Intelligent sorting method for defective red dates

Abstract

The invention provides an intelligent sorting method for defective red dates, and relates to the technical field of agricultural product screening, based on a feeding transmission module, target materials are uniformly laid, initial transmission materials are obtained, real-time image acquisition is performed on the initial transmission materials through a visual detection module, transmission material images are obtained, and based on red date defect feature data, the defective red dates are sorted. And training the defect identification module, performing material defect identification according to the defect identification module to obtain a defect identification result, obtaining information of the red dates to be screened according to the defect identification result, and sorting the defective red dates through the sorting control module. The technical problems that in the prior art, due to the fact that the intelligent degree of red date sorting is low, the red date sorting efficiency is low, the cost is high, and the sorting precision is not high are solved, the intelligent degree and the recognition precision of the red date sorting process are improved, efficient and accurate intelligent screening and sorting of red dates are achieved, manual intervention is greatly reduced, and the working efficiency is improved. And the labor cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural product screening, and in particular to an intelligent picking method for defective red dates. Background Art

[0002] In the red date processing industry, ensuring the quality of red dates and effectively removing defective products are the core elements to ensure product quality, meet market demand and improve consumer satisfaction. As a healthy food that is deeply loved by people, the quality of red dates directly affects consumers' willingness to buy and eating experience. For a long time, manual picking of defective red dates has been the main method in the red date processing industry. However, this method has obvious disadvantages: first, manual picking is inefficient and difficult to meet the needs of large-scale production; second, the cost of manual picking is high, which increases the operating costs of enterprises; third, due to the subjectivity and fatigue factors of manual picking, the sorting accuracy of defective red dates is difficult to guarantee, and it is easy to misjudge and miss. In order to overcome the disadvantages of manual picking, more and more red date processing companies have begun to explore mechanical screening methods. However, the existing mechanical screening equipment is generally less intelligent, and can only perform simple physical screening, such as size, weight, etc., and cannot accurately identify defects such as skin damage, insect eyes, and mildew of red dates, resulting in low efficiency, high cost, and low recognition accuracy in the process of red date defect sorting.

[0003] The existing technology has defects. The intelligent level of red date sorting is low, which leads to technical problems such as low efficiency and high cost of red date sorting and low sorting accuracy.

[0004] The technical effect is achieved by improving the intelligence level and recognition accuracy of the red dates sorting process, realizing efficient and accurate intelligent screening and picking of red dates, and greatly reducing manual intervention and reducing labor costs. Summary of the invention

[0005] The present application provides an intelligent picking method for defective red dates, which is used to solve the technical problems in the prior art that the intelligence level of red date sorting is low, resulting in low red date sorting efficiency, high cost, and low sorting accuracy.

[0006] In view of the above problems, the present application provides an intelligent picking method for defective red dates, the method comprising: based on a feeding transmission module, evenly laying the target material to obtain an initial transmission material; through a visual detection module, real-time image acquisition of the initial transmission material to obtain a transmission material image, wherein the visual detection module is arranged directly above the feeding transmission module; acquiring red date defect feature data, and training a defect recognition module based on the red date defect feature data; according to the defect recognition module, performing material defect recognition on the transmission material image to obtain a defect recognition result; through the defect recognition result, obtaining information on red dates to be screened out, and picking defective red dates through a sorting control module.

[0007] In a possible implementation, the following processing is performed: the feeding transmission module is composed of a transmission device and a vibration device; basic information of the target material is obtained; based on the basic information of the target material, the material conveying rate and vibration frequency are configured; according to the material conveying rate and vibration frequency, the target material is evenly laid to obtain the initial transmission material.

[0008] In a possible implementation, the following processing is performed: the visual inspection module includes a plurality of image sensor devices, which are evenly arranged above the feeding transmission module, and the plurality of image sensor devices are provided with position codes; based on the plurality of image sensor devices, real-time image acquisition is performed on the initial transmission material to obtain a transmission material image, wherein the transmission material image includes a plurality of sensor image sets, and each sensor image set is provided with a position identifier.

