A peanut shelling quality image recognition system and method capable of realizing multi-precision measurement

Abstract

The application provides a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling quality image recognition system and method capable of realizing multi-precision measurement, and belongs to the technical field of agricultural product detection. The application discloses a peanut shelling

Description

Technical Field

[0001] This invention relates to the field of agricultural product testing technology, specifically to a peanut shelling quality image recognition system and method capable of multi-precision measurement. Background Technology

[0002] Peanut shelling is a crucial step in the peanut processing industry chain, and its quality directly affects subsequent peanut kernel grading, deep processing, and edible quality. During the shelling process, factors such as improper equipment parameter settings, fluctuations in raw material moisture content, or excessive mechanical impact can easily cause peanut kernels to break, be damaged, or have their skins scratched. Incomplete shelling and shell-clamping can also occur. Therefore, rapid and accurate detection and evaluation of peanut shelling quality are of great significance for optimizing the shelling process and controlling product quality.

[0003] Currently, the detection of peanut shelling quality mainly relies on the following technical methods:

[0004] Firstly, manual sampling inspection. Traditional shelling quality assessment mainly relies on manual visual observation and sorting. Inspectors randomly sample the shelled products, visually identifying broken, incompletely shelled, and intact particles, and then calculating the percentage of each category. This method is inefficient, highly subjective, lacks standardized criteria, and is an offline sampling inspection, making it impossible to achieve real-time quality monitoring of the production line and failing to meet the timeliness requirements of modern automated production for quality control.

[0005] Secondly, there are conventional machine vision-based inspection systems. With the maturity of image processing technology, some peanut shelling quality inspection devices have incorporated machine vision solutions. These systems typically lay the shelled peanut mixture on a conveyor belt, where an industrial camera captures images at a fixed station. Image segmentation and recognition algorithms then statistically analyze indicators such as breakage rate and incomplete shelling rate. However, such systems have significant limitations: peanuts often stack or stick together on the conveyor belt, causing mutual occlusion of targets in the image and reducing detection accuracy; simultaneously, peanut kernels are randomly distributed on the conveyor belt, meaning the camera can only acquire images from a single top-down angle. Many breaks, cracks, or skin damage occur precisely on the sides or bottom of the peanut kernels, and a single-view image cannot capture all of them, leading to missed detections and statistical bias.

[0006] Thirdly, there are specialized testing devices for specific indicators. For example, for detecting peanut kernel breakage rate, there are already technical solutions based on weighing and sieving or near-infrared spectroscopy analysis; for detecting peanut integrity indicators, a dedicated peanut image database can be built, and convolutional neural network (CNN) models can be trained and optimized to effectively improve the accuracy and real-time performance of peanut integrity detection; for detecting incomplete shelling rate, there are also devices that use color sorting principles for sorting. However, these technical approaches are often single-function, only detecting a specific quality indicator, making it difficult to achieve a comprehensive multi-dimensional assessment of shelling quality within a single system.

[0007] More importantly, in actual peanut shelling quality testing scenarios, different process control requirements correspond to different precision requirements. In the rough inspection stage, it's necessary to quickly determine whether the overall quality of the shelled product meets standards, focusing on comprehensive indicators such as breakage rate and incomplete shelling rate at the macroscopic statistical level. However, in the fine analysis stage, it's necessary to accurately identify minute cracks, hidden cracks, and tiny shell-clamping particles in individual peanut kernels, placing higher demands on detection accuracy and imaging detail. However, most existing detection systems use fixed image acquisition resolutions and recognition strategies, unable to flexibly switch detection precision according to testing needs. This results in either excessive computational resource consumption and slow processing speed in routine testing, or insufficient image clarity and recognition capability in fine testing.

[0008] In summary, the existing peanut shelling quality detection technology has the following shortcomings: (1) The detection method is simple, mostly static single-angle imaging, which makes it difficult to fully capture the appearance quality information of the shelled product; (2) There is a lack of effective separation and clear imaging methods for the mutual occlusion of peanut kernels and peanut shell fragments in the stacked state; (3) The detection accuracy is fixed, and it is impossible to achieve multi-precision flexible measurement from macroscopic statistics to microscopic identification in the same system, making it difficult to balance the dynamic balance between detection speed and detection accuracy. Summary of the Invention

[0009] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies by providing a peanut shelling quality image recognition system and method capable of multi-precision measurement. This invention integrates three modes on the same device: low general precision stacking detection, medium-to-high precision discrete tiling detection, and high precision dynamic flipping detection. By integrating the stacking detection mode, the dynamic flipping detection mode, and the clear detection mode under a ring light source, and constructing a multi-precision measurement mechanism, it can quickly obtain macroscopic statistical indicators of shelling quality and perform high-precision fine review of abnormal samples, effectively overcoming the technical defects of traditional detection systems that are single in function and have fixed precision.

[0010] Technical solution: The peanut shelling quality image recognition system of the present invention, which can realize multi-precision measurement, includes a frame, on which a general precision stacking detection system, a high precision dynamic flipping detection system and a medium-to-high precision discrete tiling detection system are set.

[0011] The general-precision stacking detection system is located on the left side of the worktable of the rack, used to acquire images of peanut kernels in a stacked state, enabling rapid detection of shelling quality such as impurity rate, breakage rate, and surface damage rate. The medium-to-high precision discrete tiling detection system is located on the right side of the worktable of the rack, used to acquire images of peanut kernels in a discrete tiling distribution. The high-precision dynamic flipping detection system is located in the middle of the worktable of the rack, used to drive the peanut kernels to flip and acquire surface images of the peanut kernels from all directions. The control system is electrically connected to the general-precision stacking detection system, the medium-to-high precision discrete tiling detection system, and the high-precision dynamic flipping detection system, respectively, and is used to control each system to work together in a predetermined sequence.

[0012] The rack serves as the supporting foundation, with three independent detection systems arranged in sections on the worktable above it. The three systems share the same rack and the same interface, but each has its own independent image acquisition channel and detection logic, all centrally coordinated by the control system. After the operator issues a detection command through the interface, the control system, based on the selected detection mode, activates the corresponding detection system to perform image acquisition and data processing, ultimately transmitting the detection results back to the interface for display and export. The three systems can operate independently or be combined according to the required detection accuracy, thus enabling flexible multi-precision measurements from coarse to fine inspection.

