A multi-stage drop difference discrete and edge cloud collaborative based crop grain intelligent counting and internal quality multi-dimensional measurement system

Abstract

The application discloses a kind of crop kernel intelligent counting and internal quality multidimensional measurement system based on multistage drop difference dispersion and edge cloud cooperation, system includes multi-channel intelligent transfer warehouse machine electric feeding device, multistage ladder dynamics drop difference conveying device, inverted multispectral transmission imaging perception module, deep learning adaptive measurement module and automatic weighing recovery module.Parameter hot load is realized by scanning code and internet of things, adaptive anti-blocking feeding is realized by using rudder rotating baffle and two-stage vibration feeding disc coupling;Grain physical dispersion is realized by using multistage ladder gravity and collision tumbling;By inverted linear array CMOS camera and multi-band transmission light source, the characteristics of external contour and internal transmittance are synchronously acquired when grain falls.Lightweight YOLOv11 network is deployed on edge to accurately detect dense small targets, and gray threshold is combined to realize full and empty grain classification and statistics.The application can realize high-throughput, high-precision and automatic measurement of grain number, grain shape and quality parameters of multiple varieties of crops.

Description

Technical Field

[0001] This invention belongs to the field of agricultural automation, specifically relating to a crop grain intelligent counting and internal quality multidimensional measurement system based on multi-level drop discreteness and edge-cloud collaboration. Background Technology

[0002] my country is the world's largest producer and consumer of grain [1]. The yields of crops such as soybeans, corn, rice, and wheat are directly related to my country's food security. With the development of smart agriculture, grain quality testing has become a key link in the fields of breeding research and precision planting. According to the survey, the commonly used grain counting devices are mainly divided into four categories: traditional manual counting, photoelectric counting, piezoelectric counting, and computer vision counting.

[0003] Although traditional manual counting is inexpensive, it has significant disadvantages such as being time-consuming, labor-intensive, having large subjective errors, and high labor costs, making it difficult to meet the needs of intelligent seed counting in modern agriculture. To solve this problem, researchers have developed a variety of automated devices based on physical sensors: Lu Zexin et al. designed a photoelectric seed counter based on an STC-89C52 microcontroller, which uses counting grooves on the surface of a conveyor belt to transport seeds and collects signals through photoelectric sensors. This method achieves relatively accurate and fast seed counting with minimal damage to the seeds [2]. Zhang Lin et al. proposed a counting system based on piezoelectric sensors, which uses piezoelectric ceramic sheets to detect the weak vibrations generated by seeds hitting a cantilever beam, converts them into voltage waveforms, and processes them into pulse signals for counting [3]. Although the above physical sensor devices have improved the level of automation to a certain extent, they still face problems such as high equipment costs, poor generalization to different crops, limited accuracy, and long time consumption. Therefore, seed counting methods based on computer vision have gradually become a research hotspot in recent years.

[0004] With the development of visual detection technology, scholars at home and abroad have carried out in-depth research from two dimensions: "traditional image processing" and "deep learning object detection". Early research mainly relied on image segmentation and morphological algorithms: Monlica Wattana et al. used distance transformation and region growing to separate soybean seeds, and separated damaged seeds through HSV color space and morphological parameters, which effectively saved counting time [4]. Kang Shiying et al. used a method based on the connected component labeling of binary images to solve the counting problem of corn images with severe adhesion, ensuring high accuracy [5]. Zohaib Mussadiq et al. evaluated the performance of open source image analysis programs such as ImageJ and found that combining the Renyi Entropy thresholding algorithm can significantly improve the accuracy of seed counting in digital camera images [6]. Since Joseph Redmon et al. proposed the YOLO (You Only Look Once) algorithm framework in 2015, it has been widely used due to its efficient real-time detection capability. In response to the special characteristics of grain counting scenarios, researchers have made several improvements to the model [7]: Na Ma et al. proposed a lightweight real-time wheat grain detection model based on YOLOv8 (YOLOv8-HD), which improved the detection performance by introducing a shared convolutional layer and incorporating a vision transformer with a deformable attention mechanism into the backbone network [8]. Pang Zhengjun et al. addressed the problem of severe occlusion and adhesion of wheat grains by adding a CBAM attention module to the backbone network of the YOLOX algorithm and proposed the YOLOX-P model, which effectively solved the problem of unclear feature information [9].

[0005] While existing visual counting methods have achieved some success, they still have the following significant limitations: First, traditional feeding and conveying mechanisms lack physical adaptability to multiple crop varieties, making them prone to bridging blockages or uncontrolled material overflow, and they lack robust physical anti-stacking designs; second, conventional two-dimensional visual imaging can only extract external contours and cannot identify grains with normal projected area but shriveled interiors, lacking methods for evaluating internal quality such as seed fullness; finally, standard target detection algorithms suffer from high false negative rates when dealing with high-throughput, extremely dense, tiny grains due to the disappearance of deep features, and excessive computational redundancy, making it difficult to meet the real-time inference requirements of edge devices. Therefore, to overcome these technical bottlenecks, this invention proposes a crop grain intelligent counting and internal quality multi-dimensional measurement system based on multi-level drop discretization and edge-cloud collaboration. This invention solves the problems of cross-variety adaptation and grain overlap from a mechanical perspective by innovatively designing a multi-channel electromechanical coupling feeding system and a multi-level drop dynamic discrete structure. At the same time, it introduces multi-band adjustable transmission imaging technology and a reconstructed lightweight high-resolution YOLOv11 algorithm to extend the traditional two-dimensional morphological seed assessment to multi-dimensional detection that also includes internal quality evaluation. Furthermore, it relies on the Internet of Things to build a parameter hot reload and data archiving platform, creating an intelligent visual measurement equipment that is not limited by grain type, has high generalization ability, and supports continuous automated detection of multiple batches.

