering model can be used to model the color correction model of ordinary frames (ie, non-key frames). The atmospheric scattering model relationship can be shown as follows:

I(x,λ)=t(x,λ)*R(x,λ)+A(1-t(x,λ))

Among them, t(x,λ) is the projection rate, R(x,λ) is the image after removing interference, A(1-t(x,λ)) is the atmospheric light intensity under natural illumination, and λ is the wavelength. Because the wavelengths of the three components R, G, and B are different, it can be considered that the λ of each channel is different. x is the distance from the light source to the imaging device. It can be considered a fixed value for each pixel. Then, we need to estimate The parameters are: t(x,λ) and the atmospheric light intensity A at infinity. Generally, A can be obtained through calibration and unified calculation in advance under the natural environment in the oven cavity. In this embodiment, it is obtained by calculating the pixel value at the highest brightness point, because:

A＝A(∞)

It can be considered that under most conditions, the strongest natural light is at infinity.

In an exemplary embodiment of the present invention, the method may further include: inputting the non-key frame into a preset atmospheric scattering model, and performing color distortion in the non-key frame through the atmospheric scattering model. During the restoration process, the parameters of the current atmospheric scattering model are corrected based on the parameters of the atmospheric scattering model used in historical non-key frames before the current non-key frame.

In an exemplary embodiment of the present invention, this partial embodiment is a non-keyframe adjustment on a closed loop. Because the video stream is data that is related to the timeline, based on this consideration, in the color correction of non-key frames, instead of using a set of parameters of the atmospheric scattering model alone, the historical data of the previous frames is integrated. (or observation results) to modify the parameters of the current model and save it. In actual tests, it was found that such closed-loop adjustment can increase the robustness of the color correction algorithm to noise according to the embodiment of the present invention.

In an exemplary embodiment of the present invention, modifying the parameters of the current atmospheric scattering model based on the parameters of the atmospheric scattering model used in historical non-key frames before the current non-key frame may include:

The parameters of the atmospheric scattering model used in the historical non-key frames and the parameters of the current atmospheric scattering model are weighted and accumulated to obtain corrected atmospheric scattering model parameters.

In an exemplary embodiment of the present invention, the atmospheric scattering model parameters can be updated using a weighted accumulation of historical values and current values. In the experiment, the weighted value can be 0.8, giving a weight of 0.8 to the historical value and a weight of 0.2 to the current value. Then update the current parameter value.

An embodiment of the present invention also provides a color correction device 1 for food material images. As shown in Figure 9, it may include a processor 11 and a computer-readable storage medium 12. Instructions are stored in the computer-readable storage medium 12. When the instruction is executed by the processor 11, any one of the color correction methods for food material images described above is implemented.

Embodiments of the present invention include at least the following beneficial effects:

1. Based on the convolutional neural network, the color distortion in different food ingredients scenes in the oven is adjusted to increase the reliability of the color correction algorithm of the embodiment of the present invention.

2. Based on the continuity and timeliness requirements of oven-based video shooting, for redundant video frames, instead of the convolutional neural network, the atmospheric scattering model is used as the adjustment model for color distortion to simplify the color adjustment process of non-key frames.

3. To deal with the complex lighting environment inside the oven cavity, the parameters of the atmospheric scattering model use historical data in time series as a reference, which increases the stability of the color correction algorithm of the embodiment of the present invention.

4. Balance timeliness and algorithm accuracy. The plan uses local adjustment and global adjustment to intersperse each other to reduce the amount of calculation and enhance real-time performance.

Those of ordinary skill in the art can understand that all or some steps, systems, and functional modules/units in the devices disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may consist of several physical components. Components execute cooperatively. Some or all of the components may be implemented as software executed by a processor, such as a digital signal processor or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer-readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). As is known to those of ordinary skill in the art, the term computer storage media includes volatile and nonvolatile media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. removable, removable and non-removable media. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disk (DVD) or other optical disk storage, magnetic cassettes, tapes, disk storage or other magnetic storage devices, or may Any other medium used to store the desired information and that can be accessed by a computer. Additionally, it is known to those of ordinary skill in the art that communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

Claims (9)

1. A method of color correction of an image of food material, the method comprising:

acquiring a video stream related to food materials, and determining key frames and non-key frames in the video stream;

inputting the key frame into a preset color correction model, and performing color correction on the key frame through the color correction model;

inputting the non-key frames into a preset atmospheric scattering model, reducing color distortion in the non-key frames through the atmospheric scattering model, correcting parameters of the current atmospheric scattering model according to parameters of an atmospheric scattering model adopted by historical non-key frames before the current non-key frames, wherein the atmospheric scattering model is used for simulating illumination intensity so as to cope with the illumination environment in the oven cavity;

the color correction model is obtained by training with food material sample pictures in the simulated oven cavity as data;

obtaining a food material picture set in the simulation oven cavity, extracting the food material sample picture from the food material picture set, wherein the food material picture set comprises one or more of the following picture data:

in a sample collection environment of the simulated oven cavity, collecting picture data of various food materials under a standard light source;

in a sample collection environment of the simulated oven cavity, collecting picture data of various food materials under the oven light; the method comprises the steps of,

in the sample collection environment of the simulation oven cavity, the collected picture data of various food materials are collected at different collection temperatures and in the uniform change process of the collection temperatures.

2. The method for correcting color of food material image according to claim 1, wherein the color correction model is obtained by training a convolutional neural network model by using food material sample pictures in a simulated oven cavity as training data and test data.

3. The method for color correction of food material images according to claim 2, wherein training the convolutional neural network model with food material sample pictures in the simulated oven cavity as training data and test data comprises:

acquiring a food material picture set in the simulated oven cavity;

extracting the food material sample picture from the food material picture set, and dividing the food material sample picture into two parts; one part is used as the training data, and the other part is used as the test data;

training the convolutional neural network model through the training data and a preset deep learning algorithm, and verifying a training result by adopting the test data.

4. A method of colour correction of an image of food material according to claim 2 or 3, characterised in that the method further comprises: when training the convolutional neural network model, the middle layer convolution in the convolutional network adopts a variable-shape convolution Deformable Convolution as a basic layer structure.

5. A method of color correction of an image of food material according to claim 3, the method further comprising: separating the food material image from the background image in the extracted food material sample image, establishing an image binary segmentation sample set related to the food material image, and dividing the image binary segmentation sample set into the training data and the test data.

6. The method for color correction of an image of a food material according to claim 1, wherein,

the correcting the parameters of the current atmospheric scattering model according to the parameters of the atmospheric scattering model adopted by the historical non-key frames before the current non-key frame comprises the following steps:

and carrying out weighted accumulation calculation on the parameters of the atmospheric scattering model adopted by the historical non-key frame and the parameters of the current atmospheric scattering model to obtain corrected atmospheric scattering model parameters.

7. The method of claim 1, wherein the determining key frames and non-key frames in the video stream comprises:

inputting the video stream into a preset key frame judging module to judge key frames according to a judging strategy in the key frame judging module, and taking frames except the key frames in the video stream as the non-key frames.

8. The method of claim 7, wherein the determining strategy comprises:

wherein thr is a preset threshold value, an initial value is 0, g is a gradient map of an image obtained by calculation of the first two frames, n represents an nth frame, n is a positive integer, S is the saturation of the image, num is the number of integral pixels of the image, and alpha is a balance parameter.

9. A color correction device for a food material image, comprising a processor and a computer readable storage medium having instructions stored therein, wherein the instructions, when executed by the processor, implement the color correction method for a food material image according to any one of claims 1-8.