[0013] Furthermore, the general precision stacking inspection system includes a feed hopper fixed above the frame, a transparent acrylic material box located below the feed hopper's outlet, a light source channel connected to the transparent acrylic material box facing the first inspection camera, and a first inspection camera mounted on a second support frame. The side wall of the transparent acrylic material box facing the first inspection camera serves as the observation surface, and LED intelligent color-adjusting light sources are symmetrically installed on both sides of the observation surface. After being powered on, the LED intelligent color-adjusting light sources illuminate the inside of the material box evenly. The second support frame is equipped with a slider and a longitudinal slide rail, allowing the first inspection camera to move up and down along the second support frame and rotate 360°. The second support frame is detachably connected to the worktable by bolts, and several mounting holes are provided on the worktable. The second support frame can be adjusted to move left and right along the worktable by selecting different mounting holes.

[0014] The feeding hopper is vertically fixed above the frame. Peanuts fall from the bottom outlet of the hopper into the transparent acrylic material box, where they naturally accumulate under gravity. LED intelligent color-adjusting light sources are symmetrically installed on the left and right sides of the observation surface of the transparent acrylic material box. When powered on, they emit uniform yellow stripe light into the material box, eliminating shadow areas caused by stacking and ensuring uniform illumination of the surface material of the stacked peanuts. The first detection camera is mounted on the second support frame and can be adjusted up and down via a slider and longitudinal rail. The camera body can rotate 360° to ensure that the camera can be aimed at the observation surface of the transparent acrylic material box at the optimal angle. When switching to stacking detection mode, the first detection camera captures a single image of the surface of the stacked peanuts inside the box. The control system uses an improved watershed algorithm to segment the stacked peanut kernels in the image, extracting the color feature values ​​of each closed region in the RGB and HSV color spaces, and classifying and identifying each region based on the color feature values. In this mode, because the materials are stacked and obstructed by each other inside the material box, the peanut kernels located at the bottom of the inner side of the acrylic material box cannot be imaged, so the detection accuracy is the lowest. However, the image acquisition does not require waiting for the materials to disperse or flip, so the speed is the fastest and it is suitable for rapid rough inspection of large batches.

[0015] Furthermore, the high-precision dynamic flipping detection system includes: multiple parallel conveyor belts, laid out parallel and spaced on the worktable, with materials placed between the gaps of two conveyor belts. The conveyor belts are driven by a first drive motor to transport the materials forward at a uniform speed; a limit stop bar, extending upward from the gap between the two conveyor belts, fixed below the worktable to block the materials and keep them at the shooting position; a dial wheel, arranged laterally below the gap between the conveyor belts, the dial wheel having a hollow cylindrical structure with a fixed shaft passing through its center. The two ends of the fixed shaft are fixed to the support plates of the flipping mechanism on both sides, the fixed shaft remains stationary, and the dial wheel rotates independently around the fixed shaft; the outer wall of the dial wheel is uniformly and integrally formed with several flexible silicone teeth, the teeth are evenly arranged along the circumference of the dial wheel, and three rows are equidistantly arranged in the axial direction of the dial wheel, each row of teeth being evenly distributed along the circumference, the teeth intermittently extending upward from the gap between the conveyor belts to move the materials; and a sprocket drive group, independent of the conveyor belt drive system, used to drive the dial wheel to rotate.

[0016] Furthermore, the gap between the two conveyor belts is 5mm.

[0017] Furthermore, the sprocket drive assembly includes a second drive motor, a driving sprocket, driven sprockets, and an annular drive chain; the second drive motor is fixedly mounted on the support plate of the tilting mechanism, and its output shaft is connected to the driving sprocket; a driven sprocket is welded to one axial end of each derailleur; the annular drive chain surrounds the driving sprocket and all the driven sprockets; when the second drive motor is running, the driving sprocket synchronously drives all the driven sprockets and their corresponding derailleurs to rotate continuously in the same direction through the annular drive chain.

[0018] Furthermore, the high-precision dynamic flipping detection system also includes a crossbar support, which includes two legs symmetrically arranged on both sides of the conveyor belt. The two legs are fixedly connected by a set of crossbars. The crossbars are horizontally arranged above the conveyor belt along the conveying direction. The spacing between each pair of adjacent crossbars is 25mm, and the bottom of each crossbar is 1mm above the upper surface of the conveyor belt. When the peanut kernels are conveyed with the conveyor belt, they are clamped between two adjacent crossbars. The crossbars are used to restrict the left and right movement of the peanut kernels in the width direction of the conveyor belt.

[0019] Furthermore, the medium-to-high precision discrete tiling detection system includes a third support frame fixed on the worktable. The third support frame is equipped with a second detection camera, a ring light source, a discrete bearing detection table, and a bottom transmissive surface light source from top to bottom. The second detection camera is adjustable in height, the ring light source provides supplementary lighting from top to bottom, and the bottom transmissive surface light source transmits light from bottom to top. The two light sources work together to acquire images of the tiling material.

[0020] During the inspection process, peanuts are manually laid out evenly on the surface of the discrete bearing inspection platform, with no stacking or obstruction between the materials. The top surface of each material is fully exposed to the field of view of the second inspection camera. A bottom-transmitting light source emits transmitted light upwards, which passes through the discrete bearing inspection platform and illuminates the bottom of the material. For materials with good light transmittance, the light can penetrate the interior, revealing differences in transmittance in internal cracks or hollow areas in the image. A top-ring light source emits a ring of light downwards, providing uniform illumination to the material surface and eliminating bright spots caused by localized reflections. The combination of dual light sources allows the second inspection camera to capture images of the laid-out material vertically downwards, exhibiting both clear surface details and discernible internal structures. In this mode, the material is in a static, discrete state, without stacking or flipping, resulting in stable and reliable image acquisition and moderate inspection accuracy. However, for peanuts, only surface defects can be detected; bottom damage cannot be detected.

[0021] A method using the above-mentioned image recognition system includes three detection modes: stacked state detection, discrete distribution state detection, and dynamic alternating rotation state detection.

[0022] Stacked state detection includes: acquiring images of shelled peanuts in a stacked state, segmenting the image using an improved watershed algorithm to separate the stuck peanut kernels, extracting color feature values ​​of each segmented region in the RGB and HSV color spaces, and classifying and identifying each region based on the color feature values;

[0023] Discrete distribution detection includes: acquiring original images of shelled peanuts under discrete distribution conditions, using annotation tools to annotate the bounding boxes of targets in the images and constructing a dataset, using the YOLO v11 model to train the dataset, inputting the images to be detected into the trained model for inference, outputting the bounding box coordinates, category and confidence score of each target, counting the number and proportion of each type of target and determining the shelling quality in combination with a preset damage rate threshold.

