A visual intelligent identification and control system for doneness of large-capacity stir-fried meat

Abstract

The present application relates to a kind of visual intelligent identification and control system for the doneness of large-capacity stir-frying meat, belongs to the field of food automatic stir-frying, machine vision, intelligent control technology.The system includes: collecting laser speckle and visual image during the process of meat stir-frying;After pre-processing, the light spot area, color and texture features are extracted;The logarithmic correlation model of surface maturity value and center maturity value is used to invert the center true doneness from surface features;Combined with deep learning model, doneness classification and working condition verification are carried out, and heating and stir-frying are closed-loop controlled accordingly.The present application solves the problem of accurate detection of internal doneness caused by oil smoke interference and material stacking in large-capacity stir-frying by laser speckle imaging combined with maturity value kinetics model, realizes non-destructive, real-time monitoring and automatic accurate control of meat doneness, effectively avoids overcooked outside or local burnt, and can be widely applied to large-capacity stir-frying scenes such as group meal, prepared dishes and meat product processing.

Description

Technical Field

[0001] This invention relates to a visual intelligent recognition and control system for meat doneness that is adapted for large-capacity stir-frying, belonging to the fields of automated food stir-frying, machine vision, and intelligent control technology. Background Technology

[0002] Currently, large-capacity stir-frying processes are commonly used in central kitchens for group meals, pre-prepared food processing plants, and industrialized meat production lines. Single-batch feed amounts can reach 10 kg or even several tons, making it a core process in large-scale meat production. The doneness of the meat directly determines the product's taste, flavor, food safety, and shelf life, making it a crucial control indicator in the production process. At present, most large-capacity meat stir-frying production lines still rely on operators' visual judgment combined with personal experience to determine the cooking endpoint. This method is highly subjective, and standards are difficult to unify between different operators and different production shifts, easily leading to significant fluctuations in product batch quality. Furthermore, the stir-frying area is constantly exposed to high temperatures and pervasive fumes, significantly increasing labor intensity and increasing the risk of burns and smoke-related injuries, contradicting the modern food factory's development direction of automation, unmanned operation, and standardization.

[0003] Numerous studies and technological explorations have been conducted both domestically and internationally regarding the detection of meat cooking doneness. Currently, mainstream detection methods include contact temperature measurement, traditional physicochemical testing, and computer vision detection. However, none of these technologies are well-suited to the complex on-site conditions of large-volume stir-frying. The industry commonly uses contact temperature sensors such as thermocouples to determine meat doneness, relying on single-point temperature data as a judgment criterion. However, in large-volume stir-frying, materials are stacked and continuously turned, resulting in extremely uneven temperature distribution within the pot. The surface of the food heats up rapidly in contact with hot oil and air, while the bottom and geometric center of the material stack heat up more slowly. Single-point temperature measurement cannot reflect the true doneness of the entire batch. Furthermore, contact probes, in prolonged contact with oil and food, easily accumulate residue and grease, gradually reducing temperature measurement accuracy. There is also a risk of cross-contamination of food, making them unsuitable for continuous industrial mass production. Traditional physicochemical testing determines the doneness of meat by measuring physicochemical indicators such as the degree of protein denaturation, moisture content, and myoglobin oxidation rate. The test results are accurate, but this method is a destructive sampling test. The operation process is cumbersome and the test is time-consuming. It can only be used for laboratory sampling and cannot achieve real-time online monitoring during dynamic cooking.

