Potato chip production equipment

Abstract

The invention provides potato chip production equipment. The potato chip production equipment comprises a machine body and a screening unit mounted on the machine body, according to the invention, firstly, the screening unit is adopted to screen out potatoes with proper sizes, potato chips discharged from the third discharge port are vertically cut by the cutter, and the cutter is controlled by the air cylinder, so that proper materials are automatically screened out for cutting; an industrial camera is matched with a feedback control method based on image analysis detection, the acceleration of the air cylinder and the tightness of the cutter are fed back, the acceleration of the air cylinder and the fixing tightness of the cutter are repeatedly adjusted, so that the fragment rate of slices is reduced, and the cutter rapidly cuts out complete potato slices; on-line detection precision of 100 thousand pieces per hour is achieved based on an image analysis detection feedback control method, efficiency is improved by 200 times compared with traditional manual sampling inspection, and meanwhile raw material loss is reduced by 30% or above.

Description

Technical Field

[0001] The invention relates to potato slice production equipment and food processing equipment. Background Art

[0002] Potatoes, also known as potatoes, are annual herbaceous plants of the genus Solanaceae, Solanaceae. The processing procedures of potato pieces include washing, peeling, cutting, blanching, dehydration, frying, freezing, packaging, etc. Among them, in the cutting process, a motor may be selected to drive the tool for cutting, or a cylinder may be used to drive the tool up and down for fast cutting. The shape of the tool may be a single-edged tool, a double-edged tool, or even a curved tool may be used to cut wavy potato chips. Due to the different sizes and shapes of potatoes, and the fact that potatoes contain hard particles such as sand and gravel, the tool is easily blunted. At the same time, the shear force of the contact between the tool and the potato may cause severe rupture of potato cells, rough surface, and increased free starch. Too low a shear force will cause incomplete cutting, resulting in burrs or adhesion, affecting the subsequent blanching and drying uniformity.

[0003] Therefore, the present invention provides a potato slice production device with a potato slice cross-section monitoring method for feedback of the shear force of the cutter. Summary of the Invention

[0004] The present invention provides a potato chip production device, which can effectively solve the above problems.

[0005] The present invention is achieved in that:

[0006] A potato chip production device includes a machine body and a screening unit mounted on the machine body; the present invention also includes a cylinder, a cutter, a cutter holder, and an industrial camera. The cylinder is located on one side of the machine body and also above a third discharge port. The output end of the cylinder is mechanically connected to the cutter holder, and the cutter holder is detachably fixedly connected to the cutter.

[0007] As a further improvement, the tool is locked on the tool holder by using bolts.

[0008] As a further improvement, a ring-shaped LED light is installed on the machine body around the third discharge port.

[0009] As a further improvement, the number of the industrial camera is one, which is used to shoot a front video file of the potato slice cross section.

[0010] As a further improvement, a controller is also included for controlling the acceleration and vibration frequency of the cylinder.

[0011] An image analysis-based detection feedback control method for detecting feedback of cylinder acceleration and tool tightness on potato slicing includes the following steps:

[0012] Step 1: Build a video monitoring system to obtain an image dataset of at least 1,000 potato slice sections;

[0013] Step 2: Build a target detection and instance segmentation system;

[0014] Step 3: Train the object detection and instance segmentation system using the image dataset to obtain the best object detection and instance segmentation model;

[0015] Step 4: Input the image data to be detected into the object detection and instance segmentation system. According to the current scene and requirements, a corresponding trained object detection model is selected as the best model. After the image data to be detected is detected by the best model, the final instance segmentation result is obtained.

[0016] Step 5: The continuously generated target detection results are accumulated and integrated through the cumulative confidence algorithm to obtain the final instance segmentation result;

[0017] Step 6: Calculate the potato fragmentation rate of the current frame based on the final instance segmentation result of the object detection and instance segmentation system;

[0018] Step 7, set the dynamic threshold value of the potato chip fragmentation rate. When the fragmentation rate is ≤3%, the cutting is qualified. When the fragmentation rate is >3%, an alarm is triggered. It is necessary to detect the tightness of the tool and the acceleration of the cylinder, and reset the tightness of the tool and the acceleration of the cylinder according to the fragmentation rate.

