Mountain land mobile intelligent vegetable sorting and fresh-keeping system based on AI vision

Abstract

The invention relates to a mountain mobile intelligent vegetable sorting and fresh-keeping system based on AI vision, and belongs to the technical field of agricultural machinery and agricultural product postharvest treatment. The system aims at solving the problems that the sorting efficiency of vegetables in mountainous areas is low and the sorting standard adaptability is poor. According to the technical scheme, the system is characterized by comprising a pre-cooling unit for quickly cooling harvested vegetables, a sorting unit for performing automatic quality grading and flaw detection, a packaging unit for performing sealed packaging, a temporary storage unit for providing a temporary fresh-keeping storage environment, a moving platform for supporting and moving the units and an AI visual control system. According to the system, field immediate processing is achieved through mobile integrated design, images are collected through a visual detection channel integrating a high-definition camera and a spectral imager, the image processing process is executed through an embedded processor, a sorting model is updated regularly in combination with a continuous learning module, and therefore the sorting precision and the localized adaptive capacity are improved, and the sorting efficiency is improved. The device is mainly used for on-site sorting and fresh-keeping operation in mountain vegetable producing areas.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of agricultural machinery and postharvest processing of agricultural products, and particularly relates to a mountainous mobile vegetable intelligent sorting and fresh-keeping system based on AI vision. BACKGROUND

[0002] At present, the mode of centralized harvesting and then transporting to a fixed place for sorting is generally adopted in mountainous vegetable production areas. Due to the complex terrain and limited road conditions in mountainous areas, the vegetables often need a long transportation time from the harvesting site to the sorting center. During this period, the vegetables continue to respire due to the high environmental temperature, which easily leads to a decrease in freshness, loss of nutritional components, and even wilting or deterioration. The failure to perform pre-cooling treatment in time after harvesting is an important factor affecting the fresh-keeping quality of vegetables.

[0003] In the sorting link, most production areas still mainly rely on manual quality grading and defect recognition of vegetables. The manual sorting method has the problems of low efficiency, high labor intensity, and difficulty in unifying the sorting standards. The operators are prone to visual fatigue in long-term repetitive labor, which leads to an increased missed detection rate of surface defects of vegetables, or inconsistent grading results due to individual judgment differences. In addition, different regions and different varieties of vegetables often have differences in appearance characteristics and quality standards, and the traditional sorting method lacks the flexibility to adjust to such regional and variety differences.

[0004] To solve the above problems, some production areas attempt to introduce fixed automatic sorting equipment. However, such equipment is usually fixedly installed and is difficult to adapt to the characteristics of scattered plots and limited scale of single-point harvesting in mountainous vegetable production areas. The equipment has high deployment cost and is difficult to move, and cannot realize immediate processing at the harvesting site. Its sorting model is often trained based on general data sets, and the recognition accuracy may decrease when facing unique appearance characteristics formed by special varieties or planting environments in specific regions. Therefore, in the postharvest processing of mountainous vegetables, how to realize efficient, accurate and adaptive sorting operation to local characteristics is still a difficulty to be further solved. SUMMARY

[0005] It is an object of the application to solve at least the above problems and to provide at least the advantages described later.

[0006] The application also aims to provide an AI vision-based mountain mobile vegetable intelligent sorting and fresh-keeping system, which can realize instant processing of post-harvest vegetables at the field head in the mountain vegetable production area through mobile integrated design, rapidly reduce the temperature of the vegetables by using a pre-cooling unit to delay quality deterioration, and automatically grade the quality, detect defects and judge the maturity of the vegetables by means of an AI vision sorting system integrated with visible light and spectral imaging, while continuously optimizing the sorting model through a continuous learning mechanism, thereby effectively improving the efficiency and accuracy of the sorting operation and enhancing the adaptability of the system to different vegetable varieties in different regions, and ultimately ensuring the post-harvest quality and commodity consistency of the vegetables.

[0007] In order to achieve these objects and other advantages of the present application, an AI vision-based mountain mobile vegetable intelligent sorting and fresh-keeping system is provided, which comprises a pre-cooling unit for rapidly cooling post-harvest vegetables, a sorting unit for automatically grading the quality and detecting defects of the pre-cooled vegetables, a packaging unit for sealing and packaging the sorted vegetables, a temporary storage unit for providing a temporary fresh-keeping storage environment for the packaged vegetables, a mobile platform for supporting the movement of the above-mentioned units, and an AI vision control system. The sorting unit comprises a conveyor belt and an AI vision detection channel, the conveyor belt being used to convey the pre-cooled vegetables to the AI vision detection channel, and the AI vision detection channel being internally provided with a high-definition camera and a spectral imager. The AI vision control system comprises an embedded processor and a solid-state memory, the embedded processor being connected to the high-definition camera and the spectral imager through data lines, and the solid-state memory being stored with an initial vegetable sorting model trained based on a convolutional neural network. The embedded processor executes an image data processing procedure, which comprises an image acquisition step, a feature extraction step and a classification output step; the image acquisition step synchronously acquires a visible light image collected by the high-definition camera and a spectral image collected by the spectral imager; the classification output step calls the initial vegetable sorting model, takes the feature information extracted in the feature extraction step as input, performs comprehensive judgment and simultaneously executes automatic grading, defect detection and maturity judgment operations; The AI vision control system is also integrated with a continuous learning module, which monitors the confirmation operation and correction operation of the operator on the AI sorting result on the touch screen of the sorting unit; when the correction operation occurs, the continuous learning module adds the current image data, feature extraction data and corrected classification label to the local training set; the continuous learning module starts an incremental training procedure at a preset period, such as once a week, updates the parameters of the initial vegetable sorting model using the local training set, and generates an optimized sorting model adapted to the local vegetable varieties.

[0008] Preferably, the feature extraction step extracts color features, shape features and texture features from the visible light image, and extracts reflectance features and absorption peak features from the spectral image.

[0009] Preferably, in the classification output step, the embedded processor executes a comprehensive decision flow, which includes performing a defect detection operation and a maturity judgment operation in parallel, then performing an automatic grading operation with the defect detection result as a priority condition, and associating the grading result with the maturity judgment result to generate a comprehensive control instruction, and the packaging unit adapts the packaging operation according to the comprehensive control instruction.