[0004] Computer vision, with its advantages of being non-destructive, non-contact, and capable of continuous data acquisition, has become a popular technology for online food quality inspection and is now widely used in fields such as fresh meat quality screening and small-sample cooking doneness identification. Traditional machine vision solutions mostly use color cameras to acquire images, extract color features such as RGB, HSV, and CIE-LAB, and combine them with texture and contour information, along with threshold segmentation and traditional machine learning algorithms to distinguish between raw and cooked foods. These methods can be used normally under ideal conditions in a static laboratory environment with no oil fume interference and uniform meat sample sizes. However, large-scale cooking environments involve a large amount of oil fume, high-temperature steam, and hot oil reflections, which can easily cause image color shifts, increased noise, and abnormal highlight areas. In addition, the continuous movement and mutual occlusion of materials mean that relying solely on color and shallow texture features will lead to a significant decrease in recognition accuracy. Existing traditional vision solutions have not been specifically optimized for various interferences in industrial environments, resulting in poor engineering implementation results. With the continuous development of convolutional neural network technology, deep learning models such as Inception, EfficientNet, and MobileNet have been gradually introduced into the field of food ripeness recognition. However, most existing related studies are based on static laboratory samples to construct datasets. The sample scenes are clean, and the material forms are uniform. They have not been optimized with algorithms and datasets in combination with actual working conditions such as large-scale dynamic stir-frying, material stacking, and image shaking. More importantly, the true ripeness of meat is determined by the state of the cold spot at the geometric center of the food. Most existing visual recognition studies only classify the surface features of the food and have not established a correlation model between surface visual features and the ripeness of the internal center, which cannot solve problems such as the outside being cooked but the inside being raw, or localized burning. At the same time, most studies have not clarified the applicable conditions of the model, such as the thickness of the meat, the shooting light, and the exposure. Once the on-site working conditions change, the model's recognition ability will rapidly decline, making it difficult to operate stably in the long term.

[0005] Furthermore, some cutting-edge research has begun to explore food photophysical property detection technologies such as laser speckle and spectroscopy. These technologies utilize changes in the reflection, scattering, and refraction of light after meat is heated to determine its doneness. While these technologies offer better resistance to color interference than ordinary color vision, current research remains largely theoretical, lacking both in-depth application based on food ripening kinetics and integration with large-capacity industrial cooking equipment to form a complete closed-loop control system. Consequently, they have yet to be commercially viable. Overall, existing technologies reveal a significant disconnect between laboratory research and actual industrial needs. Industrial testing methods suffer from low accuracy and inconsistent performance, and most technologies fail to establish a clear link between the surface condition of the food and its true core doneness. The industry urgently needs an integrated solution that balances interference resistance, real-time performance, comprehensive doneness assessment, and automated control. Summary of the Invention

[0006] To address the technical challenges of achieving precise, non-destructive, real-time online detection and automated closed-loop control during large-volume meat stir-frying due to interference from oil fumes and steam, material stacking, and inconsistent cooking temperatures, this invention provides a visual intelligent recognition and control system for meat cooking temperature adapted for large-volume stir-frying. The technical solution is as follows: In a first aspect, the present invention provides a visual intelligent recognition and control system for meat doneness that is adapted to large-capacity stir-frying, comprising: The hardware acquisition unit is deployed around the high-capacity frying equipment to collect multimodal data during the meat frying process; An image preprocessing unit, connected to the hardware acquisition unit, is used to preprocess the acquired image data; The feature parsing unit, connected to the image preprocessing unit, is used to extract optical property features and visual texture features from the preprocessed image; The maturity value conversion unit is connected to the feature analysis unit and has a built-in surface maturity value. M S With center maturity value M C The correlation model is used to calculate the surface maturity value based on the photophysical characteristics and visual texture characteristics, and to invert the meat center maturity value. A deep learning recognition unit, connected to the maturity value conversion unit, is used to classify and identify the meat's maturity status based on the meat's center maturity value and to perform working condition boundary verification. A logic control unit, connected to the deep learning recognition unit, is used to generate control commands based on the recognition results; The actuator unit, connected to the logic control unit, is used to perform heating, stirring and discharging actions.

