Intelligent fine sorting and drying equipment for apple slices and fine sorting method of intelligent fine sorting and drying equipment

Abstract

The intelligent apple slice fine sorting and drying equipment comprises a conveying mechanism, a height limiting mechanism is mounted at the feeding end of the conveying mechanism, the conveying mechanism comprises a mounting side plate, and a conveying belt is mounted on the inner side of the mounting side plate; a plurality of protruding strip-shaped structures are arranged on the outer surface of the conveying belt at equal intervals. The intelligent selection area comprises a multispectral imaging mechanism, a color grading mechanism and a shape screening mechanism. Multi-spectral fusion detection: RGB + NIR + HDR imaging is combined to improve a YOLOv7 model, the defect detection rate is improved, the error rejection rate is reduced, dynamic color grading, HSV and spectral feature splicing + FCM clustering are adopted, incremental learning is supported, variety differences are adapted, the color sorting accuracy is improved, and the color sorting efficiency is improved. And meanwhile, 3D shape screening, laser triangulation and Poisson curved surface reconstruction are adopted, so that the warping degree detection precision is improved, and the rejection rate of deformed pieces is increased.

Description

Technical Field

[0001] This application relates to the field of apple coarse processing technology, specifically to an intelligent fine sorting and drying device for apple slices and its fine sorting method. Background Technology

[0002] The production process of dried apples usually includes raw material pretreatment: washing and peeling: apples are washed with a brush roller washing machine to remove surface dirt, peeled mechanically or by steam, slicing and core removal: rotating blades cut into ring slices with a thickness of 4±0.5mm, and water jetting to remove the core;

[0003] Color protection and dehydration: Rinsing and enzyme inactivation: Soaking in citric acid + ascorbic acid solution to inhibit polyphenol oxidase activity; Centrifugation and dehydration: Centrifugation at 600 r / min for 3 minutes to reduce moisture content to 65%-70%;

[0004] Drying and shaping in stages: First stage: 60℃ hot air drying for 2 hours to soften cell walls; Second stage: 55℃ gradient cooling drying for 4 hours to prevent surface hardening;

[0005] Sorting and Packaging: Manual Sorting: Traditional production lines rely on workers to remove defective pieces (black spots / insect holes / deformation) and finally package them in fixed quantities.

[0006] In the processing of dried apples, traditional sorting mainly relies on manual labor or single optical sorting equipment, which has the following drawbacks:

[0007] Insufficient efficiency and accuracy: Manual sorting is slow, highly subjective, and prone to missing minor defects (such as wormholes and black spots); single RGB sorting is difficult to identify internal moisture distribution or texture defects, resulting in a high defect rate.

[0008] Missing multidimensional features: Existing equipment struggles to simultaneously analyze color, shape, and spectral characteristics. For example, color grading relies on fixed thresholds, making it unable to adapt to color gamut variations among different apple varieties (such as the color difference between Fuji and Gala).

[0009] Shape detection often uses 2D images, ignoring thickness and warpage (such as uneven drying caused by deformed sheets);

[0010] Process fragmentation: Sorting and drying are carried out in separate stages. Secondary handling after sorting increases the fragmentation rate, and deteriorated pieces are not removed before drying, increasing energy consumption. Summary of the Invention

[0011] Therefore, this application provides an intelligent fine sorting and drying device for apple slices and a fine sorting method thereof to solve the problems existing in the prior art.

[0012] To achieve the above objectives, this application provides the following technical solution:

[0013] Firstly, an intelligent fine-selection and drying device for apple slices includes:

[0014] The conveying mechanism includes a height limiting mechanism installed at its feed end. The conveying mechanism includes a mounting side plate, and a conveyor belt is installed on the inner side of the mounting side plate. The outer surface of the conveyor belt is provided with several raised strip structures at equal intervals, and the height of the raised strip structures is smaller than the thickness of the apple slices. The outer side of the conveyor belt of the height limiting mechanism is provided with several raised strip structures. The upper sides of the height limiting mechanism and the conveying mechanism are both conveyed to the same side, so that the apple slices are input into the intelligent selection area in a single layer.