[0010] Preferably, in the comprehensive decision flow, more specifically: The defect detection operation identifies vegetable individuals with pest and disease spots, mechanical damage cracks, moldy areas, and abnormal shapes by comparing texture features with preset defect templates, and calculates the proportion of defect area to total vegetable surface area; The maturity judgment operation analyzes the red component ratio in color features and the chlorophyll absorption peak intensity in spectral features to output maturity grades for tomatoes and peppers; The embedded processor then takes the result of the defect detection operation as a priority judgment condition; when pest and disease spots, mechanical damage cracks, moldy areas, or abnormal shapes are identified, and the proportion of defect area to total vegetable surface area reaches 5% to 15%, the embedded processor marks the vegetable individual as substandard and generates a rejection instruction; when pest and disease spots, mechanical damage cracks, moldy areas, or abnormal shapes are identified, and the proportion of defect area to total vegetable surface area is between 1% and 5%, the embedded processor records the defect mark; The embedded processor then performs an automatic grading operation to calculate a basic grade based on color feature values, shape contour area values, and texture uniformity values; color feature values are determined by calculating the proportion of pixels in the H component falling within the range of 60 to 120 in the HSV color space; shape contour area values are obtained by pixel statistics and graded within the range of 8000 to 25000 pixels square; texture uniformity values are divided within the range of 0.1 to 0.5 by the contrast feature of the gray level co-occurrence matrix; If the vegetable individual has been marked as substandard, the rejection instruction is maintained; if the vegetable individual has a defect mark but has not been marked as substandard, at least one grade is forcibly lowered based on the basic grade to generate a final grade; if the vegetable individual has neither a substandard mark nor a defect mark, the basic grade is the final grade; The embedded processor finally associates the final grade with the maturity grade output by the maturity judgment operation to generate a comprehensive control instruction containing grading information and maturity information, and the packaging unit selects the corresponding packaging material and determines the stacking method according to the comprehensive control instruction.

[0011] Preferably, in the comprehensive decision-making process, the embedded processor also performs multi-defect comprehensive judgment; the multi-defect comprehensive judgment includes: when multiple types of defects exist in the same vegetable individual, the embedded processor calculates the defect area proportion of each type of defect, and counts the number of types of defects that exist at the same time; If the number of types of defects that exist is two or more, and the defect area proportion of each type of defect is between 0.5% and 3%, the embedded processor will add up the defect area proportions of these defects; If the total defect area proportion after addition reaches 3% to 8%, the grading result of the vegetable individual will be forcibly adjusted downward by one grade based on the basic grading; If the total defect area proportion after addition exceeds 8%, the vegetable individual will be marked as a defective product.

[0012] Preferably, in the comprehensive decision-making process, the embedded processor also performs dynamic weight allocation; the dynamic weight allocation includes: the embedded processor dynamically adjusts the weight of the defect detection operation in the grading decision according to the maturity level output by the maturity judgment operation; When the maturity level is mature, the embedded processor adjusts the defect area proportion threshold of the mold area in the defect detection operation from 5% to 3%; When the maturity level is unripe, the embedded processor adjusts the defect area proportion threshold of mechanical damage cracks from 5% to 8%.

[0013] Preferably, in the incremental training process of the continuous learning module, the samples in the local training set are weighted according to defect complexity; The embedded processor counts the number of types of defects that exist in each training sample, and for training samples that have two or more types of defects, an additional weight coefficient of 0.1 to 0.3 is added to the original weight, and the embedded processor uses the weighted training samples to update the parameters of the initial vegetable sorting model.

[0014] Preferably, in the multi-defect comprehensive judgment, the embedded processor performs weighted addition calculation; The embedded processor assigns a weight coefficient to each type of defect, where the weight coefficient of disease and insect spot is 1.2 to 1.5, the weight coefficient of mechanical damage crack is 1.0 to 1.2, the weight coefficient of mold area is 1.5 to 1.8, and the weight coefficient of abnormal morphology is 0.8 to 1.0; The embedded processor uses the following formula when calculating the total defect area proportion: Total defect area proportion = ∑(defect area proportion of the ith type of defect × corresponding weight coefficient); where i represents the serial number of the type of defect; The embedded processor performs hierarchical adjustment according to the total defect area ratio: When the total defect area ratio reaches 3% to 6%, the grading result of the vegetable individual is forcibly lowered by one grade based on the basic grading; when the total defect area ratio reaches 6% to 10%, the grading result of the vegetable individual is forcibly lowered by two grades based on the basic grading; when the total defect area ratio exceeds 10%, the vegetable individual is marked as a defective product; In the incremental training process of the continuous learning module, the embedded processor also records the misjudgment rate of each defect type; When the misjudgment rate of a certain defect type is higher than 15% for three consecutive incremental training, the embedded processor automatically adjusts the weight coefficient of the defect type, and the adjustment range is 0.1 to 0.3 times of the original weight coefficient, and the embedded processor uses the adjusted weight coefficient for subsequent weighted accumulation calculation.

[0015] Preferably, the mobile platform comprises a traction connecting mechanism and a wheel suspension system, the traction connecting mechanism is matched and connected with the traction device of the tractor or light truck, and the moving speed of the mobile platform is 5 to 15 kilometers per hour.

[0016] Preferably, the pre-cooling unit is fixedly installed in the front area of the mobile platform, and a temperature sensor and a refrigeration unit are arranged in the pre-cooling unit, the temperature sensor monitors the temperature in the pre-cooling chamber and feeds back to the refrigeration unit, so that the temperature in the pre-cooling chamber is maintained at 2 to 8 degrees Celsius.

[0017] The present application at least includes the following beneficial effects: Firstly, the present application integrates the functions of pre-cooling, sorting, packaging and temporary storage on a mobile platform, which can be directly deployed in the field of mountainous vegetable production area, realizing the instant processing of harvested vegetables, effectively shortening the operation time from harvesting to pre-cooling and sorting, and reducing the quality decline of vegetables caused by long-distance transportation and waiting for processing.

[0018] Secondly, the present application uses visible light and spectral image fusion analysis, and combines with the preset decision logic, so that the system can simultaneously complete the automatic grading, defect recognition and maturity judgment of vegetables, reduces the dependence on artificial experience, improves the objectivity and efficiency of sorting operation, and ensures the consistency of sorting standard.

[0019] Thirdly, the system of the present application has a continuous learning ability, which can record the correction feedback of the operator and periodically update the model incrementally, so that the initial universal sorting model can gradually adapt to the characteristics of vegetable varieties and localized sorting standards in specific production areas, improving the accuracy and adaptability of the system in different application scenarios.