[0007] Optionally, the hardware acquisition unit includes: The laser speckle imaging module consists of a He-Ne laser, an industrial CCD camera, a neutral density filter, and a high-temperature resistant, oil-proof, and sealed protective cover. It adopts a 45° fixed incident light path and is used to acquire laser speckle images of meat surfaces to extract the speckle area features. Auxiliary color industrial cameras, no less than 2 channels, are deployed from multiple angles and equipped with ring anti-reflective supplementary lights to capture color and texture images of meat surfaces. The environmental monitoring module is used to collect the equivalent EV value of on-site illumination and ambient temperature and humidity in real time. The material parameter input module is used to input the thickness of the meat chunks, the amount of material added at one time, and the process parameters of the frying temperature.

[0008] Optionally, the image preprocessing unit uses multi-frame mean filtering and specular masking algorithms to eliminate noise and reflections caused by oil fumes and steam, and sharpens and repairs motion blur caused by stir-frying.

[0009] Optionally, the feature parsing unit includes: The optical property feature extraction subunit is used to perform ellipse fitting on the laser speckle image and calculate the laser spot area. The visual feature extraction subunit is used to extract HSV color histogram, LBP texture features and image singular value features from color images, and to perform feature dimensionality reduction through PCA principal component analysis.

[0010] Optionally, in the maturity value conversion unit, the surface maturity value M S With center maturity value M C The association model is a logarithmic association model; The maturity value conversion unit is based on the meat block thickness, stir-frying temperature, and surface heat transfer coefficient. H fp Grouping and applying the fitting formula, using ln (M C ) and M S The logarithmic linear relationship, derived from the surface maturity value M S Inversion calculation yields the central maturity value of meat. M C .

[0011] Optionally, the deep learning recognition unit is equipped with a MobileNet-V2 convolutional neural network trained based on transfer learning.

[0012] Optionally, the system is configured with working condition boundary verification rules: the thickness of a single piece of meat to be stir-fried is ≤1cm, and the image equivalent exposure EV value is between -2 and +2; if any threshold is exceeded, the system triggers an alarm signal.

[0013] Optionally, the logic control unit adopts an industrial PLC and embedded industrial computer architecture, with built-in segmented control logic: When the area is identified as immature, maintain the preset heating power and standard stir-frying speed; When the area is identified as being close to overcooked, reduce the heating power and increase the stir-frying speed; When the target maturity value is detected, a stop heating command is immediately issued and the discharge procedure is triggered.

[0014] Secondly, the present invention provides a visual intelligent recognition and control method for the doneness of meat adapted to large-capacity stir-frying, applied to the system described in any of the above claims, comprising the following steps: Step 1: Start the system and check the hardware status. Enter the meat block thickness, feed amount, target maturity value and stir-frying temperature process parameters. The system will call the corresponding surface maturity value and center maturity value fitting model according to the parameters. Step 2: Start the high-capacity stir-frying equipment and continuously collect laser spot images, surface visual images, and equivalent EV values ​​of ambient light during the meat stir-frying process through the hardware acquisition unit; Step 3: Denoise, remove reflections, repair motion blur, and crop the effective area of ​​the acquired image, and extract the laser spot area, color, texture, and singular value features; Step 4: Calculate the surface maturity value of the current material. M S The central maturity value was obtained by inversion based on the logarithmic correlation model. M C The maturity classification and working condition boundary verification are completed using a deep learning model. Step 5: Issue instructions based on the ripeness recognition results, dynamically adjust the heating power and stirring speed, and automatically stop heating and start discharging when the preset ripeness endpoint is reached.

[0015] Optionally, in step 4, maturity classification is performed using a MobileNet-V2 convolutional neural network.

