An AI-based apple de-coring adaptive positioning system and method

Abstract

The application relates to the technical field of product processing, and discloses an apple kernel removing self-adaptive positioning system and method based on AI, which synchronously acquires an apple appearance image sequence and an internal transmission image through a multi-modal nondestructive acquisition unit, establishes a single-fruit exclusive three-dimensional reference coordinate system to complete spatial registration of multi-modal data, further mines deep correlation rules of fruit appearance geometric features and internal kernel distribution through a multi-modal fusion deep learning network, and realizes high-precision nondestructive perception of kernel three-dimensional space parameters; minimum safe cutting allowance is matched based on kernel accurate positioning results, self-adaptive selection of kernel removing cutter specifications, planning of cutting feed tracks and operation parameters are realized, single-fruit exclusive customized kernel removing control is realized, the cutting range of fruit pulp is maximally reduced on the premise of ensuring complete kernel removal and meeting the residual rate standard, and the technical contradiction between kernel residue and fruit pulp loss is solved from the cutting control level.

Description

Technical Field

[0001] This invention relates to the technical field of product processing, specifically to an AI-based adaptive positioning system and method for apple core removal. Background Technology

[0002] Apple pitting is a crucial step in the apple deep processing industry chain, and its accuracy directly affects the yield of finished fruit pulp and product quality. As the apple processing industry upgrades towards large-scale and refined production, traditional fixed-parameter pitting equipment can no longer meet the demands for high-quality and high-utilization production. Intelligent sensing-based adaptive pitting technology has become a core direction for industry technological upgrading, and is of great significance for improving the economic efficiency and product standardization of apple processing. Existing apple pitting technologies have several shortcomings, making it difficult to adapt to the demands of industrial production for high precision, low loss, and high flexibility. Traditional pitting equipment uses standardized cutters and fixed cutting strokes, failing to consider the natural variations in apple shape, pit position, and size. To cover the largest pit size, excessive cutting allowance is required, resulting in a high pulp loss rate. Reducing the cutting range easily leads to pit residue, causing… The existing seed detection methods rely heavily on visual perception or manual experience to determine the fruit's external dimensions, failing to accurately identify the three-dimensional spatial distribution and size parameters of the internal seed without causing damage. This hinders precise cutting control. Furthermore, the detection and seeding processes are independent, lacking a closed-loop feedback correction mechanism. This prevents dynamic adjustment of positioning and cutting parameters based on actual seeding results, leading to a continuous decrease in seeding accuracy due to equipment wear and changes in raw material batches. Additionally, the system lacks flexibility in adapting to various apple varieties and shapes. Changing apple varieties or fruit diameter grades requires repeated manual adjustments to equipment parameters, resulting in long production cycles and low efficiency. This makes it unsuitable for large-scale processing of multiple varieties, potentially leading to raw material waste and product quality risks in deep processing such as concentrated juice and apple chips.

[0003] Chinese invention patent application CN121027095A, published on November 28, 2025, discloses a hawthorn pitting and positioning detection system and method. The system includes a visual inspection module, a controller, a calculation module, and a pitting device. It acquires and preprocesses hawthorn images through a dual-trigger mechanism; the calculation module stores a standard image database based on ≥10 manually inspected samples, and uses a feature matching algorithm to compare the images to determine the number and angle of deviations in the positioning module; the controller implements graded control based on preset thresholds, triggering pauses or complete stops in the preceding process and controlling the positioning module to correct its angle; the pitting device uses a pointed hollow tubular cutter, dynamically adjusting its position and depth based on the visually detected fruit diameter to accommodate differences in hawthorn size. The method, through image acquisition, comparative analysis, threshold judgment, and cutter adaptation steps, achieves real-time monitoring and adjustment of hawthorn positioning deviations and precise control of the pitting process, effectively improving processing quality and efficiency. However, the above technical solutions cannot accurately locate the spatial position of the pit, increasing the pit residue rate and fruit pulp loss rate. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the numerous shortcomings of existing apple pitting technologies, which struggle to meet the demands of industrial production for high precision, low loss, and high flexibility; traditional pitting equipment employs standardized blades and fixed cutting strokes, failing to consider the natural variations in apple shape, pit position, and size between individual fruits. To cover the largest pit size, excessive cutting allowance is required, resulting in high pulp loss rates. Reducing the cutting range easily leads to pit residue, creating a technical contradiction where pit residue rate and pulp loss rate are mutually exclusive. Existing pit detection solutions rely heavily on visual perception or manual experience, only providing external fruit dimensions and failing to accurately identify the three-dimensional spatial distribution and size parameters of the internal pit without damage, hindering precise cutting control. Furthermore, the detection and pitting execution processes are independent, lacking a closed-loop feedback correction mechanism, making it impossible to address the root causes of pit loss. The positioning and cutting parameters are dynamically adjusted based on the actual pitting effect. As the equipment wears down and raw material batches change, the pitting accuracy will continue to decrease. At the same time, the system lacks flexibility in adapting to multiple varieties and different shapes of apples. When changing apple varieties or fruit diameter grades, manual repeated adjustments to equipment parameters are required, resulting in long production changeover cycles and low efficiency. This makes it difficult to adapt to production scenarios involving large-scale processing of multiple varieties. In deep processing production such as concentrated juice and apple chips, it can easily lead to waste of raw material costs and product quality risks. The solution is to achieve the above-mentioned multimodal non-destructive perception and AI deep learning to achieve three-dimensional precise positioning of the fruit pit, adaptively match cutting parameters and running trajectory, and solve the technical contradiction between the increase and decrease of fruit pit residue and fruit pulp loss. By relying on a closed-loop feedback mechanism to compensate for execution errors and adapt to multiple varieties of raw materials, the system can significantly improve pitting accuracy, raw material utilization rate and production line flexibility.

[0006] (II) Technical Solution

[0007] This invention is achieved through the following technical solution: an AI-based adaptive positioning method for apple core removal, the method comprising:

[0008] Step S101: Obtain an external image sequence and an internal transmission image of a single apple through a multimodal acquisition unit, perform image preprocessing on the external image sequence and the internal transmission image to generate a standardized multimodal image dataset.

[0009] Step S102: Extract pixel-level features from the standardized multimodal image dataset to obtain the geometric parameters of the single fruit appearance and the projection feature parameters of the fruit core;

[0010] Step S103: Establish a three-dimensional reference coordinate system with the geometric center of the single fruit as the origin, register the spatial mapping relationship between the appearance geometric parameters and the projection feature parameters of the fruit core, and construct the initial feature dataset of the three-dimensional space of the fruit core.

[0011] Step S104: Input the initial feature dataset of the three-dimensional space of the fruit pit into the AI ​​positioning model, analyze the correlation between the appearance features and the distribution of the internal fruit pit, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate positioning result of the fruit pit.