[0020] Additional advantages, objects, and features of the application will be apparent from the following description, taken in conjunction with the accompanying drawings. DETAILED DESCRIPTION

[0021] The application is further described in detail by the following, to enable those skilled in the art to carry out the application according to the description.

[0022] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0023] An AI vision-based mountain mobile vegetable intelligent sorting and fresh-keeping system, comprising a pre-cooling unit for rapidly cooling harvested vegetables, a sorting unit for automatically grading and defect detecting the pre-cooled vegetables, a packaging unit for sealing and packaging the sorted vegetables, a temporary storage unit for providing a temporary fresh-keeping storage environment for the packaged vegetables, a mobile platform for supporting the movement of the above-mentioned units, and an AI vision control system; The sorting unit comprises a conveyor belt and an AI vision detection channel, the conveyor belt is used to convey the pre-cooled vegetables to the AI vision detection channel, and the AI vision detection channel is internally provided with a high-definition camera and a spectral imager; The AI vision control system comprises an embedded processor and a solid-state memory, the embedded processor is connected with the high-definition camera and the spectral imager through a data line, and the solid-state memory stores an initial vegetable sorting model trained based on a convolutional neural network; The embedded processor executes an image data processing flow, which comprises an image acquisition step, a feature extraction step, and a classification output step; the image acquisition step synchronously acquires a visible light image collected by the high-definition camera and a spectral image collected by the spectral imager; the classification output step calls the initial vegetable sorting model, takes the feature information extracted in the feature extraction step as input, performs comprehensive judgment, and synchronously executes automatic grading, defect detection, and maturity judgment operations; The AI vision control system is also integrated with a continuous learning module, which monitors the operator's confirmation operation and correction operation on the AI sorting result on the touch screen of the sorting unit, when the correction operation occurs, the continuous learning module adds the current image data, feature extraction data, and corrected classification label to the local training set, and the continuous learning module starts an incremental training process once every preset period, such as a week, updates the parameters of the initial vegetable sorting model using the local training set, and generates an optimized sorting model adapted to the local vegetable varieties.

[0024] Among the existing post-harvest processing techniques for vegetables, fixed sorting equipment is a common solution. This type of equipment is usually installed in centralized processing plants, requiring vegetables harvested from scattered mountain plots to be transported before processing. This mode faces challenges in mountainous areas, as the long transportation distance from scattered mountain plots to fixed plants can easily lead to a decline in vegetable quality during transportation. In addition, the sorting model built into these fixed equipment is usually trained based on general data sets, and when it comes to specific mountain vegetable varieties due to differences in climate and soil, its sorting accuracy may not be ideal, and once the model is deployed, it is difficult to effectively adjust it according to the actual situation.

[0025] The system of the present technical solution is built on a mobile platform, which can be a flat trailer with heavy suspension wheels, connected to a tractor or light truck through standard towing connection mechanisms, allowing movement on mountain field roads at a speed of 5 to 15 kilometers per hour. The pre-cooling unit is fixedly installed in the front area of the mobile platform, and can be equipped with a compressor refrigeration unit and a PT100 temperature sensor inside, maintaining the temperature in the pre-cooling room between 2 and 8 degrees Celsius, and rapidly cooling the harvested vegetables. The sorting unit is located behind the pre-cooling unit, and its conveyor belt can be made of food-grade rubber material to transport pre-cooled vegetables to the AI vision detection channel. Inside the channel are installed devices such as a Hikvision 2 million pixel high-definition global shutter camera and a Specim FX10 series hyperspectral imager, which are used to simultaneously capture visible light images and spectral images of the vegetables.

[0026] The core of the AI vision control system can be selected from NVIDIA Jetson AGX Orin embedded processors and Samsung 860EVO solid state drives. The embedded processor is connected to the camera and the spectral imager through data lines, and the solid state drive stores an initial vegetable sorting model trained based on the TensorFlow framework and convolutional neural network architecture. The image data processing flow executed by the embedded processor starts with the image acquisition step, which synchronously acquires image data from the two imaging devices. Then it enters the feature extraction step, which extracts color, shape and texture features from the visible light image, such as calculating the proportion of pixels in the H component of the HSV color space falling within the range of 60 to 120 to assess color, counting the total number of pixels within the contour to assess size, and calculating the contrast value within the range of 0.1 to 0.5 through the gray level co-occurrence matrix to assess texture uniformity; at the same time, it extracts reflectance features and chlorophyll feature absorption peak intensity from the spectral image. Then comes the classification output step, which calls the initial sorting model and inputs the extracted multi-dimensional feature information for comprehensive judgment, and simultaneously outputs the automatic grading, defect detection and maturity judgment results of the vegetables.

[0027] The continuous learning module collects data by monitoring the operator's confirmation or correction operations on the AI sorting results on the sorting unit touch screen. Each time a correction operation occurs, the module adds the current image data that triggered the correction, the corresponding feature extraction data, and the operator's corrected correct classification label to the local training set. The continuous learning module is set to automatically start an incremental training process once a week, using the local training set data accumulated this week to fine-tune the initial vegetable sorting model parameters, gradually generating an optimized sorting model that better adapts to the appearance characteristics of local specific vegetable varieties. Through the combination of this mobile integrated design and adaptive learning mechanism, the system can directly complete the series of operations from pre-cooling to sorting in the field, and continuously improve the recognition accuracy of local characteristics.

[0028] Further, the feature extraction step extracts color features, shape features, and texture features from the visible light image, and extracts reflectance features and absorption peak features from the spectral image.

[0029] In the field of vision-based agricultural product sorting technology, traditional image analysis methods usually rely on single visible light imaging. This method mainly captures the external color and shape information of vegetables through a color camera, and has certain recognition effect for obvious surface defects such as large area damage or color abnormalities. However, visible light images cannot penetrate the surface of vegetables and cannot effectively obtain information related to internal quality or chemical composition. For example, for tomatoes with similar surface color but different internal maturity, or slight internal mold, relying solely on visible light features can easily lead to misjudgment or missed detection, limiting further improvement of sorting accuracy.