[0016] The beneficial effects of this invention are: This invention employs a technical solution combining laser speckle imaging and deep learning, and establishes a logarithmic correlation model between surface maturity and center maturity values. This effectively solves the technical challenges of achieving accurate, non-destructive, real-time online detection and automated closed-loop control in large-volume meat stir-frying due to interference from oil fumes and steam, material stacking, and inconsistent surface and interior maturity. Specifically, this invention uses laser speckle imaging as the core detection method, leveraging the strong anti-interference capability of optical properties to overcome the shortcomings of traditional color vision schemes, such as image color shift, increased noise, and significantly reduced recognition accuracy under high temperature, oil fumes, steam, and hot oil reflection environments. Furthermore, by constructing… M S - M C The logarithmic correlation model can accurately invert the true maturity of the internal central area of ​​meat based on the surface characteristics of the meat, fundamentally solving the industry pain point of "cooked on the outside and raw on the inside" or "partially burnt" when frying large quantities of materials, and making up for the shortcomings of existing technologies that can only detect the surface characteristics of food.

[0017] This invention not only achieves non-destructive, continuous, and real-time detection of meat doneness, with detection stability and accuracy far exceeding traditional methods, but also realizes fully automated closed-loop control through the linkage of heating, stir-frying, and discharging mechanisms, eliminating the differences in human experience and ensuring a high degree of consistency in product quality. Furthermore, the system uses a lightweight deep learning model with fast computing speed, which can be deployed on industrial control computers and other equipment, meeting the low-latency requirements of industrial sites. It also incorporates multiple boundary condition checks and early warning mechanisms, providing strong fault tolerance and making it widely applicable to large-capacity cooking scenarios such as group catering, pre-prepared dishes, and meat product processing. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the visual intelligent recognition and control method for the doneness of meat adapted to large-capacity stir-frying, as described in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0021] Example 1: This embodiment provides a visual intelligent recognition and control system for meat cooking with large capacity, including: a hardware acquisition unit, an image preprocessing unit, a feature analysis unit, a maturity value conversion unit, a deep learning recognition unit, a logic control unit, and an actuator unit.

[0022] The hardware acquisition unit is the front-end sensing hardware of the system, deployed around large-capacity woks / continuous cooking production lines, and includes: (1) Laser speckle imaging module: The core detection equipment consists of a 632.8 nm He-Ne laser, an industrial CCD camera, a neutral density filter, and a high-temperature resistant and oil-proof sealed protective cover; it adopts a 45° incident optical path, with a fixed shooting distance and focal length, to collect images of laser spots on meat, which are used to extract the core photophysical characteristics of the spot area; (2) Auxiliary color industrial camera: multi-view deployment (no less than 2 channels), equipped with ring anti-reflective supplementary light, to collect images of the color and texture of meat surface as auxiliary verification basis; (3) Environmental monitoring module: Real-time collection of on-site light equivalent EV value and ambient temperature and humidity to determine the working status of the equipment; (4) Material parameter input module: Manual / automatic input of process parameters such as meat block thickness, single feeding amount, preset frying temperature, heat exchange coefficient, etc.

[0023] The image preprocessing unit interfaces with the hardware acquisition unit to perform image correction for large volumes of cooking steam, fumes, motion blur, and localized reflections. Its functions include: (1) Image noise reduction: Multi-frame mean filtering and specular masking algorithm are used to eliminate image noise and reflection caused by oil fumes and steam; (2) Motion blur repair: Sharpening is applied to the dynamic images of the stir-frying process; (3) ROI region cropping: Automatically crop the effective meat sample area and remove invalid pixels from the pot body, flame, and background; (4) Image standardization: unify image size and brightness to match the input specifications of deep learning models.

[0024] The feature parsing unit is divided into two sub-units: optical property feature extraction and image visual feature extraction. (1) Optical property feature extraction subunit: ellipse fitting is performed on the laser speckle image to calculate the laser spot area; this parameter deviates from the dynamic parameter judged by human visual subjective familiarity by only 4%, and is the core judgment index of this system; (2) Visual feature extraction subunit: Extract HSV color histogram, LBP texture features and image singular value features from color images, and use them as auxiliary discriminant features after dimensionality reduction by PCA principal component analysis.