[0015] The intelligent selection area includes a multispectral imaging mechanism, a color grading mechanism, and a shape screening mechanism to implement intelligent selection of apple slices. The corresponding mechanism drives the spray valve installed on the inner wall of the installation side plate to reject defective products, while good products enter the drying mechanism installed at the discharge end for drying.

[0016] Optionally, the spray valve is characterized by being inclined.

[0017] Optionally, the multispectral imaging mechanism is characterized by comprising an RGB camera, a NIR camera, an HDR camera, and a synchronization trigger, and employs a spray valve correspondingly disposed on the lower side of the multispectral imaging mechanism as the actuator.

[0018] Secondly, an intelligent and precise method for slicing apples includes the following steps:

[0019] S1: Apple slice pretreatment and conveying. After slicing, the apple slices are rinsed with water, centrifuged to dehydrate and air-dried to remove surface moisture and impurities. The initial sugar content and firmness are quickly detected by a near-infrared spectrometer, and slices with potential spoilage risk are marked and conveyed in a single layer to the intelligent selection area.

[0020] S2: First-level sorting, multispectral imaging mechanism captures apple slice images, YOLOv7 model marks defect areas, and spray valve precisely removes unqualified slices;

[0021] S3: Secondary sorting, the FCM algorithm performs color clustering on qualified pieces and sorts them into compartments according to preset grades;

[0022] S4: Three-level sorting, combining 3D structured light data with warpage threshold to remove deformed pieces;

[0023] S5: Standard-compliant apple slice input drying mechanism.

[0024] Optionally, the first-level sorting in S2 includes the following steps:

[0025] S2.1: Data Preprocessing:

[0026] RGB image: The CLAHE algorithm is applied to enhance local contrast and eliminate shadow interference;

[0027] NIR image: The absorption peak at 970 nm wavelength was extracted by derivative spectroscopy to generate a thermogram of water distribution.

[0028] HDR images: merge images with different exposure times to generate a high dynamic range grayscale image, highlighting texture details;

[0029] Then, the weight map for each pixel is calculated:

[0030] Weight = Contrast Weight ×Saturation Weight ×Exposure Weight

[0031] The two images are merged by weighted average to generate an HDR grayscale image;

[0032] S2.2: Constructing a defect detection model:

[0033] Model architecture: An improved YOLOv7-tiny network is adopted, and the backbone network is replaced with ShuffleNetV2 to reduce the amount of computation. The detection head outputs three types of defects.

[0034] S2.3: Implementation of Decisions:

[0035] The model outputs a defect bounding box, which is mapped to the valve array coordinate system to determine the valve number to be triggered. When the defect area accounts for ≥5% or a critical defect is detected, the corresponding valve is triggered immediately.

[0036] Optionally, the secondary sorting in S3 includes the following steps:

[0037] S3.1: Color space conversion:

[0038] The RGB image is converted to the HSV color space, and the hue (H), saturation (S), and brightness (V) features are extracted. Standardization processing is then performed: the H channel is normalized to 0–1, and the S and V channels are linearly scaled to 0–1. The H and S features are then extracted. normalized V normalized A 3-dimensional feature vector is formed by collecting reflectance in the wavelength range of 450-650nm, taking one point every 10nm, for a total of 21 features. The HSV features and spectral features are concatenated into a 24-dimensional vector (3+21), which is then input into the FCM clustering algorithm.

[0039] S3.2: FCM clustering:

[0040] Parameter settings: number of clusters C=5, dynamically adjusted by elbow rule and contour coefficient, fuzziness factor m=2, iteration stopping threshold ε=1e-5, K-means++ algorithm is used to generate initial cluster centers to avoid local optima, and enhanced feature vectors are constructed by combining reflectance data measured by spectrometer.