[0030] The technical solution deepens the feature extraction step. When the vegetables enter the AI vision detection channel through the conveyor belt, the system synchronously starts the high-definition camera and the spectral imager. The high-definition camera can select a 2 million pixel industrial color camera to collect the visible light image of the vegetables. Then, the embedded processor extracts three types of key features from these visible light images: in terms of color features, the processor converts the image from the RGB color space to the HSV color space and counts the proportion of pixels in the H component falling within the green to yellow interval of 60 to 120, to quantify the color features; in terms of shape features, the outline of the vegetables is outlined through an edge detection algorithm, and the total area of the pixels enclosed by the outline is calculated, which is usually in the range of 8000 to 25000 pixels as a basis for judging the size; in terms of texture features, the gray level co-occurrence matrix of the image is calculated, and the contrast index is extracted, which is within the interval of 0.1 to 0.5 to reflect the uniformity of the texture. At the same time, the spectral image collected by the spectral imager (which can use a 400-1000 nanometer wavelength range hyperspectral device) is used to extract deeper features. The system analyzes the reflectivity value of a specific waveband (such as near 675 nanometers) and calculates the intensity of the chlorophyll characteristic absorption peak (such as at 670 nanometers). This dual extraction mechanism of visible light and spectral features provides a more comprehensive and reliable data basis for subsequent intelligent sorting decisions.

[0031] Further, in the classification output step, the embedded processor executes a comprehensive decision-making process, which includes parallel execution of flaw detection operation and maturity judgment operation, and then executes automatic grading operation with flaw detection result as priority condition, and generates comprehensive control instruction by associating grading result and maturity judgment result, and the packaging unit adapts packaging operation according to the comprehensive control instruction.

[0032] The technical solution effectively improves the coordination and decision-making efficiency of the sorting system by establishing a comprehensive decision-making process combining parallel processing and priority judgment. This solution enables flaw detection and maturity judgment to be performed simultaneously, avoiding the response delay that may be caused by traditional serial processing, and sets the flaw result as a precondition for grading operation, ensuring that quality problems are given priority, and preventing serious defect products from entering subsequent links and causing resource waste. Finally, by associating grading and maturity information to generate a unified instruction, the packaging unit can accurately match the packaging strategy, realizing the coordination of instructions from detection to packaging and smooth connection of the system operation process.

[0033] Further, in the comprehensive decision-making process, more specifically: The flaw detection operation identifies vegetable individuals with pest and disease spots, mechanical damage cracks, mold areas, and abnormal shapes by comparing texture features with preset defect templates, and calculates the proportion of defect area to total vegetable surface area. The maturity judgment operation outputs a maturity grade by analyzing the red component ratio in the color feature and the chlorophyll absorption peak intensity in the spectral feature for tomatoes and pepper vegetables; The embedded processor then takes the result of the flaw detection operation as a priority judgment condition; when a disease and pest spot, mechanical damage crack, mold area, or abnormal shape is identified, and the proportion of the defect area to the total area of the vegetable surface is between 5% and 15%, the embedded processor marks the vegetable individual as substandard and generates a rejection instruction; when a disease and pest spot, mechanical damage crack, mold area, or abnormal shape is identified, and the proportion of the defect area to the total area of the vegetable surface is between 1% and 5%, the embedded processor records the flaw mark; The embedded processor then performs an automatic grading operation to calculate a basic grade based on the color feature value, shape contour area value, and texture uniformity value; the color feature value is determined by calculating the proportion of pixels in the H component falling within the range of 60 to 120 in the HSV color space; the shape contour area value is obtained by pixel statistics and graded within the range of 8000 to 25000 pixels square; the texture uniformity value is divided within the range of 0.1 to 0.5 by the contrast feature of the gray level co-occurrence matrix; If the vegetable individual has been marked as substandard, the rejection instruction is maintained; if the vegetable individual has a flaw mark but has not been marked as substandard, the final grade is generated by forcibly downgrading at least one grade based on the basic grade; if the vegetable individual has neither a substandard mark nor a flaw mark, the basic grade is the final grade; The embedded processor finally associates the final grade with the maturity grade output by the maturity judgment operation to generate a comprehensive control instruction containing grading information and maturity information, and the packaging unit selects the corresponding packaging material and determines the stacking method according to the comprehensive control instruction.

[0034] In the decision mechanism of existing automated sorting systems, simple threshold judgment or single feature analysis is usually used for quality grading. For example, the system may only divide the grade according to color saturation, or only reject according to the surface flaw area. Such single-dimensional judgment often cannot accurately reflect the overall quality of the vegetables. Especially when there are multiple feature changes in the vegetables, such as a tomato with bright color but with slight mechanical damage, the traditional system has difficulty in coordinating the weight relationship between different features, and is prone to problems of inconsistent grading results with actual quality.

[0035] The technical solution refines the processing logic of the comprehensive decision-making process: after performing defect detection and maturity judgment in parallel, the embedded processor starts a priority judgment mechanism. For detected pest spots, mechanical damage cracks, moldy areas, or abnormal shapes, the system accurately calculates the area proportion. When the proportion reaches the threshold range of 5% to 15%, the processor immediately marks the vegetable individual as substandard and generates a rejection instruction; when the defect area is within the range of 1% to 5%, the defect mark is recorded but not immediately eliminated.

[0036] In the subsequent automatic grading stage, the processor calculates the basic grade based on multi-dimensional features. The color feature is quantified by counting the proportion of pixels with H component values in the HSV color space within the interval of 60 to 120; the shape feature is calculated by contour analysis to calculate the pixel area, and is divided into levels within the range of 8000 to 25000 pixels; the texture uniformity is analyzed by the contrast feature value of the gray level co-occurrence matrix, which represents the texture quality within the range of 0.1 to 0.5.

[0037] In the final grading determination link, the system implements a dynamic adjustment strategy: for individuals marked as substandard, maintain the rejection instruction; for individuals with defect marks, forcibly downgrade at least one grade based on the basic grade; and for individuals without defect marks, directly use the basic grade. Finally, the processor associates the grading results with the maturity information to generate comprehensive control instructions containing complete quality parameters. The packaging unit automatically selects the corresponding packaging material and determines the best stacking method according to the instructions, such as selecting packaging materials with better cushioning performance for vegetables with higher maturity, to achieve fine work of quality-based packaging.

[0038] Further, in the comprehensive decision-making process, the embedded processor also performs multi-defect comprehensive judgment; the multi-defect comprehensive judgment includes: when the same vegetable individual has multiple types of defects, the embedded processor calculates the defect area proportion of each type of defect and counts the number of types of defects that exist simultaneously; If the number of types of defects is two or more, and the defect area proportion of each type of defect is between 0.5% and 3%, the embedded processor adds the defect area proportions of these defects; If the total defect area proportion after addition reaches 3% to 8%, the grading result of the vegetable individual is forcibly downgraded by one grade based on the basic grade; If the total defect area proportion after addition exceeds 8%, the vegetable individual is marked as substandard.