[0009] Wherein, the image sensing device also includes a TOF camera; The 3D depth image of red dates is collected based on the TOF camera, which captures the depth information of the red dates, that is, the distance from each pixel to the camera, and can accurately capture the shape, three-dimensional structure and surface height changes of the object. This makes it possible to detect morphological features such as bulges, depressions, cracks, and deformations.

[0010] In a possible implementation, the following processing is performed: based on big data, red date defect feature data is crawled; based on the red date defect feature data, image feature data is extracted for preprocessing to obtain sample defect images; based on the sample defect images, a target extraction unit and a defect judgment unit are trained, and the target extraction unit and the defect judgment unit constitute the defect recognition module.

[0011] In a possible implementation, the following processing is performed: pre-labeling the sample defect image to obtain a red date image data set; based on the red date image data set, supervised training is performed in combination with a target detection algorithm to obtain the target extraction unit.

[0012] In a possible implementation, the following processing is performed: based on the red date defect feature data, a plurality of defect judgment features are obtained; based on the plurality of defect judgment features, defects are annotated on the red date image data set to obtain a sample calibration image set, wherein the sample calibration image set includes positive samples and negative samples; based on the sample calibration image set, a convolutional neural network is used for training to obtain the defect judgment unit.

[0013] In a possible implementation, the following processing is performed: the transmitted material image is input into the target extraction unit of the defect recognition module to obtain multiple defect recognition targets; the defect judgment unit receives the multiple defect recognition targets, and performs defect recognition judgment to obtain a defect recognition result, which includes the defective red dates location, defect type and defect level.

[0014] The defect determination unit in the defect recognition module receives multiple defect recognition targets from the target extraction unit, and uses the pre-trained convolutional neural network to make detailed determinations on each defect recognition target, determine the location of defective red dates, the type of defect and the level of the defect, such as color-related defects such as cracks, black spots, insect pests, discoloration, and morphological defects such as protrusions, depressions, cracks, deformations, and the level of the defect such as slight, medium, and severe, and then outputs the defect recognition results, which record the location, defect type and defect level of each defective red date in detail. The model realizes the rapid and accurate identification of red date defects, reduces labor costs, and ensures the intelligence and efficiency of the red date defect recognition process, greatly improving the subsequent sorting efficiency and accuracy.

[0015] In a possible implementation, the following processing is performed: through the defect identification result, the defective red dates location, defect type and defect level are obtained, the screening standards are traversed, and the information of the red dates to be screened out is obtained; based on the information of the red dates to be screened out, a sorting instruction is generated; according to the sorting instruction, the sorting control module is activated to pick the defective red dates.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in the embodiment of the present application is based on a feeding transmission module, which evenly lays the target material to obtain the initial transmission material; through a visual detection module, real-time image acquisition is performed on the initial transmission material to obtain the transmission material image; the defect feature data of red dates is obtained, and a defect recognition module is trained based on the defect feature data of red dates; according to the defect recognition module, material defects are identified for the transmission material image to obtain defect recognition results; through the defect recognition results, information on red dates to be screened out is obtained, and defective red dates are picked through a sorting control module. The intelligent degree and recognition accuracy of the red date sorting process are improved, and efficient and accurate intelligent screening and picking of red dates are realized, and the technical effect of greatly reducing manual intervention and reducing labor costs is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the process flow of an intelligent picking method for defective red dates provided in this application.

[0018] Figure 2 A schematic diagram of the process of training a defect recognition module in an intelligent picking method for defective red dates provided in this application. DETAILED DESCRIPTION

[0019] The present application provides an intelligent picking method for defective red dates, which is used to solve the technical problems of low intelligent sorting of red dates in the prior art, resulting in low efficiency, high cost and low sorting accuracy. The method achieves the technical effect of improving the intelligent level and recognition accuracy of the red date sorting process, realizing efficient and accurate intelligent screening and picking of red dates, and greatly reducing manual intervention and labor costs.

[0020] Below, the technical solutions in the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all of them.