[0041] S3.3: Hierarchical decision-making:

[0042] The clustering results are mapped to preset color levels. Apple slices with cluster centers located within the preset HSV range, H<0.1 or S<0.2 (i.e., gray or black), are marked as needing re-inspection.

[0043] S3.4: Execute decision, the model outputs the defect bounding box, maps it to the valve array coordinate system, determines the valve number to be triggered, and immediately triggers the corresponding valve when the defect area accounts for ≥5% or a critical defect is detected.

[0044] Optionally, the process of generating initial cluster centers using the K-means++ algorithm includes:

[0045] Membership matrix calculation:

[0046] u ij =1 / Σ k (||x i -c k || / ||x i -c j ||)^(2 / (m-1))

[0047] Among them, u ij c represents the membership degree of sample i to cluster j. j As cluster center;

[0048] Clustering center update:

[0049] c j =Σ i (u ij ^m*x i ) / Σ i (u ij ^m)

[0050] Repeat the above calculations until the stopping condition is met.

[0051] Optionally, the three-level sorting in S4 includes the following steps:

[0052] S4.1: Point Cloud Generation:

[0053] Using the principle of laser triangulation, the deformation of laser stripes captured by the camera is converted into height data to generate point clouds on the surface of the apple slice.

[0054] S4.2: Surface Reconstruction

[0055] Statistical filtering and radius filtering are applied to remove outliers whose average distance from the neighborhood exceeds 5σ. A Poisson surface reconstruction algorithm is used to generate a mesh model and calculate the thickness and warpage.

[0056] S4.3: Shape Execution Decision:

[0057] Threshold setting: thickness 4±0.5mm, warpage ≤1mm. When the thickness or warpage exceeds the standard, it is mapped to the spray valve array coordinate system to determine the spray valve number to be triggered, and the corresponding spray valve is triggered immediately.

[0058] Optionally, S3.3 further includes a dynamic adjustment strategy, including:

[0059] Establish a mapping table between apple varieties and color thresholds, monitor the color of apple slices after sorting in real time using a colorimeter, and automatically adjust the cluster centers if the deviation exceeds ΔE2.0.

[0060] Meanwhile, an incremental FCM algorithm is used to update the cluster centers every 1000 apples processed, adapting to color differences between batches.

[0061] Compared with the prior art, this application has at least the following beneficial effects:

[0062] 1. Three-level intelligent sorting, improving accuracy:

[0063] Multispectral fusion detection: RGB+NIR+HDR imaging combined with an improved YOLOv7 model improves defect detection rate (black spots / insect holes / scorching), reduces false rejection rate, and adopts dynamic color grading, HSV and spectral feature splicing + FCM clustering, supports incremental learning, adapts to variety differences, and improves color sorting accuracy. At the same time, 3D shape screening is adopted: laser triangulation + Poisson surface reconstruction improves warpage detection accuracy and increases the rejection rate of deformed pieces.

[0064] 2. Anti-stacking design: The height limiting mechanism, together with the partitioned protrusions of the conveyor belt, ensures single-layer feeding, direct connection between sorting and drying, and the tilted setting of the spray valve accurately removes defective products, while good products directly enter the drying mechanism. Attached Figure Description

[0065] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0066] Figure 1 A schematic diagram of the overall structure of an intelligent fine sorting and drying device for apple slices provided in one embodiment of this application;

[0067] Figure 2A schematic diagram of the overall structure of the connection between the height limiting mechanism and the mounting side plate of an intelligent fine sorting and drying device for apple slices provided in one embodiment of this application;

[0068] Figure 3 A flowchart illustrating an intelligent and precise method for slicing apples, provided as an embodiment of this application;

[0069] Figure 4 A flowchart of a first-stage sorting method for intelligent and precise apple slices provided in one embodiment of this application;

[0070] Figure 5 This is a flowchart illustrating a two-stage sorting method for intelligent and precise apple slices, provided as an embodiment of this application.