[0039] In the existing defect judgment logic of the vegetable sorting system, an independent judgment threshold is usually set for a single defect type. When there is only one kind of defect on the surface of a vegetable, this method is still applicable. However, when multiple minor defects occur simultaneously on the same vegetable individual, such as a 2% area ratio of pest and disease spots and a 1.5% area ratio of mechanical damage, the traditional system will judge it as a qualified product because each defect does not reach the separately set substandard threshold. This processing method ignores the cumulative negative impact of multiple defects on the overall quality of the vegetable, which may cause some substandard products to flow into the market.

[0040] The technical solution adds a multi-defect comprehensive judgment mechanism in the established comprehensive decision-making process. When the embedded processor identifies that two or more types of defects exist simultaneously on the same vegetable individual, such as the coexistence of pest and disease spots and mechanical damage cracks, or the simultaneous occurrence of mold area and abnormal shape, the system will start a specific processing program. The processor first accurately calculates the defect area ratio of each type of defect. When these ratio values are all within the lower range of 0.5% to 3%, the system will not process them individually, but will accumulate the defect area ratios of each type of defect. If the total defect area ratio obtained by accumulation reaches the interval of 3% to 8%, even if any single defect does not reach the substandard standard, the system will forcibly downgrade the grading result of the vegetable individual by one grade based on the basic grading. If the total defect area ratio after accumulation exceeds 8%, the vegetable individual is directly marked as substandard and rejected. This comprehensive judgment mechanism fully considers the total amount of defects that have a substantial impact on the value of vegetable commodities, rather than just focusing on the severity of a single defect, making the sorting decision more in line with the actual quality status.

[0041] Further, in the comprehensive decision-making process, the embedded processor also performs dynamic weight allocation; the dynamic weight allocation includes: the embedded processor dynamically adjusts the weight of the defect detection operation in the grading decision according to the maturity level output by the maturity judgment operation; When the maturity level is mature, the embedded processor adjusts the defect area ratio judgment threshold of the mold area in the defect detection operation from 5% to 3%; When the maturity level is unripe, the embedded processor adjusts the defect area ratio threshold of mechanical damage cracks from 5% to 8%.

[0042] In existing vegetable sorting systems, the defect judgment standard usually adopts a fixed threshold. This one-size-fits-all approach is difficult to adapt to the quality characteristics of vegetables of different maturity. For example, for a mature tomato, surface mold may mean accelerated deterioration of internal quality, but the fixed threshold may not be able to identify this risk in time; on the contrary, for an immature green pepper, a slight mechanical damage on the surface may have little impact during subsequent maturation, but the fixed standard may cause excessive elimination. This inflexible judgment mechanism is prone to misjudgment of some maturity vegetables, affecting the overall sorting accuracy.

[0043] The technical solution further introduces a dynamic weight distribution mechanism on the basis of the established multi-defect comprehensive judgment. The embedded processor synchronously obtains the result output by the maturity judgment operation when performing defect detection, and dynamically adjusts the judgment threshold of a specific defect type according to different maturity levels. Specifically, when the system judges that the vegetable maturity level is mature, the processor will moderately tighten the defect area proportion judgment threshold of the mold area from the baseline 5% to 3%. This adjustment is based on the consideration that mature vegetables have lower tolerance to mold erosion, and can more sensitively identify potential safety risks. Conversely, when the vegetable is judged to be unripe, the processor will appropriately relax the defect area proportion judgment threshold of mechanical damage cracks from 5% to 8%. This is because unripe vegetable tissues are usually more flexible and have certain self-repair potential during subsequent storage, and slight mechanical damage has relatively small impact on the final quality. This dynamic threshold adjustment based on maturity enables the system to implement differentiated sorting strategies according to the physiological characteristics of different vegetables, effectively reducing unnecessary losses caused by maturity differences under the premise of ensuring basic quality, and improving the scientificity and accuracy of sorting decisions.

[0044] Further, in the incremental training process of the continuous learning module, the samples in the local training set are weighted according to the defect complexity; The embedded processor counts the number of defect types present in each training sample, and for training samples with two or more defect types, an additional weight coefficient of 0.1 to 0.3 is added to the original weight. The embedded processor uses the weighted training samples to update the parameters of the initial vegetable sorting model.

[0045] In the process of continuous optimization of machine learning models, traditional incremental training methods usually assign the same weight to all training samples. This approach, while gradually absorbing new data, ignores the complexity differences of samples themselves. In the context of vegetable sorting, samples containing only a single, obvious defect differ from those with multiple minor defects in their contribution to the improvement of model generalization ability. Equal treatment of all samples may lead to insufficient learning of complex and atypical defect combinations by the model, resulting in limited accuracy improvement in actual sorting when dealing with complex situations involving multiple overlapping defects.

[0046] In the incremental training process of the continuous learning module, the technical solution introduces a sample weighting mechanism based on defect complexity. Before performing the weekly incremental training, the embedded processor first analyzes each sample in the local training set to count the number of different defect types present on the corresponding vegetable individual. For training samples with two or more defect types, the system adds an additional weight coefficient of 0.1 to 0.3 based on the original weight. For example, a tomato image sample that records both pest and mechanical damage will have its importance increased, resulting in greater influence during model parameter updating. This weighting method enables the optimized sorting model to focus more on learning difficult cases with complex features and ambiguous boundaries during the iteration process, rather than simply repeating mastered feature patterns. By forcing the model to better understand the relationships and differences between multiple defect features, the system enhances the ability to handle complex situations in real-world sorting environments and promotes robust improvement in sorting accuracy.

[0047] Further, in the multi-defect comprehensive judgment, the embedded processor performs weighted accumulation calculation; The embedded processor assigns a weight coefficient to each defect type, with the weight coefficient of pest and disease spots being 1.2 to 1.5, the weight coefficient of mechanical damage cracks being 1.0 to 1.2, the weight coefficient of mold area being 1.5 to 1.8, and the weight coefficient of abnormal shape being 0.8 to 1.0; When calculating the total defect area ratio, the embedded processor uses the following formula: Total defect area ratio = ∑(defect area ratio of the ith defect type × corresponding weight coefficient); where i represents the serial number of the defect type; The embedded processor performs hierarchical adjustment based on the total defect area ratio: When the total defect area ratio reaches 3% to 6%, the hierarchical result of the vegetable individual is forcibly downgraded by one level based on the basic grading; when the total defect area ratio reaches 6% to 10%, the hierarchical result of the vegetable individual is forcibly downgraded by two levels based on the basic grading; when the total defect area ratio exceeds 10%, the vegetable individual is marked as a defective product; In the incremental training process of the continuous learning module, the embedded processor also records the misjudgment rate of each defect type; When the misjudgment rate of a certain defect type is higher than 15% for three consecutive incremental training, the embedded processor automatically adjusts the weight coefficient of this defect type, and the adjustment range is 0.1 to 0.3 times of the original weight coefficient. The embedded processor uses the adjusted weight coefficient for subsequent weighted accumulation calculation.