[0025] The maturity value conversion unit has a built-in "maturity value conversion" function based on extensive experimental calibration. M C -M S The "logarithmic correlation model" works as follows: it calculates the current material surface maturity value based on the laser spot area and visual characteristics. M S Combined with the entered meat block thickness, stir-frying temperature, and surface heat transfer coefficient H fp Call the corresponding group fitting formula, from M S Inversion calculation of material center maturity value M C , M C This serves as a standard for determining the true doneness of meat in large-capacity stacks.

[0026] The deep learning recognition unit is the core computing module of the system, equipped with the MobileNet-V2 convolutional neural network trained based on transfer learning. This model uses 22708 standard meat images with different central maturity values as the dataset to complete transfer learning training, achieving a classification accuracy of 99.8% and a single iteration time of 2.13s, which meets the industrial real-time performance requirement; according to M C the value, it is automatically divided into three states: immature ( M C <250min), mature (250~550min), and over-mature ( M C >550min); it also compares the image EV value and meat block thickness in real time, and outputs an early warning signal when the threshold is exceeded.

[0027] The logic control unit adopts an industrial PLC + embedded industrial personal computer architecture, receives the output result of the deep learning recognition unit, and has built-in segmented control logic: (1) Immature interval: maintain the preset heating power and standard stir-frying rotation speed; (2) When the target maturity value is reached: immediately issue instructions to stop heating and reduce the stir-frying speed, and trigger the discharging procedure; (3) Near the over-mature interval: reduce the heating power in advance and increase the stir-frying rotation speed for early warning to prevent burning; (4) Abnormality determination: when the EV value < -2, the EV value > 2 or the meat block thickness > 1cm, the equipment alarm is triggered to prompt the staff to adjust the working conditions.

[0028] The actuator unit is a supporting executive component for the large-capacity frying equipment, which receives instructions from the control unit, including: (1) Heating mechanism: electromagnetic / gas heating module, supporting continuous power adjustment; (2) Stir-frying mechanism: stirring paddle / pan-turning mechanism with adjustable rotation speed and forward / reverse rotation; (3) Automatic discharging mechanism, oil fume exhaust fan, and cleaning mechanism.

[0029] Embodiment 2: This embodiment provides a control method of the visual intelligent recognition and control system for meat maturity adapted to large-capacity frying, comprising: Step 1: Equipment initialization and process parameter entry.

[0030] Start the system, check the working status of the laser imaging module, industrial camera and light source; enter process parameters such as meat block thickness, feeding amount, target maturity value, and frying temperature, and the system automatically invokes the corresponding M C - M S fitting model.

[0031] Step 2: Real-time acquisition of images of the stir-frying process.

[0032] The large-capacity wok features a heating and stirring mechanism, a laser speckle module and an auxiliary camera that continuously acquire images, and an environmental module that synchronously monitors the equivalent EV value.

[0033] Step 3: Preprocess and extract features from the acquired images.

[0034] The preprocessing unit performs noise reduction, de-glare removal, blur repair, and region extraction; the feature parsing unit extracts the laser spot area, color, texture, and singular value features.

[0035] Step 4: Calculation of maturity value and identification of maturity level.

[0036] Maturity value conversion unit calculation M S And inversion to obtain M C The deep learning model completes the ripeness classification and verifies whether the lighting and meat thickness meet the usage boundaries.

[0037] Step 5: Closed-loop control execution The control unit issues instructions based on the recognition results to dynamically adjust the heating power and stirring speed; when the preset maturity endpoint is reached, heating is automatically stopped and the material is discharged.