[0071] Explanation of reference numerals in the attached figures:

[0072] 1. Conveying mechanism; 101. Mounting side plate; 102. Conveyor belt; 2. Height limiting mechanism; 3. Multispectral imaging mechanism; 4. Color grading mechanism; 5. Shape screening mechanism; 6. Drying mechanism; 7. Spray valve. Detailed Implementation

[0073] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0074] This invention provides an intelligent fine sorting and drying device for apple slices, comprising: a conveying mechanism 1, a height limiting mechanism 2 installed at the feed end of the conveying mechanism 1, the conveying mechanism 1 including a mounting side plate 101, a conveyor belt 102 installed on the inner side of the mounting side plate 101, and a plurality of raised strip-shaped structures evenly spaced on the outer surface of the conveyor belt 102, wherein the height dimension of the raised strip-shaped structures is smaller than the thickness dimension of the apple slice, thereby dividing the conveyor belt 102 into several regions, so that only a single apple slice can be laid in each region. The height limiting mechanism 2 has a plurality of raised strip structures on the outer side of the conveyor belt, limiting... The upper sides of both the height limiting mechanism 2 and the conveying mechanism 1 are conveyed to the same side, so that the lower side of the height limiting mechanism 2 moves in the opposite direction to the upper side of the conveying mechanism 1. This allows the height limiting mechanism 2 to move apple slices whose stacking height is higher than the outer protruding strip structure of the conveyor belt 102 in the opposite direction of the conveying mechanism 1. When the apple slices are conveyed to the outer protruding strip structure of the next conveyor belt 102, the apple slices that are higher than the outer protruding strip structure of the conveyor belt 102 fall into the area formed there, thereby avoiding the stacking of apple slices and enabling them to be input into the intelligent selection area in a single layer.

[0075] The intelligent selection area includes a multispectral imaging mechanism 3, a color grading mechanism 4, and a shape screening mechanism 5, which enables intelligent selection of apple slices. The spray valve 7 installed on the inner wall of the installation side plate 101 drives the defective products to be rejected, and the good products enter the drying mechanism 6 installed at the discharge end of 1 for drying.

[0076] The spray valve 7 is set at an angle, which makes it easy to bring the airflow close to the surface of the conveyor belt 102, so that the airflow can blow into the gap between the apple slices and the conveyor belt 102, thereby blowing the apple slices up and blowing them into the waste trough more effectively.

[0077] Specifically, the multispectral imaging mechanism 3 consists of an RGB camera, an NIR camera, an HDR camera, and a synchronization trigger, and uses the spray valve 7 correspondingly set on the lower side of the multispectral imaging mechanism 3 as the actuator.

[0078] Furthermore, the multispectral imaging mechanism 3 is also equipped with a light source optimization mechanism, which uses an LED ring light source. The vertical distance between the light source and the apple slice conveyor belt is 150mm to ensure that the light uniformly covers the detection area. A diffuse reflection plate made of microbead coated polycarbonate board is installed. The light source and the apple slice are at a 45° angle to eliminate direct reflected light and form a diffuse reflection environment.

[0079] After primary sorting by the multispectral imaging mechanism 3, the product enters the color grading mechanism 4 for secondary sorting, and finally the shape screening mechanism 5 for tertiary sorting before entering the drying mechanism 6 for drying.

[0080] In one embodiment, an intelligent and precise method for slicing apples is provided, comprising the following steps:

[0081] S1: Apple slice pretreatment and conveying. After slicing, the apple slices are rinsed with water, centrifuged to dehydrate and air-dried to remove surface moisture and impurities, and to avoid surface reflection interfering with color sorting. The initial sugar content and firmness are quickly detected by a near-infrared spectrometer, and slices with potential spoilage risk are marked and conveyed to the intelligent sorting area.