[0021] like Figure 1 As shown, the present application provides an intelligent picking method for defective red dates, the method comprising: Based on the feeding transmission module, the target material is evenly laid to obtain the initial transmission material.

[0022] Specifically, before picking the target material, firstly, based on the feeding transmission module, such as a conveyor belt, the target material red dates are evenly laid on the conveyor belt through the feeding transmission module, and it is ensured that the red dates can move stably on the conveyor belt. After the red dates are evenly laid, the feeding transmission module is started to make the red dates start to move on the conveyor belt, and the red dates that have just started to be transmitted during the transmission process are obtained, that is, the initial transmission material. Evenly laying the red dates facilitates the subsequent visual inspection module to identify, which is the premise for ensuring the accuracy of the inspection. If the red dates are stacked or arranged unevenly, it will affect the effect of the visual inspection.

[0023] The initial transmission material is imaged in real time by a visual inspection module to obtain an image of the transmission material, wherein the visual inspection module is arranged directly above the feeding transmission module.

[0024] Specifically, the visual inspection module is used to collect images of the transmission material in real time, and the visual inspection module is located directly above the feeding transmission module to ensure that the target material image is captured from the best angle. Through the visual inspection module, the image of the initial transmission material is collected in real time to obtain the transmission material image, which refers to the image data obtained by real-time image collection of the initial transmission material on the conveyor belt, that is, the uniformly laid red dates, by the visual inspection module. The transmission material image contains the appearance information of the red dates. The transmission material image is obtained through the visual inspection module to facilitate the subsequent quality analysis of the red dates.

[0025] Obtain red date defect feature data, and train a defect recognition module based on the red date defect feature data.

[0026] Specifically, based on historical data and expert experience, various defect feature data of red dates are collected, including insect bites, cracks, discoloration, surface damage, etc., and the acquired red date defect feature data are summarized to construct a defect feature data set for training. Then, the defect feature data set of red dates is used to train the defect recognition module using a machine learning algorithm, such as a convolutional neural network, a support vector machine, etc. After the training is completed, a recognition module that can accurately identify red date defects in the image is formed. Obtaining a defect recognition module through training can improve the efficiency and accuracy of red date defect recognition, thereby achieving efficient and accurate recognition of red date defects.

[0027] According to the defect recognition module, material defect recognition is performed on the transmitted material image to obtain a defect recognition result.

[0028] Specifically, the transmission material image collected by the visual inspection module is input into the defect recognition module, and the trained defect recognition module is used to perform real-time analysis on the red dates in the transmission material image. According to the analysis results of the model, the various defect characteristics of the red dates are identified and judged, and the corresponding defect recognition results are generated. The defect recognition of the red dates that currently need defect recognition is performed through the model, which improves the efficiency and accuracy of the defect recognition of the red dates. Further, it can be determined which red dates are defective according to the defect recognition results, and then it can be determined which red dates need to be screened, thereby improving the screening efficiency and accuracy.

[0029] The information of red dates to be screened out is obtained through the defect identification results, and the defective red dates are sorted through the sorting control module.

[0030] Specifically, according to the defect recognition results output by the defect recognition model, the information of the red dates to be screened out, that is, the red dates with defects, is determined. After obtaining the information of the red dates to be screened, the sorting control module automatically picks the defective dates according to the information of the red dates to be screened, separates them from the normal red dates in the transmission path, and realizes the picking of defective red dates. The sorting control module is responsible for automatically screening the red dates according to the defect recognition results, including the control module. The sorting control module generates a sorting instruction based on the information of the red dates to be screened obtained from the defect recognition results, and then according to the sorting instruction, the control module executes the picking of defective red dates and separates the defective red dates. The sorting control module can quickly and accurately pick out defective red dates, reducing the workload of manual inspection and labor intensity. At the same time, it also reduces the misjudgment caused by human factors, and improves the efficiency and accuracy of red date sorting.