[0006] References Han Guangfu, Han Xiao. A Multidimensional Examination of the Practice of Ensuring National Food Security in the Process of Chinese Modernization [J / OL]. Journal of Jiangsu University (Social Sciences Edition), 1-10 [2026-01-25]. https: / / doi.org / 10.13317 / j.cnki.jdskxb.2026.01. Lü Zexin, Guo Huidong, Wu Heng, et al. Design and experiment of photoelectric digital device [J]. Journal of Tarim University, 2019, 31(01):84-88. Zhang Lin, Zhao Zuoxi, Ke Xinrong, et al. Piezoelectric seed counting system [J]. Transactions of the Chinese Society for Agricultural Machinery, 2011, 42(08):41-45. Wattana M ,Siriluk B ,Khotwit S .Counting and Separating DamagedSeeds of Soybean Seeds using Image Processing[J].International Journal onAdvanced Science, Engineering and Information Technology,2018,8(4):1366-1371. Kang Shiying, Yao Bin. A rapid grain particle counting method based on computer vision [J]. Automation Technology and Application, 2019, 38(08):98-101+106. Mussadiq Z ,Laszlo B ,Helyes L , et al.Evaluation and comparison of opensource program solutions for automatic seed counting on digital images[J].Computers and Electronics in Agriculture,2015,117:194-199. Redmon J ,Divvala KS ,Girshick BR ​​, et al.You Only Look Once:Unified, Real-Time Object Detection.[J].CoRR,2015,abs / 1506.02640. Ma N ,Su Y ,Yang L , et al.Wheat Seed Detection and Counting MethodBased on Improved YOLOv8 Model[J].Sensors,2024,24(5). Pang Zhengjun, Dong Luan, Wen Zhaofa, et al. Automatic seed counting method for crops based on YOLOX [J]. Agricultural Engineering, 2023, 13(01):29-35. Summary of the Invention

[0007] (a) Technical problems to be solved To address the limitations of existing crop visual seed testing equipment in automated adaptation to multiple varieties, physical destacking of seeds, internal quality evaluation, and accurate detection of dense, small targets, this invention aims to systematically solve the following core bottlenecks faced by traditional seed testing technologies: First, it addresses the problems of poor adaptability of traditional single or fixed-diameter feeding devices to crops of different particle sizes, and the high likelihood of mechanical "bridging" blockages or uncontrolled "sponging," thus meeting the needs of continuous automated testing of multiple varieties of samples. Second, it solves the problem that planar or single-stage conveying structures cannot effectively break the mechanical interlocking of seed clusters, utilizing multi-level spatial drop to achieve physical dispersion of seeds. Third, it addresses the problem that conventional two-dimensional visual imaging can only measure external morphology and cannot accurately identify seeds with normal projected area but shriveled interiors, filling the gap in three-dimensional seed fullness evaluation. Fourth, it solves the problems of high false negative rates and low boundary segmentation accuracy of conventional target detection algorithms when facing extremely dense, small seeds or edge-transmitted halos. In summary, by overcoming four major technical challenges—electromechanical collaborative adaptive feeding, multi-level drop dynamic discretization, multispectral transmission feature extraction, and underlying algorithm architecture reconstruction—this invention ultimately achieves high-precision automated measurement of grains from multiple crop varieties throughout the entire process, without damage, resisting stacking, and taking into account both internal and external quality.