[0024] Dynamic alternating rotation detection includes: acquiring video image sequences of peanut kernels during the dynamic flipping process; using the YOLO v11 model to perform target detection on each frame; using DeepSORT or ByteTrack algorithms to assign a unique identifier to each detected peanut kernel and track its position changes in consecutive frames; and combining the detection results of the same identifier in multiple consecutive frames to evaluate the overall damage level of each peanut kernel.

[0025] Furthermore, in the stacked state detection, the improved watershed algorithm segments the stacked peanut kernel images based on image gradient information and distance transformation; the color feature values ​​include at least hue, saturation, brightness, and the mean and variance of each RGB channel.

[0026] Furthermore, in discrete distribution detection, the training process of the YOLO v11 model includes: dividing the labeled dataset into training, validation, and test sets in a ratio of 7:2:1 or 8:1:1; iteratively training the model using the training set; optimizing the model parameters through backpropagation of the loss function until the loss function converges; and determining that the shelling quality is unqualified when the statistically obtained proportion of broken peanut kernels exceeds the breakage rate threshold.

[0027] Furthermore, in the dynamic alternating rotation state detection, the DeepSORT or ByteTrack algorithm predicts the position of the peanut in the next frame through Kalman filtering, and performs the association matching between the detection box and the existing trajectory through the Hungarian algorithm; the detection results of the same identifier in multiple consecutive frames are combined as follows: the category of the same target that is consistently determined in multiple consecutive frames is used as the final determination result, and the information of adjacent frames is used to correct the false detection in a single frame.

[0028] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows:

[0029] (1) The present invention integrates three modes on the same device: low-precision stacking detection, medium-high precision discrete tiling detection and high-precision dynamic flipping detection. Users can switch flexibly according to actual detection needs. Stacking mode is selected for large-batch rapid coarse inspection, discrete tiling mode is selected for routine sampling inspection, and dynamic flipping mode is selected for fine full-surface inspection. This realizes multi-precision measurement from coarse inspection to fine inspection, and takes into account the dynamic balance between detection speed and detection accuracy.

[0030] (2) The original "conveyor belt + limit stop + dial wheel" dynamic flipping mechanism of this invention, through the combined action of the friction of the conveyor belt, the blocking force of the limit stop, and the pushing force of the flexible dial wheel, enables the peanut kernel to repeatedly flip at multiple angles during the stopping period. Compared with the existing technology that relies on the natural rolling of the inclined plane or the complex clamping mechanism for flipping, the present invention has a simple structure, stable and controllable flipping, and is compatible with a variety of granular materials. It can ensure that all surfaces of each material are completely covered by the image acquisition device, effectively overcoming the problem of missed detection in traditional single-angle imaging.

[0031] (3) In the medium-to-high precision discrete tiling detection mode, the present invention adopts a dual light source scheme of top ring light source and bottom transmissive surface light source. The top ring light source provides supplementary light from top to bottom to eliminate the reflection on the material surface, and the bottom transmissive surface light source transmits light from bottom to top to make the internal defects of thin skin or semi-transparent materials clearly visible, which significantly improves the image acquisition quality.

[0032] (4) This invention uses a transparent acrylic material box with an LED intelligent color-adjusting light source to uniformly illuminate the peanut kernels, so that the first detection camera can capture images of peanut kernels in a stacked state at one time. It uses an improved watershed algorithm and color feature extraction to achieve rapid initial screening of indicators such as impurity rate and breakage rate. It is suitable for online rapid detection on the production site, with fast detection speed and high efficiency.

[0033] (5) In the dynamic flip detection mode, the present invention adopts a joint framework of YOLO v11 target detection and DeepSORT or ByteTrack multi-target tracking, assigns a unique identifier to each peanut kernel and continuously tracks its position change in consecutive frames, and evaluates the overall damage degree of each peanut kernel by integrating temporal information, effectively filtering out false detections in a single frame and significantly improving detection accuracy and reliability. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the image recognition system in this invention;

[0035] Figure 2 This is a schematic diagram of the general precision stacking detection system in this invention;

[0036] Figure 3 This is a schematic diagram of the structure of the high-precision discrete tiling detection system in this invention;

[0037] Figure 4 This is a schematic diagram of the high-precision dynamic flipping detection system in this invention;

[0038] Figure 5 This is a schematic diagram of the dial in this invention;

[0039] Figure 6This is a schematic diagram of the conveyor belt structure in this invention;

[0040] Figure 7 This is a schematic diagram of the crossbar support structure in this invention;

[0041] Figure 8 This is a schematic diagram of the limiting stop bar in this invention;

[0042] Figure 9 This is a schematic diagram of the structure of the prying teeth in this invention;

[0043] Figure 10 This is a schematic diagram of the sprocket drive assembly in this invention;

[0044] Figure 11 This is a schematic diagram of the frame structure in this invention;

[0045] Figure 12 This is a schematic diagram of the structure of the mobile interactive interface mounting platform in this invention;

[0046] 1. Frame; 11. Worktable; 12. Mobile interactive interface mounting platform; 13. Interactive interface; 2. Low-precision stacking detection system; 21. Feed hopper; 22. Transparent acrylic material box; 23. Light source channel; 24. LED intelligent color-adjusting light source; 3. High-precision dynamic flip detection system; 31. Conveyor belt; 311. First drive motor; 32. Limit stop bar; 33. Dial wheel; 331. Dial tooth; 34. Sprocket drive assembly; 341. Drive sprocket; 342. Sprocket sample wheel; 343. Circular transmission chain; 344. Second drive motor; 35. Fixed through shaft; 36. Horizontal support bracket; 361. Horizontal support; 362. Support leg; 37. Flipping mechanism support plate; 4. Medium-precision discrete tiling detection system; 41. Second detection camera; 42. Circular light source; 43. Discrete load-bearing detection platform; 44. Bottom transmissive surface light source; 45. Third support frame; 51. First detection camera; 52. Second support frame; 53. Vertical slide rail, 54 sliders, 6 mobile power supply, 7 light source brightness adjuster. Detailed Implementation

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described. Example 1

[0048] This embodiment provides a dynamic flipping detection system. The dynamic flipping detection system 3 is arranged in the middle of the worktable 11 and is the core innovative structure of this invention. It is suitable for high-precision full-surface defect detection of various granular materials such as peanuts, broad beans, and corn. It includes multiple parallel conveyor belts 31, limit stops 32, dial wheels 33, sprocket drive group 34, and crossbar support 36.