[0031] Furthermore, based on the feeding transmission module, the target material is evenly laid to obtain the initial transmission material, including: the feeding transmission module is composed of a transmission device and a vibration device; basic information of the target material is obtained; based on the basic information of the target material, the material conveying rate and vibration frequency are configured; according to the material conveying rate and vibration frequency, the target material is evenly laid to obtain the initial transmission material.

[0032] Specifically, the feeding transmission module is used to transmit the target material and ensure that the target material can be evenly distributed during the transmission process. The feeding transmission module includes a transmission device and a vibration device. The transmission device is the main part of the feeding transmission module, which is composed of a conveyor belt and is used to carry and transport materials. The vibration device is a vibration motor or other vibration equipment installed below or on the side of the transmission device. By generating vibration, it prevents the material from piling up on the conveyor belt and ensures that the material can be evenly distributed. The frequency and amplitude of the vibration can be adjusted according to actual needs.

[0033] Before the transmission work, the basic information of the target material is obtained. The basic information includes but is not limited to the color, size, shape, surface characteristics, etc. of the target material, that is, the basic information related to the red dates is obtained, and according to the basic information of the red dates and the actual situation, the material conveying rate and vibration frequency are automatically or manually configured. The conveying rate determines the moving speed of the material on the transmission device, and the vibration frequency affects the distribution uniformity of the material during the transmission process. After determining the appropriate conveying rate and vibration frequency of the red dates material, the feeding transmission module is started, the red dates are evenly laid out, and the initial transmission material is obtained. By configuring the feeding transmission module according to the basic information of the target material, effective management and control of the material is achieved, which lays the foundation for the efficient operation of the entire automated sorting system, so that the subsequent visual inspection and sorting process can proceed smoothly.

[0034] Further, the visual inspection module is used to collect real-time images of the initial transmission material to obtain the transmission material image, including: the visual inspection module includes a plurality of image sensor devices, which are evenly arranged just above the feeding transmission module, and the plurality of image sensor devices are provided with position codes; based on the plurality of image sensor devices, the initial transmission material is collected in real time to obtain the transmission material image, the transmission material image includes a plurality of sensor image sets, and each sensor image set is provided with a position identifier; Wherein: the image sensing device also includes a TOF camera; The 3D depth image of red dates is collected based on the TOF camera to capture the surface morphology and depth information of the red dates.

[0035] Specifically, the visual inspection module achieves subsequent rapid and accurate evaluation of defects in the initially transmitted materials through high-precision image acquisition. The visual inspection module contains multiple image sensor devices, such as CCD, CMOS cameras, etc., which can capture high-resolution RGB images to record information such as color and texture. These image sensor devices are evenly installed directly above the feeding transmission module to ensure that the entire conveyor belt width can be covered, thereby capturing all areas during the material transmission process. Each image sensor device is equipped with a position code to identify its position in the visual inspection module. The position code is a physical label or other formed identifier.

[0036] Multiple image sensor devices perform real-time image acquisition of the initial transmission material to obtain multiple image data, namely, the transmission material image. The transmission material image includes multiple sensor image sets, each of which is captured by a specific image sensor device, and each sensor image set is marked with a position. The visual inspection module including multiple image sensor devices can perform comprehensive and real-time image acquisition of the initial transmission material, generate multi-dimensional transmission material image data, including RGB images and depth images, and improve the accuracy and reliability of subsequent identification of target material defects.

[0037] In one embodiment, Figure 2 As shown, the defect feature data of red dates are obtained, and based on the defect feature data of red dates, a defect recognition module is trained, including: crawling the defect feature data of red dates based on big data; extracting image feature data based on the defect feature data of red dates for preprocessing to obtain sample defect images; based on the sample defect images, training a target extraction unit and a defect judgment unit, and the target extraction unit and the defect judgment unit constitute the defect recognition module.

[0038] Specifically, first, appropriate data sources are selected, such as agricultural product e-commerce websites, agricultural databases, professional forums, etc., and a large amount of defect feature data of red dates is collected and crawled from multiple appropriate data sources, including various defect types, such as breakage, discoloration, insect bites, cracks, etc. Then, the extracted image feature data is preprocessed, including denoising, grayscale, contrast enhancement and other operations, to obtain clear sample defect images, namely sample defect images.