[0008] (II) Technical Solution This invention proposes a multi-dimensional measurement system for intelligent counting and internal quality of crop grains based on multi-level drop discrete and edge-cloud collaboration. To address the technical bottlenecks in traditional visual seed testing, such as poor adaptability to multiple varieties, physical seed adhesion, inability to identify shriveled or shriveled grains, and missed detection of densely packed small targets, this invention achieves automated, non-destructive, and stack-resistant automated measurement of grains from multiple crop varieties throughout the entire process. The system mainly includes: a multi-channel intelligent transfer hopper electromechanical collaborative feeding device, a multi-level stepped dynamic drop discrete conveying device, an inverted multispectral transmission imaging and sensing module, a deep learning and internal quality adaptive measurement module for multiple crop varieties, and an automatic weighing and recovery module. The multi-channel intelligent transfer hopper electromechanical collaborative feeding device is used to achieve continuous automatic conveying and flow adaptive control of samples from multiple crop varieties, avoiding frequent manual material changes and mechanical adjustments. This device mainly includes a six-channel rotating hopper, an electromagnetic attraction and release mechanism, a discharge hopper, a turntable drive motor, and a single-chip microcomputer control unit. An innovative electromechanical coupling closed loop, consisting of a rotating baffle driven by a servo motor and a vibrating feeder, is configured at the flow regulation port. After acquiring crop attributes through barcode scanning, the system dynamically matches the physical opening of the servo motor baffle with the excitation frequency and amplitude of the vibrating plate: for large seeds prone to bridging, the opening is increased and matched with low-frequency, large-amplitude vibration; for small seeds prone to spurting, the opening is limited and matched with high-frequency, micro-amplitude vibration, achieving seamless adaptive feeding across varieties. The multi-stage stepped dynamic drop discrete conveying device utilizes spatial kinematics and collision dynamics mechanisms to break the mechanical interlocking between seeds. This device mainly includes a dual-stage vibrating feeder, a vibration controller, a vibrating motor, and a multi-stage stepped guide trough. As seeds flow through the multi-stage steps, on the one hand, gravity nonlinearly amplifies the spatial distance between seeds in the vertical direction; on the other hand, rigid body collisions at the step junctions cause the seeds to reset their posture and randomly tumble, converting translational kinetic energy into lateral discrete kinetic energy, thus achieving a single-seed suspension state without interference before entering the imaging area. The inverted multispectral transmission imaging and sensing module is used to acquire multi-dimensional images with both high-contrast external contours and internal transmittance characteristics during the high-speed fall of grains, while avoiding dust interference. This module mainly includes an inverted high-frame-rate linear array CMOS camera and a top multi-band tunable transmission light source (including visible light and near-infrared arrays). The multi-variety crop deep learning and internal quality adaptive measurement module deploys a deep reconstruction algorithm based on edge computing (Nvidia Jetson) for accurate counting and three-dimensional fullness evaluation.In terms of target detection networks, the basic architecture was upgraded to an improved YOLOv11 network to handle image data streams acquired by high frame rate cameras. First, to adapt to the limited computing power of edge computing industrial control computers and meet the real-time inference requirements of continuous material dropping, a lightweight backbone network based on the UIB (Universal Inverted Bottleneck) module was constructed. The IB module in the UIB variant was introduced to reconstruct the backbone for dimensionality reduction calculation, which significantly reduced the computational complexity at the edge while ensuring high-precision feature extraction capabilities. Second, a high-resolution P2 micro-target detection layer was constructed. In view of the physical phenomenon that dense micro-particles with diameters of 2-5mm are prone to deep feature loss during free fall, high-resolution P2 layer features with 4x downsampling carrying fine-grained spatial information were extracted from the shallow layer of the backbone network. A top-down strategy was adopted to perform channel splicing and fusion with deep semantic features to construct a P2-P5 four-scale detection architecture that is adaptive to multi-variety heterogeneous particle sizes, thereby outputting high-precision target bounding boxes and effective counting results. In terms of morphological processing, the OpenCV algorithm is used to calculate the average transmittance gray value within the connected domain of a single seed. A dynamic threshold is set to accurately identify shriveled seeds with abnormal internal transmittance, achieving independent classification and statistics of shriveled seeds, thus providing reliable support for the system to output high-precision population fullness indicators. The automatic weighing and recovery module is used to collect the tested seeds and obtain the total weight of the sample to convert it into thousand-seed weight. It mainly includes a drawer-type collection bin, a high-precision pressure sensor, and an MQTT data transmission unit. More specifically, the multi-stage drop discrete crop seed intelligent counting and grain shape measurement equipment operates as follows: 1) The operator loads N samples to be tested into the six-channel rotating hopper and scans the barcode. The system automatically downloads the vibration frequency, camera exposure parameters, dedicated detection model, and corresponding multispectral combination and fullness gray value threshold of the corresponding crop from the cloud. 2) The microcontroller controls the turntable to rotate above the discharge port, and the electromagnetic mechanism releases the chassis. The flow regulation servo motor is linked with the vibratory plate to perform adaptive flow regulation according to the cloud parameters, ensuring that the seeds fall into the next stage with the best flow rate and density. 3) Seeds are evenly fed to the inlet of the stepped guide chute via a vibrating feeder. As they flow through multi-stage inclined slides, they are accelerated by gravity and tumble against the step boundaries, rapidly creating physical separation in space. 4) Seeds fall freely through the imaging area. The edge computing unit controls an inverted linear array CMOS camera, working in conjunction with a top multi-band light source, to perform continuous line scanning and acquire transmitted high-brightness images. The system calls upon a depth-reconstructed multi-variety adaptive measurement module and a transmittance extraction algorithm to output in real-time the total number of falling seeds, the number of effective seeds, the percentage of shriveled seeds, and multi-dimensional phenotypic data such as length, width, and two-dimensional area. 5) After imaging, the seeds fall into the weighing and recovery module. The sensor acquires the sample mass, automatically converts it to thousand-seed weight based on the visual effective seed count, and uploads all data to the cloud platform. The hopper chassis resets, the system automatically rotates to the next hopper and reloads parameters, and begins the next round of continuous measurement.

[0009] (III) Beneficial Effects This invention addresses the technical bottlenecks of existing crop seed testing equipment in terms of multi-variety adaptability, stacking and adhesion handling, internal quality evaluation, and limited edge computing power. It proposes a crop grain intelligent counting and internal quality multi-dimensional measurement system based on multi-level drop discretization and edge-cloud collaboration.