[0038] Example 3: This embodiment uses a commercially available large-capacity electromagnetic wok with a single feed capacity of 40kg. The experimental raw material is pork tenderloin. The cooking oil temperature is controlled at 100~120℃. The experimental environment is a conventional production workshop with oil fumes and steam. The center maturity value of the food is used as the standard for judging the doneness. The system is equipped with a 632.8nm He-Ne laser speckle module and two-channel anti-reflective color cameras. Image preprocessing functions such as oil fume noise reduction and motion blur repair are enabled. It is equipped with a MobileNet-V2 model trained by transfer learning and a logarithmic correlation algorithm between surface maturity value and center maturity value. A meat sample with a thickness of 0.8cm is selected, and the image equivalent EV value is set to 0. After testing with 800 sets of dynamic samples, the recognition accuracy reaches 99.7%, the single frame response latency is 2.12 seconds, and the control error at the end of cooking is less than 3 seconds. The entire system can stably achieve automated control.

[0039] Example 4: The large-capacity electromagnetic wok, cooking oil temperature, workshop environment, hardware equipment, image preprocessing program, deep learning model, and maturity value association algorithm used in this embodiment are completely consistent with those in Embodiment 3. Only the specifications of the experimental raw materials were changed; the thickness of the pork tenderloin was adjusted to the system's specified critical value of 1.0 cm, and the image equivalent EV value was set to 1. The test was still conducted using 800 sets of dynamic cooking samples. Due to the increased thickness of the meat samples, the bottom layer occlusion phenomenon was exacerbated after the materials were stacked, and some sample feature extraction showed slight deviations. The final recognition accuracy was 85.2%, and the single-frame response latency remained at approximately 2.12 seconds, without significant change. The system did not trigger any alarms throughout the process and was able to normally complete a series of automated control actions, including meat maturity recognition, heating power and stir-fry speed adjustment, and cooking endpoint determination.

[0040] Example 5 This embodiment uses all the experimental equipment, cooking process, hardware configuration, algorithm program, and 0.8cm thick pork tenderloin as raw material from Embodiment 3. Only the on-site lighting parameters are adjusted to raise the image's equivalent EV value to the system's critical value of 2. The same number of dynamic samples are used for testing. Due to the exposure parameters being close to the upper limit, a slight anomaly occurred in the image brightness, and a few image features shifted. The final recognition accuracy was 83.9%, and the single-frame response latency remained at 2.12 seconds. Only a few samples were misjudged during the experiment; the system operated smoothly overall, and all control commands were output normally, meeting the requirements for large-capacity cooking production.

[0041] Comparative Example 1 The experimental equipment, cooking conditions, and raw material specifications in this comparative example are the same as those in Example 3. The laser speckle detection module was removed, and the ripeness identification was performed solely by extracting color features from a regular color industrial camera. This method was greatly affected by on-site oil fumes and hot oil reflections, resulting in an overall recognition accuracy of only 61.3%, which is not applicable to online detection and automatic control of the production line.

[0042] Comparative Example 2 This comparative example retains the complete hardware and MobileNet-V2 model of Example 3, but removes the logarithmic correlation algorithm between surface maturity value and center maturity value. It directly determines the maturity by relying solely on the surface image of the food, resulting in a final recognition accuracy rate of 70.1%. This approach fails to address the issue of large-volume stir-fried materials being cooked on the outside but undercooked on the inside, leading to significant deviations in the judgment results.

[0043] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying, characterized in that, include: The hardware acquisition unit is deployed around the high-capacity frying equipment to collect multimodal data during the meat frying process; An image preprocessing unit, connected to the hardware acquisition unit, is used to preprocess the acquired image data; The feature parsing unit, connected to the image preprocessing unit, is used to extract optical property features and visual texture features from the preprocessed image; The maturity value conversion unit is connected to the feature analysis unit and has a built-in surface maturity value. M S With center maturity value M C The correlation model is used to calculate the surface maturity value based on the photophysical characteristics and visual texture characteristics, and to invert the meat center maturity value. A deep learning recognition unit, connected to the maturity value conversion unit, is used to classify and identify the meat's maturity status based on the meat's center maturity value and to perform working condition boundary verification. A logic control unit, connected to the deep learning recognition unit, is used to generate control commands based on the recognition results; The actuator unit, connected to the logic control unit, is used to perform heating, stirring and discharging actions.

2. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The hardware acquisition unit includes: The laser speckle imaging module consists of a He-Ne laser, an industrial CCD camera, a neutral density filter, and a high-temperature resistant, oil-proof, and sealed protective cover. It adopts a 45° fixed incident light path and is used to acquire laser speckle images of meat surfaces to extract the speckle area features. Auxiliary color industrial cameras, no less than 2 channels, are deployed from multiple angles and equipped with ring anti-reflective supplementary lights to capture color and texture images of meat surfaces. The environmental monitoring module is used to collect the equivalent EV value of on-site illumination and ambient temperature and humidity in real time. The material parameter input module is used to input the thickness of the meat chunks, the amount of material added at one time, and the process parameters of the frying temperature.

3. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The image preprocessing unit uses multi-frame mean filtering and specular masking algorithms to eliminate noise and reflections caused by oil fumes and steam, and sharpens and repairs motion blur caused by stir-frying.

4. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The feature parsing unit includes: The optical property feature extraction subunit is used to perform ellipse fitting on the laser speckle image and calculate the laser spot area. The visual feature extraction subunit is used to extract HSV color histogram, LBP texture features and image singular value features from color images, and to perform feature dimensionality reduction through PCA principal component analysis.

5. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, In the maturity value conversion unit, the surface maturity value M S With center maturity value M C The association model is a logarithmic association model; The maturity value conversion unit is based on the meat block thickness, stir-frying temperature, and surface heat transfer coefficient. H fp Grouping and applying the fitting formula, using ln (M C ) and M S The logarithmic linear relationship, derived from the surface maturity value M S Inversion calculation yields the central maturity value of meat. M C .

6. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The deep learning recognition unit is equipped with a MobileNet-V2 convolutional neural network trained based on transfer learning.

7. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The system is equipped with working condition boundary verification rules: the thickness of a single piece of meat to be stir-fried is ≤1cm, and the image equivalent exposure EV value is between -2 and +2; if any threshold is exceeded, the system will trigger an alarm signal.

8. The visual intelligent recognition and control system for meat doneness adapted to large-capacity stir-frying as described in claim 1, characterized in that, The logic control unit adopts an industrial PLC and embedded industrial computer architecture, and has built-in segmented control logic: When the area is identified as immature, maintain the preset heating power and standard stir-frying speed; When the area is identified as being close to overcooked, reduce the heating power and increase the stir-frying speed; When the target maturity value is detected, a stop heating command is immediately issued and the discharge procedure is triggered.

9. A method for visual intelligent recognition and control of meat doneness adapted to large-capacity stir-frying, characterized in that, Applied to the system as described in any one of claims 1 to 8, comprising the following steps: Step 1: Start the system and check the hardware status. Enter the meat block thickness, feed amount, target maturity value and stir-frying temperature process parameters. The system will call the corresponding surface maturity value and center maturity value fitting model according to the parameters. Step 2: Start the high-capacity stir-frying equipment and continuously collect laser spot images, surface visual images, and equivalent EV values ​​of ambient light during the meat stir-frying process through the hardware acquisition unit; Step 3: Denoise, remove reflections, repair motion blur, and crop the effective area of ​​the acquired image, and extract the laser spot area, color, texture, and singular value features; Step 4: Calculate the surface maturity value of the current material. M S The central maturity value was obtained by inversion based on the logarithmic correlation model. M C The maturity classification and working condition boundary verification are completed using a deep learning model. Step 5: Issue instructions based on the ripeness recognition results, dynamically adjust the heating power and stirring speed, and automatically stop heating and start discharging when the preset ripeness endpoint is reached.

10. The method for visual intelligent recognition and control of meat doneness adapted for large-capacity stir-frying according to claim 9, characterized in that, In step 4, the MobileNet-V2 convolutional neural network is used for maturity classification.

Images