[0039] Furthermore, the sample defect images are used to train a target extraction unit, which uses a convolutional neural network (CNN) to perform defect judgment based on depth information and RGB image features (defect features include color abnormalities, shape abnormalities (such as depressions, protrusions), cracks, spots, etc.) to identify and extract specific defect features of red dates from the image; and the sample defect images are used to train a defect judgment unit, which uses a convolutional neural network (CNN) to analyze the extracted defect features based on depth information and RGB image features and make defect judgments.

[0040] Finally, the trained target extraction unit and defect judgment unit are integrated to form a defect recognition module, which is used to detect and identify red date defects in real time. By crawling red date defect feature data based on big data, extracting image feature data for preprocessing, obtaining sample defect images, training target extraction units and defect judgment units, and forming a defect recognition module, it is possible to achieve efficient and accurate recognition of red date defects, reduce labor costs, and ensure the intelligence and efficiency of the red date defect recognition process.

[0041] Furthermore, based on the sample defect image, a target extraction unit is trained, including: pre-labeling the sample defect image to obtain a red date image data set; based on the red date image data set, supervised training is performed in combination with a target detection algorithm to obtain the target extraction unit.

[0042] Specifically, the automatic pre-labeling process is used to label the sample defect image data obtained after preprocessing, indicating the location and category of the red date defects in the sample set, and the labeled data is automatically verified to ensure the accuracy and consistency of the labeling. After the sample defect image data is labeled, a red date image dataset with a clear boundary box is formed; It specifically uses a 3D convolutional neural network (3D CNN) model to process 3D depth image data, combined with a 2D image feature detection algorithm to process RGB images. The training method is as follows: The preprocessed RGB image and 3D depth image are combined as training samples. The input data includes shape features (such as depressions and protrusions), color features (such as lesion areas), texture features (such as surface cracks), and depth features (detecting the 3D shape of depressions and protrusions, etc.).

[0043] Object detection: Using supervised learning, 3D CNN is used to detect surface anomalies of red dates, especially deep features (such as bumps or surface morphology abnormalities), and 2D convolutional neural network (2D CNN) is used to detect color and texture anomalies. For object detection tasks, the YOLO (You Only Look Once) algorithm is used to quickly locate defects in red dates.

[0044] Model training: Use a large number of labeled red dates data sets to conduct multiple rounds of training, continuously optimize model parameters, and improve the accuracy of the model in extracting defect targets.

[0045] Verification model: Use unseen red dates samples for testing to evaluate the accuracy, recall, and F1 score of the target extraction unit to ensure that it can effectively extract the defective areas of the red dates.

[0046] The trained and verified model is used as the target extraction unit. By pre-labeling the sample defect images, the red date image dataset is obtained, and the 3D convolutional neural network (3D CNN) model is used to process the 3D depth image data. The 2D image feature detection algorithm is combined to process the RGB image and supervise the training to obtain the target extraction unit, which can realize the efficient and accurate identification of red date defects, thereby improving the efficiency and accuracy of subsequent picking.

[0047] Based on the sample defect image, a defect judgment unit is trained, including: based on the red date defect feature data, a plurality of defect judgment features are obtained; according to the plurality of defect judgment features, defects are annotated on the red date image data set to obtain a sample calibration image set, wherein the sample calibration image set includes positive samples and negative samples; based on the sample calibration image set, a convolutional neural network is used for training to obtain the defect judgment unit.

[0048] Specifically, firstly, multiple defect judgment features related to defects are extracted from the red date defect feature data through image processing technology and feature extraction algorithm, such as color, size, shape, texture and other features. The specific extraction method is as follows: Color features: The color features of an image are extracted through a color histogram or color space transformation (such as RGB to HSV) to detect color abnormal areas, such as diseased spots or moldy parts.

[0049] Texture features: Gray-level co-occurrence matrix (GLCM) or local binary pattern (LBP) is used to extract the texture features of the red dates surface for detecting surface defects such as cracks and dents.