[0010] The beneficial effects of this invention are reflected in at least the following four aspects: 1) It realizes electromechanical collaborative adaptive feeding and continuous automated operation for cross-variety crops. This invention innovatively constructs a feeding architecture of "multi-channel intelligent transfer bin + dynamic coupling of servo motor and vibratory feeder". Through cloud parameter distribution, the system can automatically calculate and execute the optimal combination of physical opening and excitation force for crops with different particle sizes and friction coefficients. Through electromechanical collaborative control, the risk of seed blockage and material spillage during the feeding process is effectively reduced. This design eliminates the need for complicated manual mechanical calibration and achieves seamless compatibility and high-throughput continuous operation for all varieties and wide ranges. 2) It overcomes the problem of target occlusion and overlap under high-throughput dense detection from the physical source. This invention abandons the passive approach of simply relying on visual algorithms for overlap segmentation and designs a multi-level stepped drop discrete structure. Utilizing the spatial amplification effect of gravitational acceleration and the high-frequency rigid body collision mechanism at the step junction, the static friction and mechanical interlocking between the grain groups are forcibly destroyed. This purely mechanical physical de-stacking design transforms a dense flow of seeds into a non-interfering array of single seeds without adding an extra power source, significantly reducing the algorithm segmentation pressure on the edge computing end. 3) Breaking through the limitations of two-dimensional morphology, it achieves accurate evaluation of the three-dimensional fullness and internal quality of seeds. Based on conventional size detection, this invention introduces an inverted multispectral adjustable transmission imaging module. Utilizing the differences in absorption and penetration characteristics of different density tissues for transmitted light bands, the transmittance of a single seed (internal grayscale value of the transmission image) is used as a key feature to measure its thickness and fullness. This scheme only requires an extremely low increase in grayscale threshold determination computing power to accurately identify seeds with normal projected area but shriveled interiors, completing a technological leap from phenotypic contour to internal three-dimensional quality evaluation. 4) Significantly improves the detection accuracy and real-time inference efficiency of the edge computing platform for small targets. To address the challenge of identifying dense, tiny particles in high-resolution images, this invention represents a deep architectural-level reconstruction of the target detection algorithm. On one hand, a UIB (Universal Inverted Bottleneck) module is introduced to reconstruct the backbone network. Through adaptive fusion of spatial modeling and channel mixing, the number of model parameters and computational complexity are significantly reduced, resulting in excellent real-time inference capabilities on edge computing platforms. On the other hand, to address the issue of feature vanishing in deep networks caused by extremely small particles, a high-resolution P2 micro-target detection layer with 4x downsampling is innovatively constructed. By extracting and fusing shallow features, the extreme scale mismatch problem is effectively overcome, significantly improving the recall rate and localization accuracy of tiny particles and impurities. This invention establishes a universal technical implementation paradigm, achieving a system-level breakthrough in addressing core technical challenges through precise adaptation of hardware modules and algorithm flows. Its standardized system architecture possesses high integrity and replicability, providing a unified model for subsequent large-scale applications. Attached Figure Description

[0011] Figure 1A schematic diagram of a crop grain intelligent counting and internal quality multidimensional measurement system based on multi-level drop discreteness and edge-cloud collaboration. Figure 2 Overall structural diagram Figure 3 Detailed image of the feeding hopper Figure 4 Workflow diagram of the present invention Figure 5 A multi-crop grain adaptive recognition network based on improved YOLOv11 Figure 6 Flowchart of Crop Grain Morphological Phenotypic Parameter Extraction Algorithm Based on OpenCV Figure 7 Correlation analysis diagram between system measurement of shriveled and shriveled grains and manual measurement Figure 8 Comparison of ablation experimental performance of the improved YOLOv11 algorithm In the figure: stepped guide chute (1), multi-band adjustable transmission light source (2), drawer-type collection bin (3), weighing scale (4), industrial control computer (5), linear array CMOS camera (6), vibration controller (7), vibration motor (8), six-channel rotating hopper (9), electromagnetic release mechanism (10), turntable motor (11), feeding bin (12), vibrating feeder (13), LCD screen (14), barcode scanner (15), servo motor (16), rotating baffle (17).

[0012] 1. Stepped feed chute: The gravity acceleration generated by the multi-stage drop causes the seeds to spread out during the falling process.

[0013] 2. Multi-band adjustable transmission light source: Provides an adjustable visible light and near-infrared composite light field, which, together with the camera, simultaneously acquires the high-contrast external contour and internal light transmittance characteristics of the kernels to accurately identify shriveled kernels.

[0014] 3. Drawer-type collection bin: Located at the bottom of the equipment, it is used to receive and recycle the crop grains after testing. The drawer-type design makes it easy for operators to pull out the bin for emptying and cleaning.

[0015] 4. Weighing scale: After the seeds have been inspected, the total mass is measured in real time to help calculate the weight of a thousand seeds.

[0016] 5. Industrial PC: The core computing hub, responsible for running AI algorithms, processing image data, and coordinating the collaborative work of all hardware.

[0017] 6. Linear CMOS camera: It adopts an inverted layout to perform high-speed line scanning on falling seeds to obtain high-contrast contour images.

[0018] 7. Vibration controller: Adjusts the frequency and amplitude of the vibrating feeder to adapt to the flow characteristics of different crop seeds.

[0019] 8. Vibrating motor: As the power source of the vibrating feeder, it generates high-frequency excitation force to initially break up the agglomerated seeds and drive the seeds to move towards the inlet of the stepped feed chute at a uniform flow rate.

[0020] 9. Six-channel rotary hopper: used for temporary storage of various crop samples in a circular distribution, and the designated hopper is precisely positioned to the unloading station by rotation to achieve continuous feeding.

[0021] 10. Electromagnetic release mechanism: Controlled by system commands, it controls the opening and closing of the bottom of the hopper to achieve on-demand seed dispensing.