[0082] Specifically, infrared drying can be used to reduce the moisture content of apple slices to ≤15%, and then the apple slices are conveyed in a single layer to the intelligent selection area;

[0083] S2: First-level sorting (defect removal), multispectral imaging mechanism captures apple slice images, YOLOv7 model marks defect areas, spray valve 7 accurately removes unqualified slices;

[0084] Specifically, primary sorting includes the following steps:

[0085] S2.1: Data Preprocessing:

[0086] RGB image: The CLAHE algorithm is applied to enhance local contrast and eliminate shadow interference. Specifically, the image is divided into 8×8 non-overlapping blocks, a histogram is calculated for each block and clipped to the CLIP LIMIT, pixel values ​​are redistributed, and bilinear interpolation is used to restore image resolution and eliminate block artifacts.

[0087] NIR Image: The absorption peak at 970nm wavelength is extracted using derivative spectroscopy to generate a moisture distribution heatmap. Specifically, the NIR image is wavelength-calibrated to ensure the accuracy of the 970nm channel, the image gradient (Sobel operator) is calculated, the absorption peak intensity is extracted, a moisture distribution heatmap is generated, and it is mapped to a grayscale value of 0-255.

[0088] HDR images: merge images with different exposure times to generate a high dynamic range grayscale image, highlighting texture details. Specifically, for low-exposure images (0.5ms), highlight details are preserved, and for high-exposure images (5ms), shadow details are preserved.

[0089] Then, the weight map for each pixel is calculated:

[0090] Weight = Contrast Weight ×Saturation Weight ×Exposure Weight

[0091] The two images are merged by weighted average to generate an HDR grayscale image.

[0092] S2.2: Constructing a defect detection model:

[0093] Model architecture: An improved YOLOv7-tiny network is adopted, and the backbone network is replaced with ShuffleNetV2 to reduce the amount of computation. The detection head outputs three types of defects (black spots, insect holes, and scorch).

[0094] The specific configuration is as follows:

[0095] Backbone network: ShuffleNetV2 (version 1.0), configured as follows:

[0096] Phase 1: Conv2d(3×3, stride=2)→BatchNorm→ReLU.

[0097] Phase 2: MaxPool(3×3, stride=2) → 2 ShuffleUnits (112 channels).

[0098] Phase 3: 4 ShuffleUnits (224 channels).

[0099] Phase 4: 4 ShuffleUnits (448 channels).

[0100] Detection head:

[0101] Small target detection head: input scale 1 / 32, convolution kernel 3×3, output number of categories 3 (black spots, insect holes, scorch).

[0102] Medium target detection head: input scale 1 / 16, convolution kernel 3×3, output number of classes 3.

[0103] Large target detection head: input scale 1 / 8, convolution kernel 3×3, output number of classes 3.

[0104] SE Attention Module: An SE module is inserted before each detector head to enhance the correlation of features between channels.

[0105] Training strategy:

[0106] Dataset: 10,000 apple slice images were collected (6,000 Fuji slices and 4,000 Gala slices), divided into a training set of 8,000 slices, a validation set of 1,000 slices, and a test set of 1,000 slices. Defect areas were manually labeled. Data augmentation included random rotation (±15°), scaling (0.8 to 1.2 times), and color jitter (brightness ±0.2, contrast ±0.3).

[0107] Pre-training: Load the weights pre-trained on the COCO dataset, freeze the first 3 layers of the backbone network, and fine-tune the detection head.

[0108] Loss function: Combining CIoU loss (localization) and Focal Loss (classification) to solve the problem of imbalanced samples.

[0109] Inference optimization: Deployed to the NVIDIA Jetson AGX Orin platform, accelerated by TensorRT, inference speed ≥50fps.

[0110] S2.3: Implementation of Decisions:

[0111] The model outputs a defect bounding box, which is mapped to the coordinate system of the spray valve 7 array to determine the number of the spray valve 7 to be triggered. When the defect area accounts for ≥5% or a critical defect (such as an insect eye) is detected, the corresponding spray valve 7 is triggered immediately.