[0012] Step S105: Based on the precise positioning results of the fruit pit and the geometric parameters of the single fruit appearance, adaptively plan the pitting cutting trajectory and tool running parameters, and generate a single fruit-specific pitting control command;

[0013] Step S106: Drive the depitting actuator to complete the apple depitting operation according to the depitting control command, collect images of the fruit after depitting and verify the state of fruit pit residue and pulp loss, generate depitting effect data and update the processing database;

[0014] Step S107: Calculate the positioning deviation value and cutting deviation value based on the core removal effect data, correct the AI ​​positioning model parameters and core removal control logic through a feedback loop mechanism, regenerate the positioning result and control command, and optimize the adaptive core removal process.

[0015] Preferably, step S101 includes:

[0016] The apple is rotated at a constant speed by a rotating platform, and a multi-view industrial camera is used to continuously capture images of the apple's appearance. Simultaneously, a near-infrared transmission imaging device is used to capture transmission images of the apple's interior, thus obtaining the original appearance image sequence and the original transmission image.

[0017] For each frame of the original appearance image sequence and the original transmission image, a filtering algorithm is used to remove image noise and background interference to generate a first preprocessed image sequence;

[0018] Each frame of the first preprocessed image sequence is subjected to contrast enhancement processing to adjust the pixel grayscale distribution range, expand the grayscale difference between the pulp and the seed region, and generate a second preprocessed image sequence.

[0019] The second preprocessed image sequence is filtered for sharpness, retaining images with a Tenengrad gradient sharpness value greater than or equal to a preset sharpness threshold, and removing blurry frames with sharpness values ​​lower than the threshold to generate a preliminary effective image set;

[0020] The initial effective image set is then standardized in size and format to unify image resolution and data format, generating a standardized multimodal image dataset.

[0021] Preferably, step S102 includes:

[0022] For the appearance images in the standardized multimodal image dataset, a convolutional neural network is used to extract the surface contour of the fruit, the position of the fruit stalk, the position of the calyx, and the curvature features of the fruit shape.

[0023] Calculate the maximum diameter of a single fruit, its axial height, and the offset angle of the fruit's central axis based on the fruit's surface contour.

[0024] For the transmission images in the standardized multimodal image dataset, an edge detection algorithm is used to extract the projection contour of the fruit pit, and the maximum projection diameter, axial projection length and projection eccentricity of the fruit pit are calculated.

[0025] By combining the fruit's central axis from the external image with the projected outline of the pit from the transmission image, the radial and axial offsets of the pit relative to the fruit are determined.

[0026] By integrating the geometric parameters of the appearance of a single fruit with the projection feature parameters of the kernel, a pixel-level feature set is generated, providing basic data for subsequent spatial registration.

[0027] Preferably, step S103 includes:

[0028] A three-dimensional reference coordinate system for a single fruit is established with the geometric center point of the fruit as the origin and the central axis of the fruit's axial direction as the Z-axis.

[0029] Based on the 3D reconstruction results of multi-view appearance images, the surface contour features of the fruit are mapped to a 3D reference coordinate system to determine the 3D point cloud distribution of the outer surface of the fruit.

[0030] The projected outline of the fruit pit in the transmission image is mapped to a three-dimensional reference coordinate system through perspective transformation, and the initial three-dimensional envelope surface of the fruit pit is fitted by combining multi-view transmission data.

[0031] Calculate the center coordinates, axial length, radial diameter and spatial orientation deviation of the initial three-dimensional envelope of the kernel, and extract the spatial distribution characteristics of the kernel.

[0032] The spatial distribution characteristics of the fruit pit are associated with the three-dimensional point cloud data of the outer surface of the fruit according to spatial coordinates to construct the initial three-dimensional spatial feature dataset of the fruit pit.

[0033] Preferably, step S104 includes:

[0034] The initial feature dataset of the three-dimensional space of the fruit kernel is input into the AI ​​localization model, which is a multimodal fusion deep learning network;

[0035] The deep correlation between the geometric features of the fruit's appearance, the surface curvature distribution, and the spatial distribution of the internal kernel is explored through the multimodal fusion deep learning network.

[0036] Based on the correlation law, the initial three-dimensional envelope surface of the kernel is corrected and finely fitted, and the complete three-dimensional spatial parameters of the kernel are output, including the three-dimensional coordinates of the kernel center, the total axial length, the maximum radial diameter, the spatial deflection angle and the compensation amount for irregular contours.

[0037] The complete three-dimensional spatial parameters of the fruit pit are bound to the three-dimensional reference coordinate system to generate the accurate positioning result of the fruit pit and output it to the subsequent trajectory planning stage.

[0038] Preferably, step S105 includes:

[0039] Retrieve the three-dimensional dimensions and spatial position parameters of the fruit pit from the precise positioning results of the fruit pit, and combine them with the preset minimum safe cutting allowance to determine the nominal diameter and total feed stroke of the pitting tool;

[0040] Based on the spatial deflection angle of the pit and the offset angle of the fruit's central axis, calculate the feed angle and centering offset of the pitting tool, and determine the tool's feed start and end coordinates.

[0041] Based on tool parameters and feed coordinates, a continuous cutting trajectory is planned, and the corresponding feed rate and spindle speed parameters are matched to generate a single-fruit-specific de-core control command.

[0042] The core removal control command includes tool diameter selection signal, servo axis feed coordinate, feed speed and spindle speed parameters, which are used to directly drive the core removal actuator.

[0043] Preferably, step S106 includes:

[0044] The de-pitting control command is transmitted to the servo control system of the de-pitting actuator, which drives the tool to complete the feed cutting and retraction actions according to the planned trajectory, thereby realizing the de-pitting of a single fruit;

[0045] After pitting, cross-sectional and internal images of the pitted fruit are captured by a re-inspection camera to identify the area of ​​residual pit and the cutting boundary;

[0046] Based on the image recognition results, the seed residue rate and pulp loss rate of a single fruit are calculated, and corresponding seed removal effect data are generated.

[0047] The kernel removal effect data is associated and bound with the original positioning parameters and cutting parameters of the corresponding apple, stored in the processing database, and a data update log is generated.

[0048] If the pit residue rate or pulp loss rate exceeds the preset threshold, an abnormal warning signal is triggered, the fruit is marked as unqualified and transferred to the re-inspection process; if neither the pit residue rate nor the pulp loss rate exceeds the preset threshold, the fruit is judged to be of qualified pitting quality, and the corresponding data is included in the normal processing sample set.

[0049] Preferably, step S107 includes:

[0050] The actual distribution parameters of the fruit pits corresponding to qualified fruits are retrieved from the processing database, and the difference between them and the predicted positioning results output by the AI ​​positioning model is calculated to obtain the positioning deviation value.

[0051] Based on the positioning deviation value and the preset deviation correction rules, the parameter adjustment amount of the AI ​​positioning model is calculated, the model weights are fine-tuned through online learning, and the fruit kernel positioning output result is updated.