[0049] Multiple conveyor belts are laid parallel and spaced on the worktable 11. Materials are placed between two conveyor belts; in this embodiment, the gap is 5mm. The conveyor belts 31 are driven forward at a constant speed by a first drive motor 311. A limiting stop bar 32 extends upward from the gap between the two conveyor belts and is fixed to the worktable 11 below. When the material is conveyed to the shooting station by the conveyor belts, the limiting stop bar 32 prevents the material from moving further forward, thus locking the material at a fixed point. The crossbar support 36 includes two symmetrically arranged legs 362, connected by a set of crossbars 361. The distance between every two crossbars is 25mm, and the bottom of each crossbar is 1mm from the conveyor belt. During operation, every two crossbars clamp the material in the middle, preventing it from moving left or right.

[0050] Multiple sets of deflector wheels 33 are arranged horizontally below the conveyor belt gap. Each deflector wheel 33 is a hollow cylindrical structure with a fixed shaft 35 passing through its axis. A chain-tooth-like rotating wheel 342 is welded to one end of each deflector wheel 33. The fixed shaft 35 is inserted into each deflector wheel, and both ends of the fixed shaft 35 are fixed to the support plates 37 of the two-sided flipping mechanism. The fixed shaft 35 remains stationary throughout the entire process, with only the deflector wheel 33 rotating independently. Several flexible silicone teeth 331 are uniformly and integrally formed on the outer wall of the deflector wheel 33. The teeth 331 are evenly arranged along the circumference of the deflector wheel 33, and three rows are equidistantly arranged in the axial direction of the deflector wheel 33. Each row of teeth 331 is evenly distributed along the circumference. The flexible silicone teeth 331 intermittently extend upward from the conveyor belt gap to agitate the material.

[0051] The sprocket drive assembly 34 has a drive system independent of the conveyor belt 31, specifically providing rotational power for the pulleys 33. It includes a second drive motor 344, a drive sprocket 341, a toothed rotating wheel 342, and an annular drive chain 343. The second drive motor 344 is fixedly mounted on the tilting mechanism support plate 37, and its output shaft is connected to the drive sprocket 341. A toothed rotating wheel 342 is welded to one axial end of each pulley 33. The annular drive chain 343 surrounds the drive sprocket 341 and all the pulleys. When the motor operates, the chain synchronously drives all the pulleys to rotate continuously in the same direction. After the material is stopped by the limit stop, the conveyor belts on both sides clamp the material, and the lower teeth continuously move the material, causing it to repeatedly flip at multiple angles under double constraints.

[0052] The interaction between the sprocket drive assembly 34, the derailleur 33, and the derailleur teeth 331, and the material turning mechanism are as follows: The output shaft of the second drive motor 344 drives the active sprocket 341 to rotate. The active sprocket 341 transmits power synchronously to the chain tooth-shaped rotating wheel 342 welded to the end of each set of derailleur wheels 33 through the annular transmission chain 343. The driven sprocket 342 drives the corresponding derailleur wheel 33 to rotate continuously in the same direction around the fixed through shaft 35. Since the derailleur teeth 331 are integrally formed on the outer circumferential surface of the derailleur wheel 33, when the derailleur wheel 33 rotates, the derailleur teeth 331 move in a circular motion around the fixed through shaft 35 synchronously with the derailleur wheel 33. The derailleur teeth 331 continuously extend upward from the gap of the conveyor belt 31, and their extended ends alternately probe into the area where the material is located during the rotation of the derailleur wheel 33.

[0053] The specific process of material flipping is as follows: When the material is conveyed to the position of the limit stop bar 32 by the conveyor belt 31, the limit stop bar 32 blocks the material and stops it at the shooting position. At this time, the conveyor belt 31 continues to operate, generating a continuous horizontal friction force between itself and the bottom surface of the material. This friction force causes the material to tend to continue moving forward with the conveyor belt 31. At the same time, as the teeth 331, which rotate continuously with the turntable 33, move from its lowest point to its highest point, the protruding ends of the teeth 331 gradually contact and embed into the bottom of the material, applying an upward digging force to the material. Under the action of this digging force, the contact point at the bottom of the material is lifted upward, while the sides and top of the material are constrained by the conveyor belt 31 and the limit stop bar 32 respectively and cannot move horizontally. Therefore, the material rotates around the contact edge with the conveyor belt 31 or the fulcrum corresponding to the direction of its own center of gravity offset, thereby achieving flipping. The dial 33 continues to rotate, and the subsequent teeth 331 repeatedly paddle the material. Under the combined action of the horizontal friction force F1 of the conveyor belt 31, the blocking force of the limit bar 32, the lateral limiting force of the crossbar 361, and the upward paddle force F2 of the teeth 331, the material is repeatedly flipped at multiple angles while being stopped by the limit bar (32) until the camera completes the all-round image acquisition of the material.

[0054] When switching to the dynamic flip detection mode, the first detection camera 51 is moved to a vertical position directly above the flip device. The device continuously collects 200-300 frames of all-round images of the material within 3 seconds. The built-in material tracking unit assigns an independent identification mark to each peanut, completely identifying small damage, insect holes, and skin damage hidden on the front and back of the material and the sides. It has the highest detection accuracy and is compatible with a variety of granular grain and oil materials.

[0055] Compared with existing magnetic lifting and turning equipment, existing magnetic structures mostly achieve turning through magnetic adsorption and mechanical structure. The mechanical structure is mostly made of metal, which can easily cause damage to the surface or internal parts of the material, affecting the authenticity of the test results. This structure relies on conveyor belt conveying and bottom flexible dial wheel to achieve turning, without the need for magnetic components. The structure is simple and stable, and the material turning is smooth without jamming. Example 2

[0056] like Figure 1 As shown, this embodiment provides a peanut shelling quality image recognition system capable of multi-precision measurement. The core supporting structure is a frame 1, which is integrally formed as a workbench structure. The flat area on the upper part of the frame 1 is the workbench surface 11. Three independent detection systems are arranged in sections on the workbench surface 11: a general precision stacking detection system 2, a high precision dynamic flipping detection system 3, and a medium-to-high precision discrete flat-lay detection system 4. A mobile interactive interface mounting platform 12 is integrally set on the side of the frame 1. An interactive interface 13 is mounted on the mobile interactive interface mounting platform 12. The mobile interactive interface mounting platform 12 can be pushed and pulled back and forth and rotated at an angle. The operator can complete all operations such as starting and stopping the equipment, switching detection modes, viewing images, and exporting defect data through the interactive interface 13. A mobile power supply 6 and a light source brightness adjuster 7 are set under the frame. The portable power supply 6 is electrically connected to each power-consuming component, providing power support for the entire system and can also be moved outdoors to achieve mobile testing operations; the light source brightness adjuster 7 is a two-channel adjuster, which is electrically connected to the light source of each testing system and is used to independently adjust the brightness of each light source.