[0019] Wherein, the step 3 specifically includes:

[0020] Step 31, configure the training parameters in the Mask R-CNN model;

[0021] Step 32: input the training set into the Mask R-CNN model;

[0022] Step 33: During the training process, a series of high-quality candidate regions are generated through the region proposal network (RPN). The candidate regions are mapped to the feature map, and a fixed-size feature vector of the corresponding region is extracted from the feature map.

[0023] Step 34: Input the feature vector obtained in step 33 into the fully connected layer to perform target classification and bounding box regression respectively, and select candidate regions with scores higher than the threshold as detection results;

[0024] Step 35: Input the feature vector obtained in step 33 into a mask branch network. The mask branch network performs pixel-level segmentation on each candidate region and predicts the probability of each pixel belonging to the target or background, thereby generating a binary mask of the same size as the input image. The binary mask is fused with the target detection result to obtain the instance segmentation result.

[0025] Step 36: Monitor the training process, input the validation set for verification, observe the performance of the lightweight object detection model and the complete object detection model on the validation set, such as accuracy, recall, precision, and other indicators, and evaluate the generalization capabilities of the lightweight object detection model and the complete object detection model respectively through the validation set;

[0026] Step 37: Repeat steps 32 to 36 for each sub-dataset one by one, until it is repeated 5 times, and average the accuracy, recall, precision and other indicators to obtain the final evaluation result of the model performance;

[0027] Step 38: When the region proposal network, lightweight target detection model, and complete target detection model achieve satisfactory performance, the corresponding trained region proposal network, lightweight target detection model, and complete target detection model are saved.

[0028] Wherein, the step 4 specifically includes:

[0029] Step 41: input the test set into the best model of the fusion target detection system;

[0030] Step 42: For each frame in the test set, use the best model to perform target detection and inference. Collect and record the target detection results for each frame. The target detection results include the bounding box, category, confidence level, and precise outline of the detected target object.

[0031] Wherein, the step 5 specifically includes:

[0032] Step 51: For each target category to be detected, initialize a cumulative confidence value, which is usually set to 0 or a lower initial value; the cumulative confidence value is used to record the detection status of the target category in consecutive frames;

[0033] Step 52: For each target category, the cumulative confidence is updated according to the target detection result of the current frame; when the cumulative confidence reaches the set threshold, the target category is considered to exist in the current scene;

[0034] Step 53: Fusing the binary mask generated in the current frame with the target detection result;

[0035] Step 54: Output the final instance segmentation result based on the accumulated confidence and the fused target detection results.

[0036] Wherein, step 6 comprises: accumulating the total area of ​​the potato slice micro-crack region and the fragment region in the current frame according to the binary mask output by the object detection and instance segmentation system, and calculating the potato slice fragmentation rate, and the calculation formula is as follows:

[0037]

[0038] The theoretical total cross-sectional area is the cross-sectional area of ​​the third discharge port.

[0039] The beneficial effects of the present invention are:

[0040] ① The present invention first uses a screening unit to screen out potatoes of appropriate size. The potato slices discharged from the third discharge port are vertically cut by a tool controlled by a cylinder, thereby automatically screening out suitable materials for cutting. An industrial camera is also used in combination with an image analysis-based detection feedback control method to provide feedback on the cylinder acceleration and tool tightness. The cylinder acceleration and the tightness of the tool fixation are repeatedly adjusted to reduce the fragmentation rate of the slices and allow the tool to quickly cut out complete potato slices.

[0041] ② The present invention achieves an online detection accuracy of 100,000 pieces per hour based on the image analysis detection feedback control method, which is 200 times more efficient than traditional manual sampling and reduces raw material loss by more than 30%. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 The figure is a schematic diagram of the three-dimensional structure of a potato chip production device provided in Example 1 of the present invention.

[0044] Figure 2 1 is a schematic diagram of the principle of a potato chip production device provided in Example 1 of the present invention.

[0045] Figure 3 1 is a schematic diagram of the structural decomposition of a knife and a knife holder of a potato chip production device provided in Example 1 of the present invention.

[0046] Figure 4 This is a structural diagram of a detection feedback control method based on image analysis provided in Example 2 of the present invention.