[0022] 11. Turntable motor: Controlled by instructions from the industrial control computer, it drives the six-channel rotating hopper to rotate, accurately positioning the designated numbered hopper directly above the unloading station, thus realizing the automatic switching of different sample varieties.

[0023] 12. Feed hopper: As a transitional guide component, it receives the seeds falling from the rotating hopper and guides them to the flow regulation area.

[0024] 13. Vibrating feeder: High-frequency micro-vibration is used to initially break up the accumulated seeds and convey them to the next stage feed chute at a uniform flow rate.

[0025] 14. LCD screen: Human-machine interface, which displays detection images, statistical data and system status in real time.

[0026] 15. Barcode scanner: Scans the QR code on the sample label, triggering the system to automatically load the corresponding detection model and hardware operating parameters.

[0027] 16. Servo motor: As an actuator, it precisely drives the rotating baffle to rotate according to the type of crop, so as to adjust the opening of the feeding channel.

[0028] 17. Rotating baffle: In conjunction with the servo motor, it physically changes the cross-sectional area of ​​the flow channel, thereby controlling the seed flow rate and distribution width falling into the vibratory feeder. Detailed Implementation

[0029] This invention proposes a crop grain intelligent counting and internal quality multidimensional measurement system based on multi-level drop discrete and edge-cloud collaboration. It mainly includes: a multi-channel intelligent transfer electromechanical collaborative feeding device, a multi-level stepped dynamic drop discrete conveying device, an inverted multispectral transmission imaging and sensing module, a multi-variety crop deep learning and internal quality adaptive measurement module, and an automatic weighing and recycling module.

[0030] The operator loads N different varieties of crop samples into the independent compartments of the six-channel rotating hopper. The industrial control computer (5) issues a command to start the turntable motor (11), driving the hopper to rotate around the central axis and precisely position the designated hopper directly above the discharge hopper (12). Each hopper is equipped with an independent electromagnetic release mechanism (10) at the bottom. When the hopper is in position, the electromagnetic attraction mechanism is de-energized and released, the bottom cover opens, and the seeds fall into the discharge hopper (12) below. The discharge hopper (12) has a flow regulation mechanism at the outlet, consisting of a servo motor (16) and a rotating baffle (17). The system innovatively adopts an electromechanical coupling flow control mechanism: the operator scans the sample batch QR code with a barcode scanner (15), the system obtains the crop attributes and communicates with the cloud via the Internet of Things, automatically downloading and calling the matching electromechanical parameters. For large-particle, high-friction crops such as corn, the microcontroller drives the servo motor (16) to increase the effective flow cross-sectional area of ​​the rotating baffle (17) to disrupt the "bridging" geometry, and simultaneously instructs the next-level vibration controller (7) to output low-frequency, high-amplitude excitation force; for small grains such as rapeseed, the baffle opening is reduced and matched with high-frequency micro-amplitude, thereby achieving fully automatic cross-variety anti-blocking and anti-surge adaptive feeding.

[0031] This module is located below the feeding device and consists of a vibrating feeder (13), a stepped guide chute (1), a vibrating motor (8), and a vibration controller (7). Seeds fall into the vibrating feeder (13) after the flow rate is adjusted. The disc is connected to a vibration motor (8) as a power source, and the vibration controller (7) is responsible for adjusting the frequency and amplitude.

[0032] After startup, the high-frequency excitation force initially disperses the seeds and conveys them to the inlet of the stepped feed chute (1) at a uniform flow rate. The stepped feed chute (1) adopts a multi-level height difference design, and its dispersion mechanism includes two physical dimensions: in the vertical direction, using the acceleration due to gravity, adjacent seeds fall with time. As it moves, its spatial spacing follow It exhibits nonlinear amplification; in the multi-level transition zone, when the clustered grains fall, they generate high-frequency rigid body collisions, forcibly disrupting the original coplanar and mechanically interlocked posture of the grain group, forcing it to undergo random tumbling, converting translational kinetic energy into lateral discrete kinetic energy, ensuring that the seeds reach a single-grain suspension state without interference before entering the imaging zone.

[0033] The imaging module is located at the end of the stepped feed chute and is inverted to avoid dust interference. A linear CMOS camera (6) is installed below the falling channel and angled upwards toward the imaging area. A multi-band adjustable transmission light source (2) is configured above the channel, integrating visible light and near-infrared LEDs in the same light-emitting array. Based on the crop parameters sent from the cloud, the system automatically switches electronic bands and adjusts light intensity via PWM signals to provide an adjustable visible light and near-infrared composite light field. When the seed stream falls at high speed through the imaging area, the edge computing terminal controls the linear CMOS camera (6) and the multi-band adjustable transmission light source (2) to perform continuous line scanning synchronously, acquiring multi-dimensional, high-contrast images that include the sharp outline of the outer surface of the seed and the internal transmittance characteristics.