[0052] By combining the positioning deviation value with the actual cutting effect, the cutting parameter compensation amount is calculated, and the tool feed depth and centering position parameters in the nucleation control command are corrected.

[0053] Based on the corrected positioning results and cutting parameters, the de-pitting control command is regenerated, driving the de-pitting actuator to complete the subsequent de-pitting operation of the fruit, forming a closed-loop optimized adaptive de-pitting process.

[0054] An AI-based adaptive positioning system for apple core removal, used to implement the AI-based adaptive positioning method for apple core removal, the system comprising:

[0055] The multimodal acquisition module is used to acquire raw images of the apple's exterior and interior through a multi-view camera and a near-infrared imaging device. The images are then processed sequentially with filtering and noise reduction, contrast enhancement, sharpness filtering, and standardization to output a standardized multimodal image dataset.

[0056] The feature extraction module is used to extract the geometric features of fruit appearance and the projection features of the core from the standardized multimodal image dataset, calculate the diameter of a single fruit, fruit shape parameters, core projection size and offset, and generate a pixel-level feature set.

[0057] The spatial registration module is used to establish a three-dimensional reference coordinate system for a single fruit, complete the spatial mapping and registration of appearance features and internal transmission features, and construct the initial three-dimensional spatial feature dataset of the fruit core.

[0058] The AI ​​localization module is used to input the initial feature dataset of the fruit pit's three-dimensional space into the multimodal fusion deep learning network, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate localization result of the fruit pit;

[0059] The trajectory planning module is used to adaptively plan the parameters of the de-pitting tool and the cutting trajectory based on the precise positioning results of the fruit pit and the appearance parameters of the single fruit, and generate a dedicated de-pitting control command for each single fruit.

[0060] The depitting verification module is used to drive the actuator to complete the depitting operation according to the depitting control command, collect images of the fruit after depitting to verify the residual and loss status, generate depitting effect data and update the processing database.

[0061] The feedback optimization module is used to calculate the positioning and cutting deviations based on the core removal effect data, and correct the AI ​​model parameters and core removal control logic through a feedback loop mechanism to optimize the adaptive core removal process.

[0062] (III) Beneficial Effects

[0063] This invention provides an AI-based adaptive positioning system and method for apple core removal. It offers the following advantages:

[0064] First, by simultaneously acquiring apple exterior image sequences and internal transmission images through a multimodal non-destructive acquisition unit, a dedicated three-dimensional reference coordinate system for each apple is established to complete the spatial registration of multimodal data. Then, through a multimodal fusion deep learning network, the deep correlation between the geometric features of the fruit's exterior and the distribution of the internal core is mined, achieving high-precision non-destructive perception of the core's three-dimensional spatial parameters. This can accurately restore the core position, size, and spatial orientation of different individual apples, effectively solving the problem that traditional technologies cannot obtain the three-dimensional distribution of the internal core non-destructively, and providing a reliable data foundation for accurate core removal.

[0065] Second, by matching the minimum safe cutting allowance based on the precise positioning result of the fruit pit, the specifications of the pitting tool are adaptively selected, and the cutting feed trajectory and operating parameters are planned to achieve customized pitting control for each fruit. Under the premise of ensuring complete removal of the fruit pit and meeting the residual rate, the cutting range of the fruit pulp is minimized. This solves the technical contradiction between the offsetting effect of fruit pit residue and fruit pulp loss from the cutting control level, and significantly improves the utilization rate of raw materials and the product processing yield.

[0066] Third, by collecting actual processing effect data through the pitting and re-inspection process, and relying on the feedback loop mechanism to reverse the weight of the AI ​​positioning model and the pitting control parameters, a closed-loop optimization system for the entire process is formed. This system can not only compensate for the execution system errors caused by tool wear and mechanism clearance, ensuring long-term stability of pitting accuracy, but also quickly adapt to different varieties and shapes of apple raw materials through online learning, greatly improving the system's flexibility in changing production and its ability to adapt to different scenarios, thus meeting the production needs of industrialized multi-variety large-scale processing. Attached Figure Description

[0067] Figure 1 A flowchart of an AI-based adaptive positioning method for apple core removal provided by the present invention;

[0068] Figure 2 A flowchart illustrating the extraction of single fruit appearance and core feature parameters in an AI-based adaptive localization method for apple core removal provided by this invention.

[0069] Figure 3 A flowchart illustrating the kernel removal effect verification and processing database update process in an AI-based adaptive positioning method for apple kernel removal provided by this invention.

[0070] Figure 4 This is a schematic diagram of a module of an AI-based Apple core-free adaptive positioning system provided by the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] An embodiment of an AI-based adaptive positioning system and method for apple core removal is as follows:

[0073] Example 1:

[0074] Please see Figure 1 An AI-based adaptive positioning method for apple core removal, the method includes:

[0075] Step S101: Obtain the appearance image sequence and internal transmission image of a single apple through the multimodal acquisition unit, perform image preprocessing on the appearance image sequence and internal transmission image to generate a standardized multimodal image dataset;

[0076] Step S102: Extract pixel-level features from the standardized multimodal image dataset to obtain the geometric parameters of the single fruit appearance and the projection feature parameters of the fruit core;

[0077] Step S103: Establish a three-dimensional reference coordinate system with the geometric center of the single fruit as the origin, register the spatial mapping relationship between the appearance geometric parameters and the projection feature parameters of the fruit core, and construct the initial feature dataset of the three-dimensional space of the fruit core.

[0078] Step S104: Input the initial feature dataset of the three-dimensional space of the fruit pit into the AI ​​positioning model, analyze the correlation between the appearance features and the distribution of the internal fruit pit, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate positioning result of the fruit pit.

[0079] Step S105: Based on the precise positioning results of the fruit pit and the geometric parameters of the single fruit appearance, adaptively plan the pitting cutting trajectory and tool running parameters to generate a single fruit-specific pitting control command;

[0080] Step S106: Drive the pitting actuator to complete the apple pitting operation according to the pitting control command, collect images of the fruit after pitting and verify the state of fruit pit residue and pulp loss, generate pitting effect data and update the processing database;

[0081] Step S107: Calculate the positioning deviation value and cutting deviation value based on the core removal effect data, correct the AI ​​positioning model parameters and core removal control logic through a feedback loop mechanism, regenerate the positioning results and control commands, and optimize the adaptive core removal process.

[0082] Step S101 includes:

[0083] The apple is rotated at a constant speed by a rotating platform, and a multi-view industrial camera is used to continuously capture images of the apple's appearance. Simultaneously, a near-infrared transmission imaging device is used to capture transmission images of the apple's interior, thus obtaining the original appearance image sequence and the original transmission image.

[0084] For each frame of the original appearance image sequence and the original transmission image, a filtering algorithm is used to remove image noise and background interference to generate the first preprocessed image sequence;

[0085] Each frame of the first preprocessed image sequence is subjected to contrast enhancement processing to adjust the pixel grayscale distribution range, expand the grayscale difference between the pulp and the seed region, and generate the second preprocessed image sequence.