[0050] Depth features: Use the 3D depth image data captured by the TOF camera to extract the morphological features of red dates such as bulges, depressions, cracks, and deformations.

[0051] Then, according to the extracted defect judgment features, automatic pre-labeling is used to label the red date image dataset. During the labeling process, the images are divided into positive samples containing defects and negative samples without defects. The labeled images are then used as the sample calibration image set. The positive samples contain defects of various types and degrees, while the negative samples are red date images without defects or with inconspicuous defects. The specific operation method is as follows: Positive and negative sample division: According to the extracted defect features, the red dates in the image are divided into positive samples and negative samples. Positive samples refer to red dates with obvious defects, such as abnormal color, surface cracks, dents, etc.; negative samples refer to red dates without obvious surface defects.

[0052] Labeling rules: Automatically label the defect type and severity of red dates through preset rules. Set classification thresholds based on different features: Color characteristics: Set the color deviation value, and those exceeding a certain range will be considered as diseased spots.

[0053] Shape features: Determine whether there are shape abnormalities based on surface curvature and morphological changes.

[0054] Depth feature: Determine whether the surface has convexities or concavities based on the concave and convex conditions.

[0055] Generate a sample calibration image set: Positive sample set: contains red date defects of different types and severity, ensuring that samples with a variety of appearance defects are covered.

[0056] Negative sample set: includes red dates without defects or with subtle defects as negative samples.

[0057] Data enhancement: Perform enhancement operations such as rotation, scaling, and translation on sample images to expand the sample set and improve the generalization ability of the model.

[0058] The method of obtaining the defect determination unit by training a convolutional neural network based on the sample calibration image set is as follows: Convolutional Neural Network (CNN) model training: Multi-feature fusion: The preprocessed RGB image and 3D depth image are fused as input data. The fused features include the color, shape, texture and depth information of the red dates, enabling comprehensive defect analysis.

[0059] Model construction: CNN structure: Design a convolutional neural network with multiple convolutional layers and pooling layers to extract features from the input image. Use convolution kernels to capture features such as color, shape, texture, and depth.

[0060] Feature fusion layer: Add a feature fusion layer to specifically handle the fusion of RGB and depth image information. Perform deep learning on the fused features through the convolution layer.

[0061] Training process: Loss function: Use the cross entropy loss function to calculate the gap between the model prediction value and the true label and optimize the model.

[0062] Optimization algorithm: Use Adam optimizer to update model parameters, adjust hyperparameters such as learning rate and weight decay, and improve the convergence speed of the model.

[0063] Training steps: Use positive and negative sample image sets as training data sets. After multiple rounds of training, the model continuously learns the defect features of red dates and improves the judgment accuracy of the model.

[0064] Defect classification: Classification layer output: The last layer of the CNN model is a fully connected layer, which outputs the classification results of multiple categories of defects, including defect types (such as color abnormalities, cracks, and dents) and defect levels (mild, moderate, and severe).

[0065] Determination result: Based on the CNN classification result, output whether the red dates have defects and the specific type and severity of the defects.

[0066] Model Validation and Evaluation: Test set division: The sample calibration image set is divided into a training set and a test set. The test set is used to evaluate model performance, and the training set is used for model training.

[0067] Model evaluation metrics: Accuracy: The consistency rate between the defect category predicted by the model and the actual labeled category.

[0068] Recall rate: The proportion of defective red dates detected by the model, ensuring a low missed detection rate for defective red dates.

[0069] F1 score: It combines the accuracy and recall of the model as a comprehensive indicator of model performance.

[0070] Model optimization: If the evaluation index of the model does not meet the preset requirements, continue to adjust the network structure, optimize the algorithm or hyperparameters until the model accuracy reaches the expected target, and use the model that reaches the expected target as the defect judgment unit. The defect judgment unit can further determine the type and score of the defect based on the location information of the red dates and their defects provided by the target extraction unit, achieving efficient and accurate identification of red date defects, and providing data support for the efficiency and accuracy of subsequent picking.