[0112] S3: Secondary sorting (color grading), the FCM algorithm performs color clustering on qualified pieces and sorts them into compartments according to preset grades;

[0113] Specifically, secondary sorting includes the following steps:

[0114] S3.1: Color space conversion:

[0115] The RGB image is converted to the HSV color space, and the hue (H), saturation (S), and brightness (V) features are extracted and normalized: the H channel is normalized to 0-1, and the S and V channels are linearly scaled to 0-1. H, S_normalized, and V_normalized are extracted to form a 3-dimensional feature vector. The reflectance in the wavelength range of 450-650nm is collected, with one point taken every 10nm, for a total of 21 features. The HSV features and spectral features are concatenated into a 24-dimensional vector (3+21), which is then input into the FCM clustering algorithm.

[0116] S3.2: FCM clustering:

[0117] Parameter settings: number of clusters C = 5 (light golden yellow, golden yellow, amber, dark amber, abnormal color), dynamically adjusted by elbow rule and contour coefficient, fuzz factor m = 2, iteration stop threshold ε = 1e-5;

[0118] Initialization: The K-means++ algorithm is used to generate initial cluster centers to avoid local optima;

[0119] The algorithm process includes:

[0120] Membership matrix calculation:

[0121] u ij =1 / Σ k (||x i -c k || / ||x i -c j ||)^(2 / (m-1))

[0122] Among them, u ij c represents the membership degree of sample i to cluster j. j It serves as the cluster center.

[0123] Clustering center update:

[0124] c j =Σ i (u ij ^m*x i ) / Σ i (u ij ^m)

[0125] Repeat the above calculations until the stopping condition is met.

[0126] Feature fusion: Combine reflectance data (450-650nm) measured by a spectrometer to construct an enhanced feature vector.

[0127] S3.3: Hierarchical decision-making:

[0128] Grade mapping: The clustering results are mapped to a preset color grade. The cluster center is located in the preset HSV range (e.g., light golden yellow: H∈[0.1,0.3], S∈[0.5,0.8]). Apple slices with H<0.1 or S<0.2, i.e. gray or black, are marked as needing re-inspection.

[0129] Dynamic adjustment: Establish a mapping table between apple varieties and color thresholds (e.g., the light golden yellow H range for Fuji apples [0.15, 0.25]). Monitor the color of apple slices after sorting in real time using a Konica Minolta CR-400 colorimeter. If the deviation exceeds ΔE2.0 (CIEDE2000 standard), automatically adjust the cluster centers.

[0130] The incremental FCM algorithm is used to update the cluster centers every 1000 apples processed to adapt to color differences between batches.

[0131] Anomaly Handling: When the percentage of abnormal colors exceeds 5% for 5 consecutive minutes, the following actions will be triggered:

[0132] Sorting was paused, and 100 samples were collected for offline analysis.

[0133] Adjust the number of clusters C (e.g., increase to 6) or the threshold range based on the analysis results.

[0134] Reinitialize the FCM algorithm and resume sorting;

[0135] S3.4: Execute decision, the model outputs the defect bounding box, maps it to the coordinate system of the spray valve 7 array, determines the number of the spray valve 7 to be triggered, and immediately triggers the corresponding spray valve 7 when the defect area accounts for ≥5% or a critical defect (such as an insect hole) is detected.

[0136] S4: Three-level sorting (shape screening), 3D structured light data combined with warp threshold, to remove deformed pieces;

[0137] Specifically, the three-level sorting includes the following steps:

[0138] S4.1: Point Cloud Generation:

[0139] Using the principle of laser triangulation, the deformation of laser stripes captured by the camera is converted into height data, generating a point cloud (0.1mm spacing) on ​​the surface of the apple slice.

[0140] S4.2: Surface Reconstruction

[0141] Noise reduction: Statistical filtering (mean 5σ) and radius filtering (radius 2mm, minimum number of points 50) are applied to remove outliers whose average distance from the neighborhood exceeds 5σ (σ is the standard deviation of the neighborhood distance).