[0047] Figure numerals: machine body 10, screening unit 20, cylinder 30, tool 40, tool holder 50, industrial camera 60, first screen 201, second screen 202, connecting rod assembly 203, vibration motor 204, first side 101, first discharge port 102, second discharge port 103, second side 104, third discharge port 105. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0049] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0050] Example 1

[0051] Reference Figures 1 to 3 As shown, this embodiment provides a specific implementation method for a potato chip production device, including a body 10 and a screening unit 20 installed on the body 10;

[0052] Among them, the screening unit 20 includes a first screen 201 and a second screen 202 parallel to each other, dividing the internal space of the body 10 from top to bottom into a first screening chamber, a second screening chamber, and a third screening chamber. The first screen 201 is located above the second screen 202, and the density of the first screen 201 is less than the density of the second screen 202, thereby screening out medium-sized potatoes. The potatoes and potatoes in the above and below have the same meanings. The first side 101 of the body 10 is provided with a first discharge port 102 and a second discharge port 103, and the second side 104 of the body 10 is provided with a third discharge port 105. The first discharge port 102 is used to discharge materials in the first screening chamber, the second discharge port 103 is used to discharge materials in the second screening chamber, and the third discharge port 105 is used to discharge materials in the second screening chamber. Through the above structure, potatoes of medium size are screened out for slicing.

[0053] Furthermore, the first screen 201 is located on the second screen 202 and is mechanically connected to the vibration motor 204 located on the bottom side of the body 10 through a connecting rod assembly 203, wherein the connecting rod assembly 203 described here includes a plurality of support rods, which can transmit the vibration generated by the vibration motor 204, thereby driving the first screen 201 and the second screen 202 to vibrate.

[0054] The present invention also includes a cylinder 30, a tool 40, a tool holder 50 and an industrial camera 60. The cylinder 30 is located on one side of the machine body 10 and is also located above the third discharge port 105. The output end of the cylinder 30 is mechanically connected to the tool holder 50. The mechanical connection here includes but is not limited to couplings, flanges, bolts and other fixing methods.

[0055] Furthermore, the tool holder 50 is detachably fixedly connected to the tool 40. The tool mentioned here is not limited to a curved tool, a single-edged tool or a double-edged tool. An acceleration sensor is installed on the cylinder head of the cylinder 30 or the side of the cylinder body of the cylinder;

[0056] In one specific embodiment, the tool 40 is locked on the tool holder 50 by a bolt. Later, the contact pressure between the adjusting bolt and the tool 40 is adjusted to adjust the tool 40 and further adjust the shearing force of the tool 40.

[0057] Furthermore, a ring-shaped LED light is installed on the machine body 10 around the third discharge port 105 .

[0058] In this embodiment, there is only one industrial camera 60 , which faces the cross section of the potato slice to capture high-tech video images.

[0059] In other embodiments, there are two or more industrial cameras 60 that are fixed at multiple angles to shoot cross-sectional video images of potatoes.

[0060] The present invention further includes a controller 70 for controlling the acceleration, vibration frequency, etc. of the cylinder 30 .

[0061] In summary, the present invention first uses the screening unit 20 to screen out potatoes of appropriate size, and the potato slices coming out of the third discharge port 105 are vertically cut by the tool 40. The tool 40 is controlled by the cylinder 30, so as to automatically screen out suitable materials for cutting. An industrial camera is also used in combination with a feedback control method based on image analysis detection to provide feedback on the acceleration of the cylinder 30 and the tightness of the tool. The acceleration of the cylinder 30 and the tightness of the tool are repeatedly adjusted to reduce the fragmentation rate of the slices and allow the tool to quickly cut out complete potato slices.

[0062] Example 2

[0063] See also Figures 1 to 4 This embodiment provides a specific implementation method for a detection feedback control method based on image analysis, which is based on the potato chip production equipment described in Example 1, including the following steps:

[0064] Step S1, obtaining an image dataset of at least 1000 potato slice sections;

[0065] In an embodiment, step 2 specifically includes:

[0066] Step 11: Setting up an industrial camera on or near a potato chip production device to obtain a video file of the initial potato cross section after cutting;

[0067] Step 12: Preprocess the target number of video files to obtain an image dataset; the details are as follows:

[0068] Decompose the video file into multiple separate image frames through video processing (such as OpenCV). Each frame is used as a time point in the video and as input to the target detection framework. By preprocessing a large number of video files into individual image frames, the data is amplified and can be used as the dataset required by the target detection framework.