[0048] In the existing multi-defect processing of vegetable sorting systems, a simple arithmetic method is usually used to accumulate the areas of different defects. This method treats all defect types as equally important, ignoring the actual impact differences of different defects on vegetable quality. For example, the impact of mold area on the edible safety of vegetables is much greater than that of mechanical damage of the same area, and the deformed shape may more affect the appearance of the commodity. This indiscriminate accumulation method may lead to a result that does not match the actual situation, such as mixing moldy vegetables with potential safety hazards with deformed vegetables that only have poor appearance.

[0049] The technical solution introduces a weighted accumulation calculation mechanism in the multi-defect comprehensive judgment. The embedded processor assigns different weight coefficients to different defect types, among which the mold area with greater impact on quality is given a higher weight of 1.5 to 1.8, the disease and pest spot weight is 1.2 to 1.5, the mechanical damage and crack weight is 1.0 to 1.2, and the deformed shape with relatively small impact on quality is assigned a weight of 0.8 to 1.0. When calculating the total defect area ratio, the system multiplies the actual area ratio of each defect by its corresponding weight coefficient and then accumulates it. This weighted accumulation method makes mold and other serious defects occupy a larger proportion in the overall evaluation, more accurately reflecting the actual impact of multi-defect superposition on vegetable quality.

[0050] Based on the total defect area ratio obtained by weighted calculation, the system performs corresponding grading adjustment: when the ratio is in the interval of 3% to 6%, the grading result of the vegetable individual is forcibly lowered by one grade based on the basic grading; when the ratio reaches 6% to 10%, it is forcibly lowered by two grades; when the ratio exceeds 10%, the vegetable individual is directly marked as a defective product. This grading method fully considers the total amount of defects that have a substantial impact on the commodity value of vegetables, rather than only focusing on the severity of a single defect, making the sorting decision more in line with the actual quality status.

[0051] In the incremental training process of the continuous learning module, the system also establishes a dynamic optimization mechanism for the weight coefficients. The embedded processor records the misjudgment rate of each defect type in recent sorting operations. When the misjudgment rate of a certain defect type is higher than 15% for three consecutive incremental training, the system will automatically adjust the weight coefficient of that defect type. The adjustment range is 0.1 to 0.3 times the original weight coefficient. This feedback adjustment based on actual sorting results enables the system to continuously optimize the weight distribution of each defect type according to the local specific environment and operation standards, continuously improving the accuracy and adaptability of sorting decisions.

[0052] Further, the mobile platform comprises a traction connecting mechanism and a wheel suspension system, the traction connecting mechanism is matched and connected with the traction device of the tractor or light truck, and the moving speed of the mobile platform is 5-15 km / h.

[0053] The traction connecting mechanism can adopt a standard trailer ball connector or a pin shaft connector, and the material can be selected from high-strength alloy steel to withstand the traction force. The mechanism is usually installed on the front frame of the mobile platform and is fixedly connected with the platform body by welding or bolts. The wheel suspension system can be selected from an independent suspension configuration or a leaf spring suspension system, and the tire can be selected from an agricultural rubber tire with deep tread to enhance the grip. The suspension assembly is installed on the axle seat and shock absorber at the axle position of the platform bottom. During movement, the traction connecting mechanism is responsible for transmitting the traction force from the tractor to the platform, and the wheel suspension system alleviates the road bumps through springs and shock absorbing elements to keep the platform stable.

[0054] The traction connecting mechanism is designed to be matched and connected with the standard traction device of the tractor or light truck, for example, its connection interface can adapt to the common 50mm diameter traction ball or similar size hook, and the connection pin material can be selected from quenched steel to ensure wear resistance. The mechanism is assembled at the central position of the front end of the platform to ensure the centering with the towing vehicle. During actual connection, the operator puts the connector of the platform into the traction ball of the towing vehicle and locks the safety pin to complete the quick and reliable mechanical connection. The whole process does not require additional tools, which is convenient for rapid deployment or transfer in the field.

[0055] The moving speed of the mobile platform can be maintained at 5-15 km / h, for example, 5, 10 or 15 km / h. The speed range is adjusted by the control system of the towing vehicle, and the platform itself is not equipped with an independent power source. During the moving operation, the towing vehicle pulls the platform at an appropriate speed on the mountain road. The lower speed is close to 5 km / h when descending or on rough roads, and can be increased to 15 km / h on flat roads to balance the moving efficiency and equipment stability. Through the above design, the mobile platform can realize flexible transfer in mountainous environment, facilitate the overall deployment of the system to different harvesting sites, and reduce the dependence on fixed facilities.

[0056] Further, the pre-cooling unit is fixedly installed in the front area of the mobile platform, and a temperature sensor and a refrigeration unit are arranged inside the pre-cooling unit. The temperature sensor monitors the temperature in the pre-cooling chamber and feeds back to the refrigeration unit, so that the temperature in the pre-cooling chamber is maintained at 2-8 degrees Celsius.

[0057] The pre-cooling unit can be fixedly installed in the front area of the mobile platform, specifically in the front third of the platform skeleton, and is connected to the platform main frame by welding or high-strength bolts. The shell material of the pre-cooling unit can be selected from stainless steel plate or galvanized steel plate, and the internal insulation layer can be filled with polyurethane foam material. This arrangement allows vegetables to be processed first in the pre-cooling unit after harvesting, utilizing natural ventilation during vehicle travel to assist in heat dissipation, while avoiding the occupation of the operation space of the subsequent sorting unit.

[0058] The pre-cooling unit can be provided with a temperature sensor and a refrigeration unit, wherein the temperature sensor can be selected from a PT100 platinum resistance thermometer or a thermocouple temperature sensor, and is installed at multiple positions on the inner side wall and top of the pre-cooling chamber. The refrigeration unit can adopt a compression refrigeration system, including a compressor, a condenser, an evaporator and an expansion valve, and the evaporator coil is arranged on the top and side wall of the pre-cooling chamber. The temperature sensor is connected to the control unit through a data line to monitor the temperature distribution of each point in the pre-cooling chamber in real time.