[0142] Reconstruction: The Poisson surface reconstruction algorithm is used to generate a mesh model and calculate the thickness (maximum, minimum, and mean) and warpage (flatness error).

[0143] The specific parameters are as follows:

[0144] Depth: Set to 8 (2^8 = 256 subdivisions) to balance reconstruction accuracy and computational cost.

[0145] Sampling density: Set to twice the point cloud density (i.e., 0.05mm spacing) to avoid holes.

[0146] Interpolation method: Cubic B-spline interpolation is used to generate a continuous surface.

[0147] S4.3: Shape Execution Decision:

[0148] Threshold setting: thickness 4±0.5mm, warpage ≤1mm (calibrated by micrometer and coordinate measuring machine). When the thickness or warpage exceeds the standard, it is mapped to the coordinate system of the spray valve 7 array to determine the number of the spray valve 7 to be triggered, and the corresponding spray valve 7 is triggered immediately.

[0149] S5: Standard-compliant apple slice input drying mechanism 6.

[0150] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. An intelligent fine sorting and drying device for apple slices, characterized in that, include: The conveying mechanism (1) has a height limiting mechanism (2) installed at the feeding end of the conveying mechanism (1). The conveying mechanism (1) includes a mounting side plate (101). A conveyor belt (102) is installed on the inner side of the mounting side plate (101). A number of raised strip structures are evenly spaced on the outer surface of the conveyor belt (102). The height dimension of the raised strip structure is smaller than the thickness dimension of the apple slice. A number of raised strip structures are provided on the outer side of the conveyor belt of the height limiting mechanism (2). The upper sides of the height limiting mechanism (2) and the conveying mechanism (1) are both conveyed to the same side, so that the apple slices are input into the intelligent selection area in a single layer. The intelligent selection area includes a multispectral imaging mechanism (3), a color grading mechanism (4), and a shape screening mechanism (5) to implement intelligent selection of apple slices. The corresponding mechanism drives the spray valve (7) installed on the inner wall of the installation side plate (101) to eliminate defective products, and good products enter the drying mechanism (6) installed at the discharge end of (1) for drying.

2. The intelligent fine sorting and drying equipment for apple slices according to claim 1, characterized in that, The spray valve (7) is set at an angle.

3. The intelligent fine sorting and drying equipment for apple slices according to claim 1, characterized in that, The multispectral imaging mechanism (3) consists of an RGB camera, a NIR camera, an HDR camera and a synchronization trigger, and uses a spray valve (7) correspondingly set on the lower side of the multispectral imaging mechanism (3) as the actuator.

4. An intelligent and refined method for apple slices, applied to the intelligent and refined apple slice sorting and drying equipment described in claims 1-3, characterized in that, Includes the following steps: S1: Apple slice pretreatment and conveying. After slicing, the apple slices are rinsed with water, centrifuged to dehydrate and air-dried to remove surface moisture and impurities. The initial sugar content and firmness are quickly detected by a near-infrared spectrometer, and slices with potential spoilage risk are marked and conveyed in a single layer to the intelligent selection area. S2: First-level sorting, multispectral imaging mechanism captures apple slice images, YOLOv7 model marks defect areas, spray valve (7) accurately removes unqualified slices; S3: Secondary sorting, the FCM algorithm performs color clustering on qualified pieces and sorts them into compartments according to preset grades; S4: Three-level sorting, combining 3D structured light data with warpage threshold to remove deformed pieces; S5: Standard-compliant apple slice input drying mechanism (6).