[0069] By performing image scaling and color channel conversion on the image frame, the image frame is scaled to s*s (e.g., 800*800) pixels and channels are in RGB format; normalization processing can be facilitated by scaling the pixel values ​​to a specific range (e.g., 0 to 1);

[0070] Each image frame is normalized to obtain multiple processed image frames as an image dataset; normalization helps reduce noise in the data, simplifies the image to reduce computational complexity, and makes the model easier to process, thereby optimizing model performance;

[0071] Step 14: annotate and enhance each image data in the image dataset;

[0072] Each image data includes a label, wherein the label in this embodiment is box annotation information, and the box annotation information includes a complete potato slice, a slightly cracked potato slice, and a fragmented potato slice.

[0073] Step 15: Randomly divide the labeled image dataset into K sub-datasets, and use a portion of them in turn to divide into training set, test set and validation set according to the set ratio to avoid the independent validation set occupying samples;

[0074] Specifically, the annotated image dataset is divided into 5 sub-datasets, and each small dataset is divided into training set, test set and validation set in a ratio of 7:2:1, and the cycle is repeated 5 times. In this example, the training set, test set and validation set are all composed of aligned RGB images and infrared images, that is, the imaging angle, position and visual range of the RGB images and infrared images in the training set, test set and validation set are the same.

[0075] Step 2: Build a target detection and instance segmentation system;

[0076] In step 21, the object detection system is established based on the Mask R-CNN model, which performs instance segmentation on the image dataset to utilize the insufficient computing resources in industrial production. The network structure of the Mask R-CNN model is a ResNet-50 backbone network with a lightweight valley-backbone structure MobileNetV3 module and a region proposal network (RPN).

[0077] The backbone network uses ResNet-50, which is a network structure that combines ResNet-50 and Feature Pyramid Network (FPN). It is mainly used for computer vision tasks, especially in the fields of object detection and image segmentation.

[0078] The MobileNetV3 adopts Neural Architecture Search (NAS) technology to optimize the network structure through an automated search algorithm. Its main function is to improve the speed and accuracy on mobile devices while maintaining low computing costs and storage requirements.

[0079] The Region Proposal Network (RPN) consists of a convolutional backbone for feature extraction over the entire image and a network head for bounding box recognition and mask prediction.

[0080] Step 3: training the object detection and instance segmentation system using the sub-dataset to obtain different detection models;

[0081] Step 31, configure the training parameters in the Mask R-CNN model;

[0082] Step 32: input the training set into the Mask R-CNN model;

[0083] Step 33: During the training process, a series of high-quality candidate regions are generated by the Region Proposal Network (RPN); the candidate regions are mapped to the feature map, and a fixed-size feature vector of the corresponding region is extracted from the feature map;

[0084] In step 34, the feature vector obtained in step 33 is input into the fully connected layer to perform target classification and bounding box regression respectively. The classification task is to determine the category to which the candidate region belongs, and the regression task is to further accurately adjust the position and size of the bounding box of the candidate region to make it fit the target more closely. Based on the classification results, the candidate regions with scores higher than the threshold are selected as the final detection results.

[0085] Step 35: Input the feature vector obtained in step 33 into a mask branch network. The mask branch network performs pixel-level segmentation on each candidate region and predicts the probability of each pixel belonging to the target or background, thereby generating a binary mask of the same size as the input image. The binary mask is then eroded and expanded to optimize the edge of the mask, making it smoother and more accurate. The binary mask is then fused with the target detection result to obtain the instance segmentation result.

[0086] Step 36: Train and iteratively optimize the lightweight object detection model and the complete object detection model according to the configured model training parameters, and update the weights of the lightweight object detection model and the complete object detection model through backpropagation algorithm and gradient descent;

[0087] Furthermore, by monitoring the training process and inputting the validation set for verification, we observed the performance of the lightweight object detection model and the complete object detection model on the validation set, such as accuracy, recall rate, precision and other indicators, and evaluated the generalization ability of the lightweight object detection model and the complete object detection model respectively through the validation set;

[0088] Step 37: Repeat steps 32 to 36 for each sub-dataset one by one, up to 5 times, and average the accuracy, recall, precision and other indicators to obtain the final evaluation result of the model performance;

[0089] Step 38: When the region proposal network, lightweight target detection model, and complete target detection model achieve satisfactory performance, the corresponding trained region proposal network, lightweight target detection model, and complete target detection model are saved.