[0034] This module runs on the edge computing end of the industrial control computer (5), and proposes a small target and internal quality detection algorithm based on the improved YOLOv11. In terms of the target detection network: the UIB (general inverted bottleneck) module of MobileNetV4 is introduced to reconstruct the backbone network, and a "narrow-wide-narrow" feature flow structure is adopted to significantly reduce the computational complexity and inference delay at the edge end while ensuring the feature extraction accuracy; secondly, in view of the defect that 32 times downsampling easily leads to the disappearance of small seed features, high-resolution P2 layer features with 4 times downsampling are extracted from the shallow layer of the backbone network, and fused with the P3 layer using a top-down strategy to construct a P2-P5 four-scale detection system, which greatly improves the recall rate of small targets. In terms of internal quality morphology processing, the two-dimensional morphological parameters of the seeds are extracted by combining the OpenCV algorithm. The specific phenotypic extraction process (such as Figure 6As shown below: First, sub-pixel-level edge contour extraction is performed on the connected domain of a single seed. Then, for the extraction of seed length and width, the minimum bounding rectangle of the sub-pixel contour is calculated. Regardless of the tilt angle and posture of the seed in the high-speed falling scene, the longer side of the rotation rectangle is defined as the actual seed length, and the shorter side is the seed width. This method effectively eliminates the two-dimensional projection error caused by the deflection of the seed's falling posture. Finally, for the extraction of seed area and perimeter, the system uses Green's formula to perform polygon area integration on the closed sub-pixel floating-point coordinate sequence to obtain a high-precision absolute seed area. At the same time, the Euclidean distance between adjacent sub-pixels is accumulated to accurately calculate the seed perimeter. Based on the above high-dimensional morphological feature data extraction, the average transmittance grayscale value within the connected domain of a single seed is further extracted. Combined with the transmittance grayscale threshold preset in the cloud, a binarization evaluation is performed to accurately classify and identify shriveled seeds with abnormally high internal transmittance. The cloud-preset transmittance grayscale threshold is set using a combination of prior physical experimental calibration and dynamic compensation: when initializing the equipment or adding new crop varieties, a grayscale mapping relationship model under different internal fullness is established by comparing the transmission imaging and physical slicing / weighing destructive tests on samples of the same batch, thereby calibrating the basic grayscale threshold boundary for distinguishing between full and shriveled states; during continuous system operation, the algorithm further combines the light intensity attenuation feedback of the transmitted light source to dynamically compensate the basic threshold, ensuring that the system maintains extremely high physical accuracy and robustness in classifying full and shriveled grains under long-term operation. The system will record the total number of falling grains and the total number of effective grains of the sample, and the measurement results will be displayed on the LCD screen (14).

[0035] This module is located at the bottom of the equipment and consists of a weighing scale (4) and a drawer-type collection bin (3). The drawer-type collection bin (3) is used to receive the grains after testing, and integrates a weighing scale (4) and a high-precision pressure sensor inside. When the Nth sample is tested and falls into the pan, the sensor obtains the total mass. Combined with the pre-stored mass of the empty pan and the number of grains measured visually, the thousand-grain weight of the grains can be automatically calculated. The system vibration and weighing data streams are uniformly coordinated and processed by the industrial control computer (5).