5. The intelligent and precise method for slicing apples according to claim 4, characterized in that, The first-level sorting in S2 includes the following steps: S2.1: Data Preprocessing: RGB image: The CLAHE algorithm is applied to enhance local contrast and eliminate shadow interference; NIR image: The absorption peak at 970 nm wavelength was extracted by derivative spectroscopy to generate a thermogram of water distribution. HDR images: merge images with different exposure times to generate a high dynamic range grayscale image, highlighting texture details; Then, the weight map for each pixel is calculated: Weight=Contrast Weight ×Saturation Weight ×Exposure Weight The two images are merged by weighted average to generate an HDR grayscale image; S2.2: Constructing a defect detection model: Model architecture: An improved YOLOv7-tiny network is adopted, and the backbone network is replaced with ShuffleNetV2 to reduce the amount of computation. The detection head outputs three types of defects. S2.3: Implementation of Decisions: The model outputs the defect bounding box, which is mapped to the coordinate system of the spray valve (7) array to determine the number of the spray valve (7) to be triggered. When the defect area accounts for ≥5% or a critical defect is detected, the corresponding spray valve (7) is triggered immediately.

6. The intelligent and precise method for slicing apples according to claim 4, characterized in that, The secondary sorting in S3 includes the following steps: S3.1: Color space conversion: The RGB image is converted to the HSV color space, and the hue (H), saturation (S), and brightness (V) features are extracted. Standardization processing is then performed: the H channel is normalized to 0–1, and the S and V channels are linearly scaled to 0–1. The H and S features are then extracted. normalized V normalized A 3-dimensional feature vector is formed by collecting reflectance in the wavelength range of 450-650nm, taking one point every 10nm, for a total of 21 features. The HSV features and spectral features are concatenated into a 24-dimensional vector (3+21), which is then input into the FCM clustering algorithm. S3.2: FCM clustering: Parameter settings: number of clusters C=5, dynamically adjusted by elbow rule and contour coefficient, fuzziness factor m=2, iteration stopping threshold ε=1e-5, K-means++ algorithm is used to generate initial cluster centers to avoid local optima, and enhanced feature vectors are constructed by combining reflectance data measured by spectrometer. S3.3: Hierarchical decision-making: The clustering results are mapped to preset color levels. Apple slices with cluster centers located within the preset HSV range, H<0.1 or S<0.2 (i.e., gray or black), are marked as needing re-inspection. S3.4: Execute decision, the model outputs the defect bounding box, maps it to the coordinate system of the spray valve (7) array, determines the number of the spray valve (7) to be triggered, and immediately triggers the corresponding spray valve (7) when the defect area accounts for ≥5% or a critical defect is detected.

7. The intelligent and precise method for slicing apples according to claim 6, characterized in that, The process of generating initial cluster centers using the K-means++ algorithm includes: Membership matrix calculation: u ij =1 / Σ k (||x i -c k || / ||x i -c j ||)^(2 / (m-1)) Among them, u ij c represents the membership degree of sample i to cluster j. j As cluster center; Clustering center update: c j =S i (u ij ^m*x i ) / S i (u ij ^m) Repeat the above calculations until the stopping condition is met.

8. The intelligent and precise method for slicing apples according to claim 4, characterized in that, The three-level sorting in S4 includes the following steps: S4.1: Point Cloud Generation: Using the principle of laser triangulation, the deformation of laser stripes captured by the camera is converted into height data to generate point clouds on the surface of the apple slice. S4.2: Surface Reconstruction Statistical filtering and radius filtering are applied to remove outliers whose average distance from the neighborhood exceeds 5σ. A Poisson surface reconstruction algorithm is used to generate a mesh model and calculate the thickness and warpage. S4.3: Shape Execution Decision: Threshold setting: thickness 4±0.5mm, warpage ≤1mm. When the thickness or warpage exceeds the standard, it is mapped to the coordinate system of the spray valve (7) array to determine the number of the spray valve (7) to be triggered, and the corresponding spray valve (7) is triggered immediately.

9. The intelligent and precise method for slicing apples according to claim 6, characterized in that, S3.3 also includes a dynamic adjustment strategy, including: Establish a mapping table between apple varieties and color thresholds, monitor the color of apple slices after sorting in real time using a colorimeter, and automatically adjust the cluster centers if the deviation exceeds ΔE2.

0. Meanwhile, an incremental FCM algorithm is used to update the cluster centers every 1000 apples processed, adapting to color differences between batches.

Images