[0086] The second preprocessed image sequence is filtered for sharpness, retaining images with a Tenengrad gradient sharpness value greater than or equal to a preset sharpness threshold, and removing blurry frames with sharpness values ​​below the threshold to generate a preliminary effective image set;

[0087] The initial effective image set is standardized in size and format to unify image resolution and data format, generating a standardized multimodal image dataset.

[0088] Specifically, in the apple pitting process, a single apple is fed into a rotating platform by a conveyor mechanism. The platform rotates the apple 360° at a constant speed of 30 r / min. Four high-definition industrial cameras are evenly arranged around the platform to continuously capture images of the apple's appearance from different angles. Each apple captures 12-16 frames of appearance images, fully covering the fruit's surface outline. At the same time, near-infrared light sources and transmission imaging sensors are arranged above and below the platform to capture near-infrared transmission images of the apple. The difference in near-infrared light absorption coefficients between the pulp and the core is used to present the projected outline of the core. This yields the original appearance image sequence and the original transmission image, providing complete raw data support for subsequent preprocessing. To address noise in the original images caused by ambient lighting, water droplets on the fruit surface, and equipment vibration, a Gaussian filtering algorithm is applied to each frame of the sequence. A 5×5 filter kernel is used to smooth the image pixel matrix, adjusting the grayscale value of noisy pixels to the weighted average of the grayscale values ​​of their surrounding valid pixels. This eliminates abrupt changes in pixel grayscale values, making the pixel distribution between the fruit and background areas, and between the pulp and the pit areas, more continuous. This generates the first preprocessed image sequence. The filtering algorithm removes noise, initially separating the fruit and pit areas from the background, laying the foundation for subsequent feature extraction.

[0089] Each frame in the first preprocessed image sequence undergoes contrast enhancement. First, the number of pixels corresponding to different gray levels in each frame is counted to determine the distribution range of gray values. Then, a histogram equalization algorithm is used to adjust the mapping relationship of pixel gray values, stretching the original pixel gray values ​​concentrated in a narrow gray range to a complete gray range of 0-255. This expands the gray value difference between the fruit pulp and the pit region; for example, the gray value difference in the pit edge region, which was originally only 8-15, is increased to 30-45, making the projected outline of the pit and the surface edge of the fruit clearer. This process generates the second preprocessed image sequence. By enhancing contrast, the visual differentiation of the target region is strengthened, ensuring accurate capture of the geometric features of the fruit and pit in subsequent processing.

[0090] The second preprocessed image sequence undergoes sharpness filtering. The Tenengrad gradient sharpness value for each frame is calculated. This sharpness value is obtained by extracting the horizontal and vertical gradient components of pixels through convolution operations using the Sobel operator. The mean of the global gradient magnitude is used to quantify the sharpness of image edge details, thus reflecting the image sharpness. The calculation formula is as follows: ,in , These represent the number of rows and columns of the image, respectively. coordinates The grayscale gradient value of the pixel after convolution with the horizontal Sobel operator; coordinates The grayscale gradient value of each pixel after convolution with the vertical Sobel operator is used. A preset sharpness threshold of 25-40 is set, and images with a Tenengrad gradient sharpness value greater than or equal to this threshold are selected to form a preliminary effective image set. Blurry frames with sharpness values ​​below the threshold due to fruit surface reflection or brief platform shaking are removed. Through sharpness threshold selection, it is ensured that each image in the preliminary effective image set has sufficient detail sharpness to meet the image quality requirements for subsequent feature extraction.

[0091] The initial valid image set was standardized by scaling all images in the sequence to a uniform size of 640×640 pixels and converting them to 8-bit grayscale format. This eliminated differences in size and format between images from different viewpoints and at different acquisition times, generating a standardized multimodal image dataset. By standardizing image size and format, the feature extraction logic of the subsequent feature extraction network remains consistent across different frames, avoiding additional errors introduced by differences in image specifications. This provides a standardized data foundation for the accurate extraction of key features such as fruit shape and core size.

[0092] Step S102 includes:

[0093] For appearance images in a standardized multimodal image dataset, a convolutional neural network is used to extract the surface contour, pedicel position, calyx position, and fruit curvature features of the fruit.

[0094] Calculate the maximum diameter of a single fruit, its axial height, and the offset angle of the fruit's central axis based on the fruit's surface contour.

[0095] For the transmission images in the standardized multimodal image dataset, the edge detection algorithm is used to extract the projection contour of the fruit pit, and the maximum projection diameter, axial projection length and projection eccentricity of the fruit pit are calculated.

[0096] By combining the fruit's central axis from the external image with the projected outline of the pit from the transmission image, the radial and axial offsets of the pit relative to the fruit are determined.

[0097] By integrating the geometric parameters of the appearance of a single fruit with the projection feature parameters of the kernel, a pixel-level feature set is generated, providing basic data for subsequent spatial registration.

[0098] Specifically, please refer to Figure 2For appearance images in a standardized multimodal image dataset, a convolutional neural network (CNN) is input. The network extracts features through the collaborative operation of multiple convolutional and pooling layers. The convolutional layers use 3×3 kernels to capture local pixel features, exploring the differences in pixel distribution between the fruit and background regions. The pooling layers, through max pooling, reduce feature dimensionality while preserving crucial geometric information for subsequent analysis. Ultimately, the network extracts surface contour curves reflecting the fruit's geometric shape, the positions of the pedicel and calyx (marking the axial endpoints), and curvature features characterizing local variations in fruit shape. Leveraging the automatic feature extraction capabilities of the neural network, image pixel information is transformed into quantifiable contour and positional parameters, providing standardized foundational data for subsequent fruit shape parameter calculations and avoiding errors caused by subjective human judgment.

[0099] The maximum fruit diameter and axial height of a single fruit are calculated based on the extracted surface contour curve. Pixels corresponding to the two edges of the fruit are determined on the cross-section of the contour. The actual fruit diameter is obtained by calculating the straight-line distance between the two points and combining it with pixel equivalent conversion. Simultaneously, the fruit stalk endpoint and calyx endpoint are located along the fruit axis, and the vertical distance between the two points is calculated to obtain the fruit axial height. The fruit's central axis is determined based on the line connecting the positions of the stalk and calyx, and the angle between the central axis and the vertical axis of the platform is calculated to obtain the fruit's central axis offset angle. For the extraction of the fruit pit projection contour from the transmission image, the Canny edge detection algorithm is used to identify the boundary contour of the fruit pit. The maximum lateral distance on the contour is taken as the maximum diameter of the fruit pit projection, and the maximum longitudinal distance is taken as the axial projection length of the fruit pit. The lateral distance between the center of the fruit pit contour and the center of the fruit is calculated to obtain the projection eccentricity.