[0090] In this embodiment, selecting training sets under different lighting conditions and different backgrounds as preprocessing training sets helps to avoid overfitting of the model. On the other hand, it also avoids the problem of being unable to recognize when a single RGB image is input, thereby ensuring that the relevant detection functions can operate normally.

[0091] In this embodiment, the size is set to 800*800 pixels. When the image size of any input image does not meet the requirement, the image size is directly adjusted to 800*800, and the feature map is sliced ​​for the next step of processing;

[0092] Step 4: Input the image data to be detected into the object detection and instance segmentation system. According to the current scene and requirements, a corresponding trained object detection model is selected as the best model. After the image data to be detected is detected by the best model, the final instance segmentation result is obtained.

[0093] In this embodiment, step 4 specifically includes:

[0094] Step 41: input the test set into the best model of the fusion target detection system;

[0095] Step 42: For consecutive frames in the test set, use the best model to detect and infer the target, collect and record the target detection results of each frame, and the target detection results include the bounding box, category, confidence level, and precise outline of the detected target object.

[0096] Step 5: Accumulate and fuse the continuously generated target detection results through the cumulative confidence algorithm to obtain the final instance segmentation result.

[0097] In this embodiment, step 5 specifically includes:

[0098] Step 51: For each target category to be detected, initialize a cumulative confidence value, which is usually set to 0 or a lower initial value; the cumulative confidence value is used to record the detection status of the target category in consecutive frames;

[0099] Step 52: For each target category, the cumulative confidence is updated according to the target detection result of the current frame; when the cumulative confidence reaches the set threshold, the target category is considered to exist in the current scene;

[0100] Step 53: Fusing the binary mask generated in the current frame with the target detection result;

[0101] Step 54: Output the final instance segmentation result based on the accumulated confidence and the fused target detection results.

[0102] Step 6: Based on the binary mask output by the object detection and instance segmentation system, the total area of ​​the potato slice micro-crack area and the fragment area in the current frame is accumulated to calculate the potato slice fragmentation rate. The calculation formula is as follows:

[0103]

[0104] The theoretical total cross-sectional area is the cross-sectional area of ​​the third discharge port.

[0105] Step 7, set the dynamic threshold value of the potato chip fragmentation rate. When the fragmentation rate is ≤3%, the cutting is qualified, that is, the tightness of the tool and the acceleration of the cylinder are qualified; when the fragmentation rate is >3%, an alarm is triggered, and it is necessary to detect the tightness of the tool and the acceleration of the cylinder, and reset the tightness of the tool and the acceleration of the cylinder according to the fragmentation rate.

[0106] In subsequent production, it is necessary to actively collect error samples and add them to the training set to expand the diversity of sample data. It is also necessary to perform incremental training on the model every month to ensure the target detection and segmentation system

[0107] In summary, this system achieves an online inspection accuracy of 100,000 pieces per hour, which is 200 times more efficient than traditional manual sampling and reduces raw material loss by more than 30%. During actual deployment, attention should be paid to ambient temperature and humidity control. The recommended ambient temperature is 25°C and the ambient humidity is RH40% to prevent starch dehumidification from affecting imaging quality.

[0108] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A potato chip production device, characterized in that: It comprises a machine body (10) and a screening unit (20) installed on the machine body (10); The present invention also includes a cylinder (30), a tool (40), a tool holder (50) and an industrial camera (60). The cylinder (30) is located on one side of the machine body (10) and is also located above the third discharge port (105). The output end of the cylinder (30) is mechanically connected to the tool holder (50), and the tool holder (50) is detachably fixedly connected to the tool (40).

2. A potato chip production device according to claim 1, characterized in that: The tool (40) is fastened to the tool holder (50) by means of bolts.

3. A potato chip production device according to claim 1, characterized in that: An annular LED light is installed on the machine body (10) around the third discharge port (105).

4. A potato chip production device according to claim 1, characterized in that: The industrial camera (60) is one in number and is used to shoot a front video file of a cross section of a potato slice.