[0057] like Figure 2 As shown, the low-precision stacking detection system 2 is fixedly installed on the left side of the workbench 11, and includes a feeding hopper 21, a transparent acrylic material box 22, and an LED intelligent color-adjusting light source 24. The feeding hopper 21 is vertically fixed, and the bottom outlet of the feeding hopper 21 is connected to the top of the transparent acrylic material box 22. The side of the acrylic material box facing the first detection camera 51 is connected to a light source channel 23. Peanuts fall from the feeding hopper 21 and are directly stacked and stored inside the transparent acrylic material box 22. The side wall of the transparent acrylic material box 22 facing the first detection camera 51 serves as the observation surface. The LED intelligent color-adjusting light sources 24 are symmetrically installed on the left and right sides of the observation surface. After the light source is powered on, it shines light evenly into the material box to eliminate the shadows of the stacked materials.

[0058] The first detection camera 51 is mounted on the second support frame 52. The second support frame 52 is equipped with a slider 54 that slides in conjunction with the longitudinal slide rail 53. The first detection camera can move up and down along the support frame, and the camera body supports 360° rotation. The second support frame 52 can also move left and right along the table 11. Specifically, the table 11 is drilled with mounting holes spaced 25mm apart and 4mm in diameter. The second support frame 52 is bolted to the table 11, and its fixed position can be adjusted as needed. When switching to the stacking detection mode, the second support frame 52 is slid to align the first detection camera 51 with the observation surface of the transparent acrylic material box 22, capturing images of the stacked peanuts inside the box in one go. This quickly completes the shelling quality detection, including impurity rate, breakage rate, and surface damage rate. This mode has a fast detection speed, but the detection accuracy is lowest due to material stacking obstruction. It is suitable for large-scale rough inspection on-site.

[0059] like Figures 4-10 As shown, the dynamic flipping detection system 3 is arranged in the middle of the worktable 11. It is the core innovative structure of this invention and is suitable for high-precision full-surface defect detection of various granular materials such as peanuts, broad beans, and corn. It includes multiple parallel conveyor belts 31, limit stops 32, dial wheels 33, sprocket drive group 34, and crossbar support 36.

[0060] Multiple conveyor belts are laid parallel and spaced on the worktable 11. Materials are placed between two conveyor belts; in this embodiment, the gap is 5mm. The conveyor belts 31 are driven forward at a constant speed by a first drive motor 311. A limiting stop bar 32 extends upward from the gap between the two conveyor belts and is fixed to the worktable 11 below. When the material is conveyed to the shooting station by the conveyor belts, the limiting stop bar 32 prevents the material from moving further forward, thus locking the material at a fixed point. The crossbar support 36 includes two symmetrically arranged legs 362, connected by a set of crossbars 361. The distance between every two crossbars is 25mm, and the bottom of each crossbar is 1mm from the conveyor belt. During operation, every two crossbars clamp the material in the middle, preventing it from moving left or right.

[0061] Multiple sets of deflector wheels 33 are arranged horizontally below the conveyor belt gap. Each deflector wheel 33 is a hollow cylindrical structure with a fixed shaft 35 passing through its axis. A chain-tooth-like rotating wheel 342 is welded to one end of each deflector wheel 33. The fixed shaft 35 is inserted into each deflector wheel, and both ends of the fixed shaft 35 are fixed to the support plates 37 of the two-sided flipping mechanism. The fixed shaft 35 remains stationary throughout the entire process, with only the deflector wheel 33 rotating independently. Several flexible silicone teeth 331 are uniformly and integrally formed on the outer wall of the deflector wheel 33. The teeth 331 are evenly arranged along the circumference of the deflector wheel 33, and three rows are equidistantly arranged in the axial direction of the deflector wheel 33. Each row of teeth 331 is evenly distributed along the circumference. The flexible silicone teeth 331 intermittently extend upward from the conveyor belt gap to agitate the material.

[0062] The sprocket drive assembly 34 has a drive system independent of the conveyor belt 31, specifically providing rotational power for the pulleys 33. It includes a second drive motor 344, a drive sprocket 341, a toothed rotating wheel 342, and an annular drive chain 343. The second drive motor 344 is fixedly mounted on the tilting mechanism support plate 37, and its output shaft is connected to the drive sprocket 341. A toothed rotating wheel 342 is welded to one axial end of each pulley 33. The annular drive chain 343 surrounds the drive sprocket 341 and all the pulleys. When the motor operates, the chain synchronously drives all the pulleys to rotate continuously in the same direction. After the material is stopped by the limit stop, the conveyor belts on both sides clamp the material, and the lower teeth continuously move the material, causing it to repeatedly flip at multiple angles under double constraints.

[0063] The interaction between the sprocket drive assembly 34, the derailleur 33, and the derailleur teeth 331, and the material turning mechanism are as follows: The output shaft of the second drive motor 344 drives the active sprocket 341 to rotate. The active sprocket 341 transmits power synchronously to the chain tooth-shaped rotating wheel 342 welded to the end of each set of derailleur wheels 33 through the annular transmission chain 343. The driven sprocket 342 drives the corresponding derailleur wheel 33 to rotate continuously in the same direction around the fixed through shaft 35. Since the derailleur teeth 331 are integrally formed on the outer circumferential surface of the derailleur wheel 33, when the derailleur wheel 33 rotates, the derailleur teeth 331 move in a circular motion around the fixed through shaft 35 synchronously with the derailleur wheel 33. The derailleur teeth 331 continuously extend upward from the gap of the conveyor belt 31, and their extended ends alternately probe into the area where the material is located during the rotation of the derailleur wheel 33.