[0059] The temperature sensor continuously monitors the temperature in the pre-cooling chamber and feeds back to the control system of the refrigeration unit, so that the temperature in the pre-cooling chamber is maintained at 2-8 degrees Celsius, for example, 2 degrees, 4 degrees, 6 degrees or 8 degrees. When the monitored temperature is higher than the set upper limit, the control system starts the refrigeration unit to run and absorbs the heat in the pre-cooling chamber through the evaporator; when the temperature drops to the set lower limit, the control system reduces the output of the refrigeration unit or suspends the operation. Through this feedback control mechanism, the pre-cooling unit can provide a stable low-temperature environment for the harvested vegetables, achieving rapid cooling and helping to delay the decline of the quality of the harvested vegetables.

[0060] Application example: picking and intelligent sorting of tomatoes In the tomato picking season, the mobile vegetable intelligent sorting and preservation system is deployed in the mountainous tomato planting area. After picking tomatoes in the morning, the picking personnel directly place the tomato basket at the entrance of the pre-cooling unit in the front of the mobile platform. The temperature sensor inside the pre-cooling unit monitors the temperature in real time and controls the operation of the refrigeration unit, so that the temperature in the chamber is stabilized at about 4 degrees Celsius. After about 15 minutes of pre-cooling treatment, the tomatoes are transferred to the conveyor belt of the sorting unit and enter the AI vision detection channel.

[0061] In the sorting unit, a high-definition camera captures the visible light image of the tomato, and a spectral imager synchronously acquires the spectral image thereof. An embedded processor executes an image data processing procedure to extract color, shape, and texture features from the visible light image, specifically including evaluating red maturity by counting the proportion of pixels with H component values in the 0-30 interval in the HSV color space, calculating shape size by counting the area of contours in the 10000-20000 pixel range, and analyzing texture uniformity by analyzing the contrast of the gray level co-occurrence matrix in the 0.2-0.4 interval. Meanwhile, the reflectance feature at the 675 nm waveband and the chlorophyll absorption peak intensity at 670 nm are extracted from the spectral image. The classification output step calls the initial tomato sorting model for comprehensive judgment, and performs parallel execution of defect detection and maturity judgment.

[0062] The system identifies that a tomato with a maturity level of "mature" has both an area of 1.2% of pest and disease spots and an area of 0.8% of mechanical damage cracks on the surface. According to the dynamic weight distribution mechanism, the system adjusts the judgment threshold of the moldy area from 5% to 3%. When performing multi-defect comprehensive judgment, the system uses weighted accumulation calculation: the pest and disease spot weight coefficient is 1.3, and the mechanical damage crack weight coefficient is 1.1, and the weighted total defect area proportion is calculated as (1.2% x 1.3) + (0.8% x 1.1) = 2.44%. According to the result, the system records the defect mark but does not reject. Subsequently, automatic grading is performed, and the color feature value of this tomato reaches the superior level standard, and the shape and texture feature values are good level, and the basic grading is calculated to be good level. Because there is a defect mark and the weighted total defect area proportion is below 3%, the system forcibly downgrades one level based on the basic grading, and generates the final grading as medium level.

[0063] Finally, the embedded processor associates the medium level grading with the maturity information and generates a comprehensive control instruction, which is transmitted to the packaging unit. The packaging unit selects packaging materials with moderate air permeability and uses single-layer stacking according to the instruction. The sorted tomatoes enter the temporary storage unit and are temporarily stored for preservation in an environment of 6 degrees Celsius.

[0064] During the continuous one-week sorting operation, the operator corrects part of the AI sorting results through the touch screen, and the continuous learning module adds these correction data to the local training set. The system automatically starts the incremental training process at the weekend, and additionally increases the weight coefficient of the training samples with two types of defects by 0.2, and updates the parameters of the initial model using the weighted local training samples. At the same time, the system records that the misjudgment rate of pest and disease spots is higher than 15% for three consecutive weeks, and automatically adjusts the weight coefficient from 1.3 to 1.5, generating an optimized sorting model that is more suitable for the characteristics of local tomato varieties.

[0065] Through the implementation of the present application example, the system realizes in-situ processing of mountainous tomato production areas through a mobile platform, the pre-cooling unit provides timely postharvest preservation, the sorting unit realizes accurate grading through multi-feature fusion and comprehensive decision-making, and the continuous learning mechanism ensures adaptive optimization of the system, thereby providing a complete postharvest processing solution for mountainous tomato production areas.

[0066] Although embodiments of the present application have been disclosed as above, they are not limited only to the applications listed in the specification and embodiments, and can be fully applied to various fields suitable for the present application, and additional modifications can be easily made by those skilled in the art, and thus the present application is not limited to specific details, without departing from the general concept defined by the claims and the equivalent scope.

Claims

1. A mobile intelligent vegetable sorting and preservation system based on AI vision, characterized in that: It includes a pre-cooling unit for rapidly cooling harvested vegetables, a sorting unit for automatically grading and detecting defects in pre-cooled vegetables, a packaging unit for sealing and packaging sorted vegetables, a temporary storage unit for providing a temporary preservation and storage environment for packaged vegetables, a mobile platform for supporting the movement of the above units, and an AI vision control system. The sorting unit includes a conveyor belt and an AI vision inspection channel. The conveyor belt is used to transport the pre-cooled vegetables to the AI ​​vision inspection channel, which is equipped with a camera and a spectral imager. The AI ​​vision control system includes an embedded processor and a solid-state memory. The embedded processor is connected to a camera and a spectral imager, and the solid-state memory stores an initial vegetable sorting model trained based on a convolutional neural network. The embedded processor executes the image data processing flow, which includes image acquisition, feature extraction, and classification output steps. The image acquisition step simultaneously acquires visible light images captured by the camera and spectral images captured by the spectral imager; The classification output step calls the initial vegetable sorting model, takes the feature information extracted in the feature extraction step as input, makes a comprehensive judgment, and simultaneously performs automatic grading, defect detection and maturity judgment operations. The AI ​​vision control system also integrates a continuous learning module. The continuous learning module monitors the operator's confirmation and correction operations on the AI ​​sorting results on the sorting unit touch screen. When a correction operation occurs, the continuous learning module adds the current image data, feature extraction data, and corrected classification labels to the local training set. The continuous learning module starts an incremental training process at a preset cycle, uses the local training set to update the parameters of the initial vegetable sorting model, and generates an optimized sorting model adapted to local vegetable varieties.

2. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 1, characterized in that, The feature extraction step extracts color features, shape features, and texture features from visible light images, and reflectance features and absorption peak features from spectral images.

3. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 1, characterized in that, In the classification output step, the embedded processor executes a comprehensive decision-making process, which includes parallel execution of defect detection and maturity judgment operations. Then, the defect detection results are used as priority conditions to execute an automatic grading operation, and the grading results are associated with the maturity judgment results to generate comprehensive control instructions. The packaging unit adapts the packaging operation according to the comprehensive control instructions.

4. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 3, characterized in that, In the comprehensive decision-making process, specifically: The defect detection operation compares texture features with preset defect templates to identify individual vegetables with disease and pest spots, mechanical damage cracks, moldy areas, and deformed shapes, and calculates the proportion of defect area to the total surface area of ​​the vegetable. The maturity assessment operation targets tomatoes and peppers, and outputs the maturity level by analyzing the ratio of red components in color features and the intensity of chlorophyll absorption peaks in spectral features. The embedded processor then uses the result of the defect detection operation as a priority criterion. When pest or disease spots, mechanical damage cracks, moldy areas, or deformities are detected, and the defective area accounts for 5% to 15% of the total surface area of ​​the vegetable, the embedded processor marks the vegetable as defective and generates a rejection instruction; when pest or disease spots, mechanical damage cracks, moldy areas, or deformities are detected, and the defective area accounts for 1% to 5% of the total surface area of ​​the vegetable, the embedded processor records the defect mark. The embedded processor then performs an automatic grading operation, calculating a basic grading based on color feature values, shape contour area values, and texture uniformity values. Color feature values ​​are determined by calculating the percentage of pixels whose H component falls within the range of 60 to 120 in the HSV color space; shape contour area values ​​are obtained through pixel statistics and graded within the range of 8000 to 25000 pixels squared; and texture uniformity values ​​are divided within the range of 0.1 to 0.5 based on the contrast characteristics of the gray-level co-occurrence matrix. If a vegetable has been marked as defective, the rejection instruction is maintained; if a vegetable has a defective mark but is not marked as defective, the grade is forcibly lowered by at least one level from the base grade to generate the final grade; if a vegetable has neither a defective mark nor a defective mark, the base grade is the final grade. Finally, the embedded processor associates and encodes the maturity level output from the final grading and maturity judgment operation to generate a comprehensive control instruction containing grading and maturity information. The packaging unit selects the corresponding packaging material and determines the stacking method according to the comprehensive control instruction.

5. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 4, characterized in that, In the integrated decision-making process, the embedded processor also performs a comprehensive judgment of multiple defects; The comprehensive judgment of multiple defects includes: when the same vegetable has multiple types of defects at the same time, the embedded processor calculates the defect area ratio of each type of defect and counts the number of defects that exist at the same time; If there are two or more types of defects, and the defect area ratio of each type of defect is between 0.5% and 3%, the embedded processor will sum up the defect area ratios of these defects. If the total defect area after accumulation reaches 3% to 8%, the grading result of the vegetable will be forcibly downgraded by one level based on the basic grading. If the total defect area exceeds 8%, the vegetable will be marked as substandard.

6. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 5, characterized in that, In the integrated decision-making process, the embedded processor also performs dynamic weight allocation; dynamic weight allocation includes: the embedded processor determines the maturity level of the operation output based on the maturity level, and dynamically adjusts the weight of the defect detection operation in the hierarchical decision-making; When the maturity level is mature, the embedded processor adjusts the threshold for determining the percentage of moldy areas in the defect detection operation from 5% to 3%. When the maturity level is immature, the embedded processor adjusts the threshold for determining the percentage of defect area of ​​mechanical damage cracks from 5% to 8%.

7. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous terrain according to claim 6, characterized in that, In the incremental training process of the continuous learning module, the samples in the local training set are weighted according to their flawed complexity. The embedded processor counts the number of defect types in each training sample. For training samples with two or more defect types, an additional weight coefficient of 0.1 to 0.3 is added to the original weight. The embedded processor then uses the weighted training samples to update the parameters of the initial vegetable sorting model.

8. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous terrain according to claim 7, characterized in that, In the comprehensive judgment of multiple defects, the embedded processor performs weighted cumulative calculation; The embedded processor assigns a weight coefficient to each defect type, with pest and disease spots having a weight coefficient of 1.2 to 1.5, mechanical damage cracks having a weight coefficient of 1.0 to 1.2, moldy areas having a weight coefficient of 1.5 to 1.8, and deformities having a weight coefficient of 0.8 to 1.

0. When calculating the percentage of total defect area, embedded processors use the following formula: Total defect area percentage = ∑(defect area percentage of the i-th type of defect × corresponding weighting coefficient); where i represents the sequence number of the defect type; Embedded processors perform tiered adjustments based on the percentage of total defect area: When the total defective area accounts for 3% to 6%, the grading result of the vegetable will be forcibly downgraded by one grade from the basic grading; when the total defective area accounts for 6% to 10%, the grading result of the vegetable will be forcibly downgraded by two grades from the basic grading; when the total defective area accounts for more than 10%, the vegetable will be marked as substandard. In the incremental training process of the continuous learning module, the embedded processor also records the false positive rate for each type of defect; When the misclassification rate of a certain defect type is higher than 15% for three consecutive incremental training iterations, the embedded processor automatically adjusts the weight coefficient of that defect type by 0.1 to 0.3 times the original weight coefficient. The embedded processor then uses the adjusted weight coefficient for subsequent weighted cumulative calculations.

9. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 1, characterized in that, The mobile platform includes a traction connection mechanism and a wheel suspension system. The traction connection mechanism is matched and connected to the traction device of a tractor or light truck. The mobile platform has a moving speed of 5 km / h to 15 km / h.

10. The AI ​​vision-based mobile intelligent vegetable sorting and preservation system for mountainous areas according to claim 1, characterized in that, The precooling unit is fixedly installed in the front area of ​​the mobile platform. The precooling unit is equipped with a temperature sensor and a refrigeration unit. The temperature sensor monitors the temperature in the precooling room and feeds it back to the refrigeration unit, so that the temperature in the precooling room is maintained between 2 degrees Celsius and 8 degrees Celsius.