5. The potato chip production equipment according to claim 1, wherein: The invention also includes a controller (70) for controlling the acceleration and vibration frequency of the cylinder (30).

6. A method for detecting feedback based on image analysis, characterized in that: The method is used to detect the feedback of the acceleration of the cylinder (30) and the tightness of the cutter (40) on potato slicing, and comprises the following steps: Step 1: Build a video monitoring system to obtain an image dataset of at least 1,000 potato slice sections; Step 2: Build a target detection and instance segmentation system; Step 3: Train the object detection and instance segmentation system using the image dataset to obtain the best object detection and instance segmentation model; Step 4: Input the image data to be detected into the object detection and instance segmentation system. According to the current scene and requirements, a corresponding trained object detection model is selected as the best model. After the image data to be detected is detected by the best model, the final instance segmentation result is obtained. Step 5: The continuously generated target detection results are accumulated and integrated through the cumulative confidence algorithm to obtain the final instance segmentation result; Step 6: Calculate the potato fragmentation rate of the current frame based on the final instance segmentation result of the object detection and instance segmentation system; Step 7, set the dynamic threshold value of the potato chip fragmentation rate. When the fragmentation rate is ≤3%, the cutting is qualified. When the fragmentation rate is >3%, an alarm is triggered. It is necessary to detect the tightness of the tool and the acceleration of the cylinder, and reset the tightness of the tool and the acceleration of the cylinder according to the fragmentation rate.

7. The image analysis detection feedback control method according to claim 6, characterized in that: The step 3 specifically includes: Step 31, configure the training parameters in the Mask R-CNN model; Step 32: input the training set into the Mask R-CNN model; Step 33: During the training process, a series of high-quality candidate regions are generated by the Region Proposal Network (RPN); the candidate regions are mapped to the feature map, and a fixed-size feature vector of the corresponding region is extracted from the feature map; Step 34: Input the feature vector obtained in step 33 into the fully connected layer to perform target classification and bounding box regression respectively, and select candidate regions with scores higher than the threshold as detection results; Step 35: Input the feature vector obtained in step 33 into a mask branch network. The mask branch network performs pixel-level segmentation on each candidate region and predicts the probability of each pixel belonging to the target or background, thereby generating a binary mask of the same size as the input image. The binary mask is fused with the target detection result to obtain the instance segmentation result. Step 36: Monitor the training process, input the validation set for verification, observe the performance of the lightweight object detection model and the complete object detection model on the validation set, such as accuracy, recall, precision, and other indicators, and evaluate the generalization capabilities of the lightweight object detection model and the complete object detection model respectively through the validation set; Step 37: Repeat steps 32 to 36 for each sub-dataset one by one, until it is repeated 5 times, and average the accuracy, recall, precision and other indicators to obtain the final evaluation result of the model performance; Step 38: When the region proposal network, lightweight target detection model, and complete target detection model achieve satisfactory performance, the corresponding trained region proposal network, lightweight target detection model, and complete target detection model are saved.

8. The image analysis detection feedback control method according to claim 6, characterized in that: The step 4 specifically includes: Step 41: input the test set into the best model of the fusion target detection system; Step 42: For consecutive frames in the test set, use the best model to detect and infer the target, collect and record the target detection results of each frame, and the target detection results include the bounding box, category, confidence level, and precise outline of the detected target object.

9. The image analysis detection feedback control method according to claim 6, characterized in that: The step 5 specifically includes: Step 51: For each target category to be detected, initialize a cumulative confidence value, which is usually set to 0 or a lower initial value; the cumulative confidence value is used to record the detection status of the target category in consecutive frames; Step 52: For each target category, the cumulative confidence is updated according to the target detection result of the current frame; when the cumulative confidence reaches the set threshold, the target category is considered to exist in the current scene; Step 53: Fusing the binary mask generated in the current frame with the target detection result; Step 54: Output the final instance segmentation result based on the accumulated confidence and the fused target detection results.

10. The image analysis detection feedback control method according to claim 6, characterized in that: Step 6 includes: accumulating the total area of ​​the potato slice micro-crack area and the fragment area in the current frame according to the binary mask output by the object detection and instance segmentation system, and calculating the potato slice fragmentation rate, and the calculation formula is as follows: The theoretical total cross-sectional area is the cross-sectional area of ​​the third discharge port.

Images