[0064] The specific process of material flipping is as follows: When the material is conveyed to the position of the limit stop bar 32 by the conveyor belt 31, the limit stop bar 32 blocks the material and stops it at the shooting position. At this time, the conveyor belt 31 continues to operate, generating a continuous horizontal friction force between itself and the bottom surface of the material. This friction force causes the material to tend to continue moving forward with the conveyor belt 31. At the same time, as the teeth 331, which rotate continuously with the turntable 33, move from its lowest point to its highest point, the protruding ends of the teeth 331 gradually contact and embed into the bottom of the material, applying an upward digging force to the material. Under the action of this digging force, the contact point at the bottom of the material is lifted upward, while the sides and top of the material are constrained by the conveyor belt 31 and the limit stop bar 32 respectively and cannot move horizontally. Therefore, the material rotates around the contact edge with the conveyor belt 31 or the fulcrum corresponding to the direction of its own center of gravity offset, thereby achieving flipping. The dial 33 continues to rotate, and the subsequent teeth 331 repeatedly paddle the material. Under the combined action of the horizontal friction force F1 of the conveyor belt 31, the blocking force of the limit bar 32, the lateral limiting force of the crossbar 361, and the upward paddle force F2 of the teeth 331, the material is repeatedly flipped at multiple angles while being stopped by the limit bar (32) until the camera completes the all-round image acquisition of the material.

[0065] When switching to the dynamic flip detection mode, the first detection camera 51 is moved to a vertical position directly above the flip device. The device continuously collects 200-300 frames of all-round images of the material within 3 seconds. The built-in material tracking unit assigns an independent identification mark to each peanut, completely identifying small damage, insect holes, and skin damage hidden on the front and back of the material and the sides. It has the highest detection accuracy and is compatible with a variety of granular grain and oil materials.

[0066] like Figure 3 As shown, the discrete tiling detection system 4 is located on the right side of the workbench 11 and includes a third support frame 45, a second detection camera 41, a ring light source 42, a discrete load-bearing detection stage 43, and a bottom transmissive surface light source 44. The third support frame 45 is vertically fixed to the workbench 11, and the adjustable second detection camera 41, the ring light source 42, the discrete load-bearing detection stage 43, and the bottom transmissive surface light source 44 are assembled sequentially from top to bottom.

[0067] During the inspection, peanuts are manually spread out evenly on the surface of the discrete bearing inspection platform 43, with no material stacking or obstruction. A bottom-transmitting light source 44 illuminates from bottom to top, while a top ring light source 42 provides supplemental lighting from top to bottom. The dual light sources work together to eliminate surface reflections. A second inspection camera 41 captures images of the spread material vertically downwards. This method eliminates the need for a flipping mechanism, resulting in a simple equipment structure and medium-to-high inspection accuracy. It is suitable for materials with good light transmittance, such as rice and thin-skinned granules, where defects can be identified from a single side. However, for peanuts, only surface defects can be detected; bottom damage cannot be detected. Example 3

[0068] This embodiment provides an image recognition method based on Embodiment 2. It includes three detection modes: stacked state detection, discrete distribution state detection, and dynamic alternating rotation state detection. Users can select the appropriate detection mode according to their actual detection needs to achieve multi-precision measurement from low to high accuracy. The specific implementation methods of the three detection modes are described in detail below.

[0069] Implementation process of stacked state online rapid detection

[0070] Stacked state detection relies on a stacked detection system on the rack to achieve rapid online detection on the production line. The detection accuracy is generally average. The entire process includes five steps: hardware imaging, image segmentation, feature extraction, classification and recognition, and quality statistics.

[0071] 1. Material feeding and image acquisition

[0072] The production line conveyor sends the shelled and mixed peanut kernels into the feed hopper of the stacking detection system. The material falls into the transparent acrylic material box by its own weight and stacks naturally. Yellow strip supplementary light sources on both sides of the observation surface are turned on to provide uniform illumination and eliminate shadows on the stacked material. The first detection camera, which can rotate 360° and move horizontally and vertically, is aimed at the observation surface of the transparent acrylic material box to capture the original image of the stacked peanut kernels inside the box and transmit it to the image recognition controller.

[0073] 2. Material segmentation based on an improved watershed algorithm

[0074] The controller retrieves a dedicated image processing program for stacking detection and uses an improved watershed algorithm to segment the acquired images of stacked peanuts, effectively separating the individual peanuts that are stuck together and distinguishing each independent material closed area within the image.

[0075] 3. Color space feature value filtering and extraction

[0076] The controller extracts color feature values ​​from the RGB and HSV color spaces of each segmented closed region image, which are used as the basis for material classification and identification.

[0077] 4. Material classification and identification

[0078] The controller uses the extracted RGB and HSV color feature values ​​as the discrimination criteria to classify and identify each closed region after segmentation, distinguishing four types of targets: whole peanut kernels, broken peanut kernels, broken shells, and external impurities.

[0079] 5. Online quality assessment output

[0080] The controller counts the number of various targets in the image, calculates the impurity rate and rough breakage rate of the material, and pushes the image and detection values ​​to the mobile interactive interface simultaneously. The system has a built-in preset threshold for coarse screening breakage rate. When the detected breakage rate and impurity rate exceed the threshold, the interactive interface will pop up a warning to complete the online rapid screening of the production line.

[0081] Implementation process of discrete distributed offline detection

[0082] Discrete distribution detection relies on a discrete detection system on a rack to achieve offline measurement. It has high detection accuracy and is divided into two fixed stages: the model offline training stage and the equipment offline detection and use stage.

[0083] Phase 1: Offline Construction and Training of the YOLOv11 Detection Model

[0084] 1. Sample Image Acquisition: Collect original images of peanut kernels, including whole peanut kernels, broken peanut kernels, broken shells, and impurities, and construct an image sample library;

[0085] 2. Sample annotation processing: Use the LabelImg annotation tool to draw bounding boxes for peanuts, broken shells, and impurities in all images in the sample library, and annotate the target category labels;

[0086] 3. Dataset partitioning: All labeled sample images are divided into three datasets: training set, validation set, and test set;

[0087] 4. Model Setup and Parameter Configuration: YOLOv11 was selected as the basic framework for object detection, and the basic network parameters and loss function parameters were configured.

[0088] 5. Iterative Model Training: The network is input with the partitioned training set for iterative training, and the network model parameters are continuously optimized through backpropagation of the loss function.