[0100] By combining the fruit's central axis determined from the external image with the projected contour of the pit extracted from the transmission image, the radial and axial offsets of the pit relative to the fruit's central axis are determined through spatial coordinate conversion, thus preliminarily quantifying the pit's positional deviation within the fruit. External geometric parameters such as the maximum fruit diameter, axial height, and central axis offset angle are integrated with projection feature parameters such as the pit's projected diameter, axial length, and eccentricity to form a structured set of pixel-level features. Each parameter corresponds to a unique spatial identifier, providing a complete feature foundation for subsequent 3D spatial registration and modeling. This process transforms pixel information in 2D images into quantifiable geometric parameters, solving the problem that traditional detection methods can only obtain a single external dimension and cannot correlate the relative position of the internal pit.

[0101] Step S103 includes:

[0102] A three-dimensional reference coordinate system for a single fruit is established with the geometric center point of the fruit as the origin and the central axis of the fruit's axial direction as the Z-axis.

[0103] Based on the 3D reconstruction results of multi-view appearance images, the surface contour features of the fruit are mapped to a 3D reference coordinate system to determine the 3D point cloud distribution of the outer surface of the fruit.

[0104] The projected outline of the fruit pit in the transmission image is mapped to a three-dimensional reference coordinate system through perspective transformation, and the initial three-dimensional envelope surface of the fruit pit is fitted by combining multi-view transmission data.

[0105] Calculate the center coordinates, axial length, radial diameter and spatial orientation deviation of the initial three-dimensional envelope of the kernel, and extract the spatial distribution characteristics of the kernel.

[0106] The spatial distribution characteristics of the fruit pit are associated with the three-dimensional point cloud data of the outer surface of the fruit according to spatial coordinates to construct the initial three-dimensional spatial feature dataset of the fruit pit.

[0107] Specifically, based on the positions of the pedicel and calyx extracted in step S102, and the surface contour of the fruit, the geometric center point of a single fruit is calculated. Using this point as the origin, the direction of the line connecting the pedicel and calyx as the Z-axis, the direction of the fruit's maximum diameter as the X-axis, and the direction perpendicular to the XZ plane as the Y-axis, a unique three-dimensional reference coordinate system is established for each single fruit. This coordinate system is completely bound to the actual posture of a single apple, eliminating the influence of differences in fruit placement posture on subsequent positioning and providing a unified spatial reference for multimodal data fusion.

[0108] Based on a multi-view image sequence, a motion recovery structure algorithm was used to reconstruct the fruit's three-dimensional structure, obtaining high-density three-dimensional point cloud data of the fruit's outer surface. All point cloud data were then transformed into a three-dimensional reference coordinate system to obtain the complete three-dimensional point cloud distribution of the fruit's outer surface, accurately restoring the actual three-dimensional morphology of a single fruit. For near-infrared transmission images, perspective distortion was first corrected using camera calibration parameters, and then the two-dimensional projection outline of the fruit pit was mapped onto the corresponding section in the three-dimensional reference coordinate system. Combining transmission projection data from multiple perspectives, an ellipsoidal fitting algorithm was used to fit the initial three-dimensional envelope of the fruit pit, preliminarily restoring the three-dimensional morphology of the pit inside the fruit.

[0109] Based on the fitted initial 3D envelope of the fruit pit, the 3D coordinates of the pit center, the total axial length along the Z-axis, the maximum radial diameter in the XY plane, and the spatial deflection angle of the pit's central axis relative to the fruit's central axis are extracted to form a complete spatial distribution feature of the pit. This spatial distribution feature is then correlated one-to-one with the 3D point cloud data of the fruit's outer surface according to spatial coordinates and stored in a structured format to construct the initial 3D spatial feature dataset of the pit. This dataset contains both complete geometric information of the fruit's exterior and preliminary 3D distribution information of the internal pit, providing an input foundation for accurate prediction by the AI ​​model.

[0110] A dedicated 3D reference coordinate system for each fruit was established, solving the problem of lacking a unified spatial reference and being unable to directly fuse multimodal data. Spatial registration of appearance and internal features was achieved through multi-view 3D reconstruction and perspective mapping, solving the problem that a single imaging method cannot obtain the true 3D parameters of the fruit core. The constructed initial feature dataset provides a complete input dimension for AI fine-grained positioning, effectively improving the accuracy of subsequent positioning results.

[0111] Step S104 includes:

[0112] The initial feature dataset of the three-dimensional space of the fruit pit is input into the AI ​​localization model, which is a multimodal fusion deep learning network.

[0113] By using a multimodal fusion deep learning network, we can uncover the deep correlation between the geometric features of fruit appearance, surface curvature distribution, and the spatial distribution of the internal kernel.

[0114] Based on the correlation law, the initial three-dimensional envelope surface of the kernel is corrected and finely fitted, and the complete three-dimensional spatial parameters of the kernel are output, including the three-dimensional coordinates of the kernel center, the total axial length, the maximum radial diameter, the spatial deflection angle and the compensation amount for irregular contours.

[0115] The complete three-dimensional spatial parameters of the fruit pit are bound to the three-dimensional reference coordinate system to generate the accurate positioning result of the fruit pit and output it to the subsequent trajectory planning stage.

[0116] Specifically, the initial 3D spatial feature dataset of the fruit pit is input into a multimodal fusion deep learning network, which includes a feature encoding layer, a cross-modal fusion layer, and a parameter regression layer. The feature encoding layer performs high-dimensional encoding on the fruit appearance point cloud features and the initial envelope surface features of the fruit pit, respectively, and extracts deep features such as local curvature and morphological gradient. The cross-modal fusion layer achieves adaptive weighted fusion of appearance features and internal features through an attention mechanism, and explores the inherent correlation between appearance morphology and internal fruit pit distribution under different varieties and fruit shapes, such as the correspondence between the depth of the stem indentation, the height of the fruit shape and the axial length of the fruit pit, and the correlation between the curvature of the fruit equatorial surface and the radial diameter of the fruit pit.

[0117] Based on the learned association patterns, the network performs point-by-point corrections on the initially fitted 3D envelope of the fruit pit, compensating for fitting deviations caused by projection imaging errors and individual morphological differences. It also performs refined fitting corrections, particularly for the edge regions of eccentric and irregular fruit pits. Finally, the network outputs complete 3D spatial parameters of the fruit pit, including the 3D coordinates of the pit's center (positioning error ≤ 0.5 mm), total axial length (measurement error ≤ 0.3 mm), maximum radial diameter, spatial deflection angle of the pit's central axis relative to the fruit's central axis, and contour compensation for irregular protrusions at both ends of the pit.