[0036] Example of the solution: The operator loads N different varieties of crop samples into the independent compartments of the six-channel rotating hopper (9) and scans the corresponding batch's QR code label with the barcode scanner (15). The system automatically downloads various operating parameters corresponding to the crop from the AI ​​cloud via the Internet of Things: including the frequency and amplitude of the vibration controller (7), the opening angle of the flow regulation servo motor (16), the band combination and light intensity of the multi-band adjustable transmission light source (2), the exposure parameters of the linear array CMOS camera (6), the edge-end improved YOLOv11 dedicated detection model, and the transmittance grayscale threshold for judging saturated grains, realizing the automatic initialization of parameters and displaying "ready" on the LCD screen (14). The operator presses the "start" button, and the edge computing industrial control computer (5) controls the turntable motor (11) to start, precisely rotating the Nth hopper to directly above the unloading hopper (12). Once in place, electromechanical coupling flow regulation is performed: the flow regulation servo (16) drives the rotating baffle (17) to rotate to a specified angle according to the preset parameters in the cloud, thus limiting the flow cross-sectional area in physical space; then the electromagnetic release mechanism (10) at the bottom of the hopper is de-energized and released, the chassis opens, and the seeds fall into the vibrating feeder (13) under the action of gravity. At the same time, the vibration controller (7) drives the vibration motor (8) to work, matching the corresponding excitation force, and uniformly conveys the seeds to the inlet of the stepped guide trough (1), thus avoiding the "bridging" blockage of large seeds and the "surging" loss of control of small seeds from the source. When the seeds flow through the multi-level inclined slide, the spacing is nonlinearly amplified in the vertical direction by the acceleration due to gravity, and rigid body collision and tumbling are generated at the junction of the steps, converting translational kinetic energy into lateral discrete kinetic energy, and the spacing is physically stretched in space. The seeds leave the end of the guide trough and enter a free fall state, passing through the imaging area. The edge computing industrial control computer (5) controls the linear array CMOS camera (6) to perform continuous scanning in conjunction with the multi-band adjustable transmission light source (2) at the top, acquiring multispectral images with both external high-contrast contours and internal transmittance characteristics. The acquired image data stream is transmitted to the edge computing industrial control computer (5) in real time, and inference is performed through the deployed adaptive measurement module: at the target detection level, the YOLOv11 algorithm, which reconstructs the UIB lightweight backbone network and the P2 high-resolution micro-target detection layer, is used to accurately locate and recall all micro-seeds; at the internal quality measurement level, the OpenCV algorithm is combined to extract the average transmission gray value in the connected domain of a single seed, and compared with the cloud threshold to accurately identify and independently classify the shriveled seeds lacking dense material. After the vision system determines that all seeds of the Nth sample have passed, they fall into the drawer-type collection chamber (3) at the bottom. The internal weighing scale (4) weighs the sample and automatically calculates the weight of a thousand seeds by combining the effective total number of seeds output by the vision algorithm.Subsequently, the system synchronously displays all multi-dimensional measurement information, such as the total number of falling grains, the number of effective grains, the number of identified shriveled grains, grain length, grain width, two-dimensional area, and apparent thousand-grain weight, on the LCD screen (14), and uploads it back to the AI ​​cloud platform in real time via IoT protocols such as MQTT for data archiving and breeding analysis. After the measurement upload is completed, the Nth hopper chassis is reset and closed. The system automatically controls the turntable motor (11) to rotate to the N+1th hopper position, and the edge computing industrial control computer (5) automatically reloads the electromechanical and algorithm parameters corresponding to the next variety, repeating the above closed-loop operation of "cloud-based distribution - electromechanical collaboration - physical discretization - transmission imaging - AI calculation - cloud upload" until all hopper samples are tested. Finally, the operator pulls open the drawer-type collection hopper (3) and takes out the recovered sample. To verify the accuracy of the system in identifying shriveled grains, a comparative experiment was conducted. The correlation analysis of the shriveled grain identification results output by the system and the manual measurement results was performed, such as... Figure 7 As shown, the two are highly correlated, verifying the effectiveness of the saturated / shriveled particle identification method based on transmission imaging. To verify the effectiveness of the depth reconstruction target detection algorithm, this system conducted a quantitative comparison of ablation experiments on the improved YOLOv11 model (e.g., Figure 8 (As shown). Adding a P2 micro-target detection layer alone to the original baseline model improves recall and mAP, but leads to a significant increase in computational cost (G) and model size (MB), making it difficult to meet the real-time inference requirements of edge devices. While introducing the UIB module alone drastically compresses the model, it results in some loss of accuracy. The final strategy adopted in this invention, simultaneously introducing a lightweight UIB backbone network and a high-resolution P2 detection layer, successfully compresses the model size from approximately 16MB to approximately 3MB while maintaining extremely low overall computational cost, while still outperforming the original baseline model in core accuracy metrics such as precision, recall, and mAP50. This set of quantitative cross-comparison data fully demonstrates that the algorithm architecture of this system perfectly breaks through the computing power limitations of edge computing industrial control computers while ensuring high-throughput, dense micro-target detection accuracy. The specific embodiments described in this application are merely illustrative examples of the spirit of the invention. Those skilled in the art can make various modifications or additions to the described specific embodiments or use similar methods to replace them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A crop grain intelligent counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration, characterized in that, The system includes: a multi-channel intelligent transfer silo electromechanical collaborative feeding device, a multi-stage stepped dynamic drop discrete conveying device, an inverted multispectral transmission imaging and sensing module, a multi-variety crop deep learning and internal quality adaptive measurement module, and an automatic weighing and recycling module. The multi-channel intelligent transfer silo electromechanical collaborative feeding device is used to store multi-variety samples and achieve adaptive flow regulation. It includes a multi-channel rotating silo controlled by a drive motor, a discharge silo below the discharge station, and a flow regulation mechanism driven by an actuator at the discharge silo outlet. The mechanism consists of a rotating baffle driven by a servo motor. It physically adjusts the feed inlet flow rate according to the crop type to prevent blockage and ensure optimal feed density. A multi-stage stepped dynamic drop discrete conveying device is located below the feeding device to physically de-stacking crop grains. It includes a vibrating feeder driven by a vibration controller and a vibration motor, and a stepped guide trough connected to its end. The vibrating feeder uses high-frequency excitation force to initially break up agglomerated grains, while the stepped guide trough utilizes the gravitational acceleration effect generated by the multi-stage height difference to disperse adjacent grains. During free fall, the grains generate a nonlinear velocity difference, forcing them to collide and tumble at the junctions of the steps, thus completely widening the gaps in physical space and forming a single-grain suspension state where they do not interfere with each other. The inverted multispectral transmission imaging and sensing module is located below the end of the multi-stage stepped guide chute, including a dust-proof, inverted high-frame-rate camera and a multi-band adjustable transmission light source located above the falling channel, used to simultaneously acquire transmission contour and internal transmittance feature images as the grains freely fall through the imaging area; the multi-variety crop deep learning and... The internal quality adaptive measurement module runs on an edge computing industrial control computer, which includes a barcode scanner for scanning sample labels. The industrial control computer triggers IoT communication via the barcode scanner to automatically acquire and load matching cloud parameters, extracts multidimensional phenotypic data using a lightweight target detection algorithm and transmission grayscale analysis, and achieves independent classification and statistics of plump and shriveled kernels. The automatic weighing and recovery module is located at the bottom of the system and is used to recover the kernels after testing and obtain the overall mass of the sample. The apparent thousand-kernel weight is calculated by combining the total number of kernels falling from the sample output by the edge computing terminal.

2. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 1, characterized in that: Each independent compartment of the multi-channel rotary hopper is equipped with an electromagnetic release mechanism controlled by the system at its bottom; the flow regulation mechanism is specifically a rotating baffle driven by a servo motor; the feeding device and the vibrating feeding mechanism form an electromechanical coupling collaborative architecture. After receiving the sample tag instruction triggered by the barcode scanner, the system dynamically adjusts the physical opening of the rotating baffle and the excitation frequency and amplitude of the vibrating feeding mechanism in conjunction with the crop attribute parameters downloaded from the Internet of Things cloud, so as to realize cross-variety adaptive continuous feeding with anti-blocking and anti-surge material feeding.

3. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 1, characterized in that: The multi-stage stepped guide trough adopts a structure that combines multi-stage inclined slides with step height differences. It uses gravity acceleration to nonlinearly amplify the spatial distance between adjacent grains in the vertical direction, and forces the grains to undergo rigid body collisions and random tumbling at the step junctions, converting translational kinetic energy into lateral discrete kinetic energy, thereby forming a single-grain suspension state that does not interfere with each other before entering the imaging area below.

4. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 1, characterized in that: The multi-crop deep learning and internal quality adaptive measurement module includes a barcode scanner connected to the edge computing industrial control computer. The industrial control computer acquires sample information and triggers IoT communication via the barcode scanner. The lightweight target detection algorithm uses an improved YOLOv11 network model automatically loaded from the cloud to process the image data stream acquired by the high frame rate camera. To adapt to the limited computing power of the edge computing industrial control computer and meet the real-time inference requirements of continuous material feeding, the model introduces a general inverted bottleneck UIB module to reconstruct the backbone network for feature flow transformation and dimensionality reduction calculation. The module achieves low-latency feature extraction on edge computing devices through continuous computation of point-based extended convolution, depthwise convolution, and projection dimensionality reduction. Addressing the physical phenomenon of deep feature loss during free fall of densely packed tiny seeds with diameters of 2-5 mm, the model extracts high-resolution P2 layer features with four times downsampling, carrying fine-grained spatial information from the shallow layers of the backbone network. It then employs a top-down strategy to fuse these features with deep semantic features, constructing an adaptive P2-P5 four-scale detection architecture for seeds of various heterogeneous sizes and varieties. This results in high-precision target bounding boxes and effective counting results.

5. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 4, characterized in that: The internal quality adaptive measurement module, based on the extraction of two-dimensional morphological parameters of a single seed using the OpenCV algorithm, further extracts the average transmittance grayscale value within the connected domain of a single seed and dynamically compares it with a transmittance grayscale threshold preset in the cloud. The transmittance grayscale threshold is obtained through experimental calibration for different crops and can be dynamically compensated and corrected according to light source attenuation. It accurately identifies shriveled seeds lacking dense material inside and precisely filters them out at the algorithm data level, excluding them from the effective count and morphological statistical sequence of high-quality plump seeds.

6. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 1, characterized in that: In the inverted multispectral transmission imaging and sensing module, the high frame rate camera is specifically an inverted, obliquely mounted linear array CMOS camera to avoid interference from falling dust on the lens; the multi-band adjustable transmission light source is located directly above the falling channel, and its light-emitting array integrates visible light and near-infrared LEDs, which are controlled by the PWM signal output by the system to switch electronic bands and adjust light intensity to provide a visible and near-infrared composite light field that matches the light transmission characteristics of different crops.

7. The intelligent crop grain counting and internal quality multidimensional measurement system based on multi-level drop discretization and edge-cloud collaboration according to claim 1, characterized in that: The automatic weighing and recycling module includes a drawer-type collection bin located at the bottom of the device and a built-in high-precision weighing sensor. The drawer-type collection bin is used to non-destructively receive the crop grains after testing. The high-precision weighing sensor obtains the total mass of the current batch of samples in real time and transmits its data stream to the edge computing industrial control computer for multi-dimensional data archiving and apparent thousand-grain weight calculation.

8. A crop grain intelligent counting and internal quality multidimensional measurement system based on multi-level drop discreteness and edge-cloud collaboration, applied to the detection system described in any one of claims 1-7, characterized in that, The process includes the following steps: 1) The operator loads N samples to be tested into the six-channel rotating hopper and scans the barcodes. The system automatically downloads the corresponding crop's vibration frequency, camera exposure parameters, dedicated detection model, and corresponding multispectral combinations and saturation grayscale thresholds from the cloud. 2) The microcontroller controls the turntable to rotate above the feeding port, and the electromagnetic mechanism releases the chassis. The flow regulation servo motor is linked with the vibratory feeder, performing adaptive flow regulation based on cloud parameters to ensure optimal flow rate and density as the seeds fall into the next stage. 3) Seeds are evenly fed to the inlet of the stepped guide chute via a vibrating feeder. As they flow through the multi-stage inclined slides, they are accelerated by gravity and tumbled by collisions with the step boundaries, rapidly creating physical separation in space. 4) Seeds fall freely through the imaging area. The edge computing terminal controls the inverted linear array CMOS camera to perform continuous line scanning in conjunction with the top multi-band light source, acquiring transmitted high-brightness images. The system calls the deep reconstruction multi-variety adaptive measurement module and transmittance extraction algorithm to output in real time the total number of falling seeds, the number of effective seeds, the proportion of shriveled seeds, and multi-dimensional phenotypic data such as length, width, and two-dimensional area. 5) After imaging is complete, the seeds fall into the weighing and recycling module. The sensor acquires the sample mass, automatically combines the visual effective particle count to calculate the thousand-particle weight, and uploads all data to the cloud platform; the hopper chassis resets, the system automatically rotates to the next hopper position and reloads parameters, and begins the next round of continuous measurement.

Images