[0089] 6. Model Validation and Solidification: During training, the model recognition accuracy is verified in real time using the validation set. After training, the final accuracy is verified using the test set. The converged model is then solidified to the local storage of the system's image recognition controller.

[0090] Phase Two: Offline Detection Workflow for Discrete Patterns

[0091] 1. Mode switching and hardware startup: The operator switches to the discrete distribution detection mode through the mobile interactive interface. The controller cuts off the power to the first detection camera and the dynamic flip detection system, and starts all the hardware of the discrete detection system.

[0092] 2. Material spreading: Peanut kernels are manually spread out evenly on the discrete load testing platform, with no stacking or mutual obstruction of materials;

[0093] 3. Dual-light source collaborative imaging: The ring light source on the third support frame and the bottom transmission surface light source are turned on simultaneously, and the upper and lower lights work together to improve the contrast of the material outline and the damaged area; the second detection camera vertically downwards to collect a complete image of the flat material and transmit it to the controller;

[0094] 4. Real-time model inference and recognition: The controller calls the locally embedded YOLOv11 model to perform real-time inference calculations on the acquired images, and outputs the bounding box coordinates, target category, and recognition confidence score for each target in the image;

[0095] 5. Automatic shelling quality determination: The controller counts the number and proportion of whole kernels, broken kernels, broken shells, and impurities in the image, retrieves the system's preset damage rate grading threshold, and automatically completes the peanut shelling quality grade determination; the detection images and statistical reports are simultaneously uploaded to the mobile interactive interface for storage and display.

[0096] Implementation process of dynamic alternating rotational offline high-precision detection

[0097] The dynamic alternating rotation detection relies on the dynamic flip detection system on the rack to achieve offline high-precision measurement. It adopts YOLOv11 in conjunction with DeepSORT or ByteTrack multi-frame temporal recognition framework. The complete process is divided into five steps: material flip imaging, frame-by-frame target detection, cross-frame target tracking, temporal information fusion evaluation, and high-precision quality inspection output.

[0098] 1. Imaging of mechanism power-on and dynamic material flipping

[0099] The operator switches to the dynamic flipping detection mode via a mobile interface. The controller activates the dynamic flipping device and the first detection camera's motion mechanism. The drive motor drives the sprocket transmission group to rotate, synchronously driving all the dial wheels to rotate continuously. Two parallel conveyor belts transport peanut kernels to the gap between the two conveyor belts. The material moves forward with the conveyor belts and is stopped at the limit bar position. The conveyor belts laterally clamp the peanut kernels, and the continuously extending teeth on the lower dial wheels cycle and move the material, causing the peanut kernels to complete multi-angle flipping within 1-2 seconds. The first detection camera moves to a position vertically above the flipping device and continuously captures 200-300 frames of video images of the entire peanut kernel flipping process at a high frame rate. All frames are cached in real time to the controller.

[0100] 2. Frame-by-frame target detection (single-frame static information acquisition)

[0101] The controller invokes the YOLOv11 model to perform real-time inference on each frame of the video stream, identifying and locating whole peanut kernels, broken peanut kernels, shell fragments, and impurities within the frame, and obtaining the static position and category information of all targets within each frame as the basis for subsequent tracking and recognition.

[0102] 3. Cross-frame target tracking (a core step in dynamic video processing)

[0103] The controller calls DeepSORT or ByteTrack tracking algorithms to assign a unique ID to each peanut kernel detected and identified in each frame; it continuously tracks the coordinate position changes of the same ID peanut kernel in consecutive frames; when the peanut kernel is briefly occluded, has motion blur, or the material posture is deformed, the tracking algorithm continuously locks onto the same target individual to avoid target loss or mismatch.

[0104] 4. Integration of temporal information and comprehensive assessment of damage level

[0105] The controller collects the detection results of all consecutive frames of images corresponding to the same unique peanut kernel, integrates all surface defect features using multi-frame temporal redundancy information, filters out false detections and missed detections caused by lighting and reflection in single-frame images through multi-frame cross-verification, and outputs the overall damage degree judgment result of the peanut kernel by comprehensively considering all external surface image information of the peanut kernel.

[0106] 5. High-precision quality inspection result output

[0107] The controller summarizes the time-series comprehensive evaluation results of all peanut kernels, accurately calculates the overall breakage rate, damage grade, and impurity content, and generates a high-precision peanut shelling quality inspection report, which is simultaneously pushed to the mobile interactive interface for viewing, storage, and export. This mode can detect hidden minor damage on the sides and bottom of peanut kernels, with the highest detection accuracy, and is also compatible with the detection of various granular materials such as peanuts, broad beans, and corn.

[0108] Unified control logic for the whole machine

[0109] 1. Mode Interlock Control: Hardware program interlock is set for three detection modes: stacked state, discrete distribution state, and dynamic alternating rotation state. At any given time, the controller only allows one detection system to be powered on and running, and the power circuits of the other mechanisms are forcibly disconnected to prevent interference between the mechanisms.

[0110] 2. Mechanism action timing interlock: During the operation of the conveyor belt and dial of the dynamic flipping device, the translation, lifting, and rotation actions of the first detection camera are locked; during the camera's movement and focusing, the flipping transmission mechanism is suspended.

[0111] 3. Unified human-machine interaction: Parameter settings, real-time image preview, quality statistics, historical test records, and quality inspection report export for the three detection modes are all completed through the mobile interactive interface on the side of the rack; the controller has a built-in local storage unit to retain all test images and statistics for each mode for a long time, and the data is not lost after power failure.

[0112] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A peanut shelling quality image recognition system capable of multi-precision measurement, characterized in that: Includes a frame (1), on which a general precision stacking detection system (2), a high precision dynamic flipping detection system (3) and a medium-high precision discrete tiling detection system (4) are provided. The general precision stacking detection system (2) is set in the left side area of ​​the worktable (11) of the frame (1) for image acquisition of peanut kernels in the stacked state, so as to realize rapid detection of shelling quality such as impurity rate, breakage rate and surface damage rate; The medium-high precision discrete tiling detection system (4) is set in the right side area of ​​the workbench (11) of the frame (1) and is used to acquire images of discretely tiled peanut kernels; The high-precision dynamic flipping detection system (3) is set in the middle area of ​​the worktable (11) of the frame (1) and is used to drive the peanut kernels to flip and collect surface images of the peanut kernels from all directions. The system and control system are electrically connected to the general precision stacking detection system (2), the medium-high precision discrete tiling detection system (4) and the high precision dynamic flipping detection system (3), respectively, for controlling each system to work together in a predetermined sequence.

2. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 1, characterized in that: The general precision stacking detection system (2) includes a feed hopper (21) fixed above the frame (1), a transparent acrylic material box (22) located below the discharge port of the feed hopper (21), a light source channel (23) connected to the side of the transparent acrylic material box (22) facing the first detection camera (51), and a first detection camera (51) mounted on a second support frame (52); the side wall of the transparent acrylic material box (22) facing the first detection camera (51) serves as the observation surface, and LED intelligent color-adjusting light sources (24) are symmetrically installed on the left and right sides of the observation surface. The LED intelligent color-adjusting light source (24) illuminates the inside of the material box evenly after being powered on; the second support frame (52) is equipped with a slider (5354) and a longitudinal slide rail (5453), the first detection camera (51) can be raised and lowered along the second support frame (52) and can rotate 360°; the second support frame (52) is detachably connected to the workbench (11) by bolts, and several mounting holes are opened on the workbench (11). The second support frame (52) can be adjusted to move in the left and right directions along the workbench (11) by selecting different mounting holes.

3. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 1, characterized in that: The high-precision dynamic flipping detection system (3) includes: Multiple parallel conveyor belts (31) are laid out parallel and spaced on the workbench (11). The material is placed between the gaps of the two conveyor belts (31). The conveyor belts (31) are driven by the first drive motor (311) to transport the material forward at a uniform speed. The limiting stop bar (32) extends upward from the gap between the two conveyor belts (31), and the lower part of the limiting stop bar (32) is fixed to the worktable (11) to block the material and keep it at the shooting position; A dial wheel (33) is arranged horizontally below the gap of the conveyor belt (31). The dial wheel (33) is a hollow cylindrical structure with a fixed shaft (35) through which the axis passes. The two ends of the fixed shaft (35) are fixed on the support plates (37) of the two-sided flipping mechanism respectively. The fixed shaft (35) remains stationary, and the dial wheel (33) rotates independently around the fixed shaft (35). The outer wall of the dial wheel (33) is uniformly integrally formed with several flexible silicone teeth (331). The teeth (331) are evenly arranged along the circumference of the dial wheel (33), and three rows are equidistantly arranged in the axial direction of the dial wheel (33). Each row of teeth (331) is evenly distributed along the circumference. The teeth (331) intermittently extend upward from the gap of the conveyor belt (31) to move the material. And a sprocket drive assembly (34), independent of the drive system of the conveyor belt (31), for driving the dial wheel (33) to rotate.

4. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 3, characterized in that: The gap between the two conveyor belts (31) is 5 mm.

5. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 3, characterized in that: The sprocket drive assembly (34) includes a second drive motor (344), a drive sprocket (341), a driven sprocket (342), and an annular drive chain (343). The second drive motor (344) is fixedly mounted on the support plate (37) of the flipping mechanism, and its output shaft is connected to the drive sprocket (341). A driven sprocket (342) is welded to one axial end of each dial wheel (33). The annular drive chain (343) surrounds the drive sprocket (341) and all the driven sprockets (342). When the second drive motor (344) is running, the drive sprocket (341) drives all the driven sprockets (342) and the corresponding dial wheels (33) to rotate continuously in the same direction through the annular drive chain (343).

6. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 3, characterized in that: The high-precision dynamic flip detection system (3) also includes a crossbar support (36), which includes two legs (362) symmetrically arranged on both sides of the conveyor belt (31). The two legs (362) are fixedly connected by a set of crossbars (361). The crossbars (361) are horizontally arranged above the conveyor belt (31) along the conveying direction of the conveyor belt (31). The distance between each two adjacent crossbars (361) is 25mm, and the bottom of each crossbar (361) is 1mm above the upper surface of the conveyor belt (31).

7. The peanut shelling quality image recognition system capable of multi-precision measurement according to claim 1, characterized in that: The medium-high precision discrete tiling detection system (4) includes a third support frame (45) fixed on the workbench (11). The third support frame (45) is provided with a second detection camera (41), a ring light source (42), a discrete bearing detection stage (43), and a bottom transmission surface light source (44) from top to bottom. The second detection camera (41) can be raised and lowered. The ring light source (42) provides supplementary light from top to bottom, and the bottom transmission surface light source (44) transmits light from bottom to top. The two light sources work together to collect images of the tiling material.

8. A method using the image recognition system according to any one of claims 1 to 7, characterized in that: It includes three detection modes: stacked state detection, discrete distribution state detection, and dynamic alternating rotation state detection; The stacking state detection includes: acquiring an image of shelled peanuts in a stacked state, segmenting the image using an improved watershed algorithm to separate the sticky peanut kernels, extracting color feature values ​​of each segmented region in the RGB color space and HSV color space, and classifying and identifying each region based on the color feature values; The discrete distribution state detection includes: acquiring original images of shelled peanuts under discrete distribution state, using annotation tools to annotate the bounding boxes of targets in the images and constructing a dataset, using the YOLO v11 model to train the dataset, inputting the image to be detected into the trained model for inference, outputting the bounding box coordinates, category and confidence score of each target, counting the number and proportion of each type of target and determining the shelling quality in combination with a preset damage rate threshold. The dynamic alternating rotation state detection includes: acquiring a video image sequence of peanut kernels during the dynamic flipping process; using the YOLO v11 model to perform target detection on each frame; using the DeepSORT or ByteTrack algorithm to assign a unique identifier to each detected peanut kernel and track its position change in consecutive frames; and comprehensively evaluating the overall damage degree of each peanut kernel by combining the detection results of the same identifier in multiple consecutive frames.

9. The method according to claim 8, characterized in that: In the stacking state detection, the improved watershed algorithm segments the stacked peanut kernel images based on image gradient information and distance transformation; the color feature values ​​include at least hue, saturation, brightness, and the mean and variance of each RGB channel.

10. The method according to claim 8, characterized in that: In the discrete distribution detection, the training process of the YOLO v11 model includes: dividing the labeled dataset into a training set, a validation set, and a test set in a ratio of 7:2:1 or 8:1:1; iteratively training the model using the training set; optimizing the model parameters through backpropagation of the loss function until the loss function converges; and determining that the shelling quality is unqualified when the statistically obtained proportion of broken peanut kernels exceeds the breakage rate threshold.

Images