[0118] All three-dimensional spatial parameters of the fruit core are bound to a dedicated three-dimensional reference coordinate system for each fruit, ensuring that the positioning results completely correspond to the actual spatial position of the fruit. This generates the final accurate positioning result of the fruit core and transmits it to the trajectory planning module. This AI positioning model is trained on tens of thousands of labeled samples from multiple mainstream apple varieties such as Fuji, Gala, and Golden Delicious. It can adapt to various fruit shapes within the diameter range of 60mm-95mm and also has high prediction accuracy for deformed fruits and eccentric cores, solving the problems of insufficient generalization ability and significant accuracy decay when faced with individual differences in traditional models.

[0119] By employing a multimodal fusion deep learning network to uncover the deep correlation between the appearance and the internal kernel, the problem of existing technologies being unable to accurately reconstruct the three-dimensional morphology of the kernel using only projection data was solved, achieving high-precision calculation of the kernel's three-dimensional parameters under non-destructive conditions. Refined envelope surface correction and irregular contour compensation further improved the positioning accuracy of irregularly shaped kernels, providing core support for subsequent low-loss and precise cutting.

[0120] Step S105 includes:

[0121] Retrieve the three-dimensional dimensions and spatial position parameters of the fruit pit from the precise positioning results of the fruit pit, and combine them with the preset minimum safe cutting allowance to determine the nominal diameter and total feed stroke of the pitting tool;

[0122] Based on the spatial deflection angle of the pit and the offset angle of the fruit's central axis, calculate the feed angle and centering offset of the pitting tool, and determine the tool's feed start and end coordinates.

[0123] Based on tool parameters and feed coordinates, a continuous cutting trajectory is planned, and the corresponding feed rate and spindle speed parameters are matched to generate a single-fruit-specific de-core control command.

[0124] The core removal control command includes tool diameter selection signal, servo axis feed coordinate, feed rate and spindle speed parameters, which are used to directly drive the core removal actuator.

[0125] Specifically, by retrieving the maximum radial diameter and total axial length from the precise pit positioning results, and adding a minimum safety cutting allowance of 0.3mm-0.5mm to the maximum pit size, the nominal diameter and total axial feed stroke of the required pitting tool are determined. Compared to traditional equipment that uses fixed large-diameter tools, this adaptive selection can match the smallest suitable tool according to the actual size of a single fruit pit, minimizing the amount of fruit pulp removed.

[0126] By combining the spatial deflection angle of the fruit pit's central axis with the offset angle of the fruit's own central axis, the feed axis angle of the pitting tool is calculated to ensure that the tool's feed direction coincides with the fruit pit's central axis. Simultaneously, based on the three-dimensional coordinates of the fruit pit's center, the tool's centering offset is calculated to determine the tool's starting and ending coordinates, ensuring that the tool completely penetrates the fruit pit along its central axis, leaving no pit tissue behind and avoiding excessive cutting of the surrounding fruit flesh. Based on the determined tool parameters and feed coordinates, a continuous motion trajectory of the tool from feeding to cutting and then to retraction is planned, matching a feed speed of 150mm / s-200mm / s and a spindle speed of 300r / min, generating structured pitting control commands.

[0127] The pitting control commands directly interface with the servo control system of the pitting actuator, including control parameters such as feed coordinates of the X / Y / Z axes, rotary axis angle parameters, feed speed, and spindle speed. This allows for direct drive of the actuator to complete customized pitting operations for individual fruits. This adaptive trajectory planning achieves precise control with a "one fruit, one strategy" approach, resolving the trade-off between pit residue and pulp loss at the cutting scheme level, maximizing pulp yield while ensuring zero residue.

[0128] Step S106 includes:

[0129] The de-pitting control command is transmitted to the servo control system of the de-pitting actuator, which drives the tool to complete the feed cutting and retraction actions according to the planned trajectory, thereby realizing the de-pitting of a single fruit;

[0130] After pitting, cross-sectional and internal images of the pitted fruit are captured by a re-inspection camera to identify the area of ​​residual pit and the cutting boundary;

[0131] Based on the image recognition results, the seed residue rate and pulp loss rate of a single fruit are calculated, and corresponding seed removal effect data are generated.

[0132] The kernel removal effect data is associated and bound with the original positioning parameters and cutting parameters of the corresponding apple, stored in the processing database, and a data update log is generated.

[0133] If the pit residue rate or pulp loss rate exceeds the preset threshold, an abnormal warning signal is triggered, the fruit is marked as unqualified and transferred to the re-inspection process; if neither the pit residue rate nor the pulp loss rate exceeds the preset threshold, the fruit is judged to be of qualified pitting quality, and the corresponding data is included in the normal processing sample set.

[0134] Specifically, please refer to Figure 3 The pitting control command is transmitted to the servo control system of the pitting actuator via the industrial bus. The control system drives the X / Y / Z three-axis servo motor and the rotary servo motor to drive the hollow tubular pitting cutter to complete the feed cutting according to the planned trajectory and parameters. After the core is completely ejected, the cutter returns to its original position, completing the pitting operation of a single apple.

[0135] After pitting, the fruit is conveyed to the re-inspection station by a conveyor mechanism. Cross-sectional images and images of the inner wall of the pitted fruit are acquired using dual-view industrial cameras. Image segmentation algorithms are used to identify residual pit tissue and the boundary of the cut inner hole in the images. The area and volume of the residual pit are calculated to determine the pit residue rate. Simultaneously, the volume of the cut fruit pulp is calculated using the diameter and depth of the cut inner hole. Combined with the total fruit volume, the pulp loss rate is calculated, generating pitting effect data that includes residual rate, loss rate, and deviation location information.

[0136] The pitting effect data is linked and bound one-to-one with the original appearance parameters, AI positioning results, and cutting control parameters of the corresponding apple, assigning a unique processing batch identifier and storing it in the processing database. Simultaneously, a database update log is automatically generated, recording the update time, data entries, and operation identifiers to ensure the traceability of all processing data. The preset thresholds for pit residue rate are 0.3% and pulp loss rate are 5%. If the test result of a single fruit exceeds either threshold, an audible and visual warning signal is triggered, marking the fruit as unqualified and automatically transferring it to the manual re-inspection process. If both indicators do not exceed the thresholds, the pitting quality of the single fruit is deemed qualified, and the corresponding data is included in the normal processing sample set for subsequent online iterative optimization of the model.

[0137] Online re-inspection and effect quantification enable real-time hierarchical control of denucleation quality, preventing unqualified products from flowing into subsequent processes; full parameter association storage and traceability management provide complete and effective data support for subsequent model optimization and process improvement, solving the problems of missing quality data and inability to trace and optimize in the traditional denucleation process.

[0138] Step S107 includes:

[0139] The actual distribution parameters of the fruit pits corresponding to qualified fruits are retrieved from the processing database, and the difference between them and the predicted positioning results output by the AI ​​positioning model is calculated to obtain the positioning deviation value.

[0140] Based on the positioning deviation value and the preset deviation correction rules, the parameter adjustment amount of the AI ​​positioning model is calculated, the model weights are fine-tuned through online learning, and the fruit kernel positioning output result is updated.

[0141] By combining the positioning deviation value with the actual cutting effect, the cutting parameter compensation amount is calculated, and the tool feed depth and centering position parameters in the nucleation control command are corrected.

[0142] Based on the corrected positioning results and cutting parameters, the de-pitting control command is regenerated, driving the de-pitting actuator to complete the subsequent de-pitting operation of the fruit, forming a closed-loop optimized adaptive de-pitting process.

[0143] Specifically, the re-inspection data corresponding to qualified fruits is retrieved from the processing database. The actual size and position parameters of the fruit pits are extracted, and the difference between these parameters and the predicted parameters output by the AI ​​positioning model is calculated dimension by dimension to obtain the three-dimensional positioning deviation value and the size deviation value. This deviation value comprehensively reflects the prediction error of the AI ​​model and the systematic error of the actuator. Using only valid data from qualified fruits for deviation calculation can avoid the abnormal data from unqualified products interfering with the model optimization direction and ensure the accuracy of the correction logic.

[0144] Based on the magnitude and direction of the positioning deviation, and combined with preset deviation correction rules, the weight adjustment amounts of the fusion layer and regression layer in the AI ​​positioning model are calculated. Using a small-batch online learning approach, actual pitted data from qualified fruits are used as fine-tuning samples to iteratively update the model weights, allowing the model to gradually adapt to the variety characteristics and morphological distribution of the current batch of apples, continuously improving positioning accuracy. For different apple varieties, the model can quickly adapt through this online feedback mechanism without manual retraining, significantly improving the system's flexibility in changing production varieties.

[0145] Simultaneously, by combining positioning deviation and actual cutting effect, the system deviation caused by the actuator is isolated, and the parameter compensation amount for tool feed depth and centering position is calculated. The control parameters of the trajectory planning module are then corrected to compensate for execution errors caused by tool wear and mechanism backlash. After correction, for subsequent apple delivery, the positioning result and core removal control command are regenerated based on the updated AI model and corrected control parameters, driving the actuator to complete the core removal operation, forming a complete closed loop of "perception-positioning-execution-verification-optimization".

[0146] The feedback loop mechanism enables online optimization of the AI ​​model and dynamic compensation of control parameters, solving the problems of fixed parameters in traditional equipment, inability to adapt to batch differences and equipment wear. The closed-loop optimization process enables the system to maintain stable kernel removal accuracy over a long period of time, while having good adaptive adaptability to different apple varieties, meeting the production needs of industrialized multi-variety large-scale processing.

[0147] Example 2:

[0148] Please see Figure 4 An AI-based adaptive positioning system for apple core removal, used to implement an AI-based adaptive positioning method for apple core removal, the system includes:

[0149] The multimodal acquisition module is used to acquire raw images of the apple's exterior and interior through a multi-view camera and a near-infrared imaging device. The images are then processed sequentially with filtering and noise reduction, contrast enhancement, sharpness filtering, and standardization to output a standardized multimodal image dataset.

[0150] The feature extraction module is used to extract the geometric features of fruit appearance and the projection features of the core from the standardized multimodal image dataset, calculate the diameter of a single fruit, fruit shape parameters, core projection size and offset, and generate a pixel-level feature set.

[0151] The spatial registration module is used to establish a three-dimensional reference coordinate system for a single fruit, complete the spatial mapping and registration of appearance features and internal transmission features, and construct the initial three-dimensional spatial feature dataset of the fruit core.

[0152] The AI ​​localization module is used to input the initial feature dataset of the fruit pit's three-dimensional space into the multimodal fusion deep learning network, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate localization result of the fruit pit;

[0153] The trajectory planning module is used to adaptively plan the parameters of the de-pitting tool and the cutting trajectory based on the precise positioning results of the fruit pit and the appearance parameters of the single fruit, and generate a dedicated de-pitting control command for each single fruit.

[0154] The depitting verification module is used to drive the actuator to complete the depitting operation according to the depitting control command, collect images of the fruit after depitting to verify the residual and loss status, generate depitting effect data and update the processing database.

[0155] The feedback optimization module is used to calculate the positioning and cutting deviations based on the core removal effect data, and correct the AI ​​model parameters and core removal control logic through a feedback loop mechanism to optimize the adaptive core removal process.

[0156] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based adaptive positioning method for apple core removal, characterized in that, The method includes: Step S101: Obtain an external image sequence and an internal transmission image of a single apple through a multimodal acquisition unit, perform image preprocessing on the external image sequence and the internal transmission image to generate a standardized multimodal image dataset. Step S102: Extract pixel-level features from the standardized multimodal image dataset to obtain the geometric parameters of the single fruit appearance and the projection feature parameters of the fruit core; Step S103: Establish a three-dimensional reference coordinate system with the geometric center of the single fruit as the origin, register the spatial mapping relationship between the appearance geometric parameters and the projection feature parameters of the fruit core, and construct the initial feature dataset of the three-dimensional space of the fruit core. Step S104: Input the initial feature dataset of the three-dimensional space of the fruit pit into the AI ​​positioning model, analyze the correlation between the appearance features and the distribution of the internal fruit pit, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate positioning result of the fruit pit. Step S105: Based on the precise positioning results of the fruit pit and the geometric parameters of the single fruit appearance, adaptively plan the pitting cutting trajectory and tool running parameters, and generate a single fruit-specific pitting control command; Step S106: Drive the depitting actuator to complete the apple depitting operation according to the depitting control command, collect images of the fruit after depitting and verify the state of fruit pit residue and pulp loss, generate depitting effect data and update the processing database; Step S107: Calculate the positioning deviation value and cutting deviation value based on the core removal effect data, correct the AI ​​positioning model parameters and core removal control logic through a feedback loop mechanism, regenerate the positioning result and control command, and optimize the adaptive core removal process.

2. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S101 includes: The apple is rotated at a constant speed by a rotating platform, and a multi-view industrial camera is used to continuously capture images of the apple's appearance. Simultaneously, a near-infrared transmission imaging device is used to capture transmission images of the apple's interior, thus obtaining the original appearance image sequence and the original transmission image. For each frame of the original appearance image sequence and the original transmission image, a filtering algorithm is used to remove image noise and background interference to generate a first preprocessed image sequence; Each frame of the first preprocessed image sequence is subjected to contrast enhancement processing to adjust the pixel grayscale distribution range, expand the grayscale difference between the pulp and the seed region, and generate a second preprocessed image sequence. The second preprocessed image sequence is filtered for sharpness, retaining images with a Tenengrad gradient sharpness value greater than or equal to a preset sharpness threshold, and removing blurry frames with sharpness values ​​lower than the threshold to generate a preliminary effective image set; The initial effective image set is then standardized in size and format to unify image resolution and data format, generating a standardized multimodal image dataset.

3. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S102 includes: For the appearance images in the standardized multimodal image dataset, a convolutional neural network is used to extract the surface contour of the fruit, the position of the fruit stalk, the position of the calyx, and the curvature features of the fruit shape. Calculate the maximum diameter of a single fruit, its axial height, and the offset angle of the fruit's central axis based on the fruit's surface contour. For the transmission images in the standardized multimodal image dataset, an edge detection algorithm is used to extract the projection contour of the fruit pit, and the maximum projection diameter, axial projection length and projection eccentricity of the fruit pit are calculated. By combining the fruit's central axis from the external image with the projected outline of the pit from the transmission image, the radial and axial offsets of the pit relative to the fruit are determined. By integrating the geometric parameters of the appearance of a single fruit with the projection feature parameters of the kernel, a pixel-level feature set is generated, providing basic data for subsequent spatial registration.

4. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S103 includes: A three-dimensional reference coordinate system for a single fruit is established with the geometric center point of the fruit as the origin and the central axis of the fruit's axial direction as the Z-axis. Based on the 3D reconstruction results of multi-view appearance images, the surface contour features of the fruit are mapped to a 3D reference coordinate system to determine the 3D point cloud distribution of the outer surface of the fruit. The projected outline of the fruit pit in the transmission image is mapped to a three-dimensional reference coordinate system through perspective transformation, and the initial three-dimensional envelope surface of the fruit pit is fitted by combining multi-view transmission data. Calculate the center coordinates, axial length, radial diameter and spatial orientation deviation of the initial three-dimensional envelope of the kernel, and extract the spatial distribution characteristics of the kernel. The spatial distribution characteristics of the fruit pit are associated with the three-dimensional point cloud data of the outer surface of the fruit according to spatial coordinates to construct the initial three-dimensional spatial feature dataset of the fruit pit.

5. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S104 includes: The initial feature dataset of the three-dimensional space of the fruit kernel is input into the AI ​​localization model, which is a multimodal fusion deep learning network; The deep correlation between the geometric features of the fruit's appearance, the surface curvature distribution, and the spatial distribution of the internal kernel is explored through the multimodal fusion deep learning network. Based on the correlation law, the initial three-dimensional envelope surface of the kernel is corrected and finely fitted, and the complete three-dimensional spatial parameters of the kernel are output, including the three-dimensional coordinates of the kernel center, the total axial length, the maximum radial diameter, the spatial deflection angle and the compensation amount for irregular contours. The complete three-dimensional spatial parameters of the fruit pit are bound to the three-dimensional reference coordinate system to generate the accurate positioning result of the fruit pit and output it to the subsequent trajectory planning stage.

6. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S105 includes: Retrieve the three-dimensional dimensions and spatial position parameters of the fruit pit from the precise positioning results of the fruit pit, and combine them with the preset minimum safe cutting allowance to determine the nominal diameter and total feed stroke of the pitting tool; Based on the spatial deflection angle of the pit and the offset angle of the fruit's central axis, calculate the feed angle and centering offset of the pitting tool, and determine the tool's feed start and end coordinates. Based on tool parameters and feed coordinates, a continuous cutting trajectory is planned, and the corresponding feed rate and spindle speed parameters are matched to generate a single-fruit-specific de-core control command. The core removal control command includes tool diameter selection signal, servo axis feed coordinate, feed speed and spindle speed parameters, which are used to directly drive the core removal actuator.

7. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S106 includes: The de-pitting control command is transmitted to the servo control system of the de-pitting actuator, which drives the tool to complete the feed cutting and retraction actions according to the planned trajectory, thereby realizing the de-pitting of a single fruit; After pitting, cross-sectional and internal images of the pitted fruit are captured by a re-inspection camera to identify the area of ​​residual pit and the cutting boundary; Based on the image recognition results, the seed residue rate and pulp loss rate of a single fruit are calculated, and corresponding seed removal effect data are generated. The kernel removal effect data is associated and bound with the original positioning parameters and cutting parameters of the corresponding apple, stored in the processing database, and a data update log is generated. If the pit residue rate or pulp loss rate exceeds the preset threshold, an abnormal warning signal is triggered, the fruit is marked as unqualified and transferred to the re-inspection process; if neither the pit residue rate nor the pulp loss rate exceeds the preset threshold, the fruit is judged to be of qualified pitting quality, and the corresponding data is included in the normal processing sample set.

8. The AI-based adaptive positioning method for apple core removal according to claim 1, characterized in that, Step S107 includes: The actual distribution parameters of the fruit pits corresponding to qualified fruits are retrieved from the processing database, and the difference between them and the predicted positioning results output by the AI ​​positioning model is calculated to obtain the positioning deviation value. Based on the positioning deviation value and the preset deviation correction rules, the parameter adjustment amount of the AI ​​positioning model is calculated, the model weights are fine-tuned through online learning, and the fruit kernel positioning output result is updated. By combining the positioning deviation value with the actual cutting effect, the cutting parameter compensation amount is calculated, and the tool feed depth and centering position parameters in the nucleation control command are corrected. Based on the corrected positioning results and cutting parameters, the de-pitting control command is regenerated, driving the de-pitting actuator to complete the subsequent de-pitting operation of the fruit, forming a closed-loop optimized adaptive de-pitting process.

9. An AI-based adaptive positioning system for apples after core removal, characterized in that, The system for implementing the AI-based apple core-removal adaptive positioning method as described in any one of claims 1-8, the system comprising: The multimodal acquisition module is used to acquire raw images of the apple's exterior and interior through a multi-view camera and a near-infrared imaging device. The images are then processed sequentially with filtering and noise reduction, contrast enhancement, sharpness filtering, and standardization to output a standardized multimodal image dataset. The feature extraction module is used to extract the geometric features of fruit appearance and the projection features of the core from the standardized multimodal image dataset, calculate the diameter of a single fruit, fruit shape parameters, core projection size and offset, and generate a pixel-level feature set. The spatial registration module is used to establish a three-dimensional reference coordinate system for a single fruit, complete the spatial mapping and registration of appearance features and internal transmission features, and construct the initial three-dimensional spatial feature dataset of the fruit core. The AI ​​localization module is used to input the initial feature dataset of the fruit pit's three-dimensional space into the multimodal fusion deep learning network, calculate the complete three-dimensional space parameters of the fruit pit, and output the accurate localization result of the fruit pit; The trajectory planning module is used to adaptively plan the parameters of the de-pitting tool and the cutting trajectory based on the precise positioning results of the fruit pit and the appearance parameters of the single fruit, and generate a dedicated de-pitting control command for each single fruit. The depitting verification module is used to drive the actuator to complete the depitting operation according to the depitting control command, collect images of the fruit after depitting to verify the residual and loss status, generate depitting effect data and update the processing database. The feedback optimization module is used to calculate the positioning and cutting deviations based on the core removal effect data, and correct the AI ​​model parameters and core removal control logic through a feedback loop mechanism to optimize the adaptive core removal process.

Images