[0071] Further, according to the defect recognition module, material defect recognition is performed on the transmitted material image to obtain a defect recognition result, including: inputting the transmitted material image into the target extraction unit of the defect recognition module to obtain multiple defect recognition targets; the defect judgment unit receives the multiple defect recognition targets and performs defect recognition judgment to obtain a defect recognition result, which includes the location of defective red dates, defect type and defect level.

[0072] Specifically, The target extraction unit uses a pre-trained target detection algorithm to process the input image, identify multiple red dates and their defects in the transmitted image, and output multiple defect recognition targets, and automatically generate bounding boxes for each red date and its defect area based on the defect recognition targets; Next, the defect judgment unit receives multiple defect recognition targets marked by bounding boxes from the target extraction unit and uses a pre-trained convolutional neural network (CNN) to make detailed judgments on these areas. Through multi-layer convolution operations, it extracts color, shape, texture, and concave-convex information from RGB and depth images, and performs multi-level feature fusion and judgment. The following are further judgment steps: Position identification: Determine the exact defect location in the red dates to facilitate subsequent operation of the robotic arm or sorting equipment.

[0073] Defect type determination: classify defects such as cracks, black spots, insect damage, discoloration, etc.

[0074] Defect grade determination: According to the severity of the defect, it is classified into slight, medium and severe grades. For example, a shallow crack is determined as a slight crack, and a deep crack is determined as a severe crack.

[0075] Obtain defect identification results: The defect identification results not only record the type and grade of the red dates defects, but also identify the location of each defective red date in detail. The following is the specific output content: Position coordinates: The position of the red dates in the bounding box for subsequent mechanical sorting.

[0076] Defect type: Output the specific defect type of each red date, such as cracks, insect damage, discoloration, etc.

[0077] Defect level: Classified according to the severity of the defect, such as minor, moderate, and severe.

[0078] Multi-target parallel processing: To improve efficiency, the defect recognition module can use parallel computing to process targets in multiple bounding boxes simultaneously, which can further speed up the detection speed.

[0079] The defect recognition results record in detail the location, defect type and defect grade of each defective red date. The model enables fast and accurate identification of red date defects, reduces labor costs, and ensures the intelligence and efficiency of the red date defect recognition process, greatly improving the subsequent sorting efficiency and accuracy.

[0080] Furthermore, the information of red dates to be screened out is obtained through the defect identification result, and the defective red dates are picked out through the sorting control module, including: obtaining the position, defect type and defect grade of the defective red dates through the defect identification result, traversing the screening standards, and obtaining the information of red dates to be screened out; generating a sorting instruction based on the information of red dates to be screened out; and activating the sorting control module according to the sorting instruction to pick the defective red dates.

[0081] Specifically, the red date defect screening standard is pre-set, and the red date defect screening standard includes which types of defects and defect levels need to be screened out. For example, the standard may stipulate that all red dates with severe cracks or large black spots must be eliminated. Through the defect recognition results output by the defect recognition model, the specific defect location, defect type and defect level of each defective red date are obtained, and the preset screening standards are traversed. According to the screening standards, the information of the identified defective red dates is compared. According to the comparison results, it is determined which red dates meet the screening conditions and need to be picked out, that is, the information of the red dates to be screened out. The information of the red dates to be screened includes the location of the red dates to be eliminated and related defects and other information. Next, based on the information of the red dates to be screened out, a specific sorting instruction is generated, and the sorting instruction is passed to the sorting control module. After receiving the instruction, the sorting control module activates the corresponding sorting mechanism to pick the defective red dates. Optionally, the sorting control module includes actuators, such as pneumatic grippers, robotic arms, etc. When the sorting control module receives a sorting instruction, the actuator picks out defective red dates from the conveyor belt and puts them into a designated collection container. This not only realizes efficient and accurate intelligent screening and picking of red dates, but also greatly reduces manual intervention and reduces labor costs.

[0082] Through the technical solutions of the above embodiments, the intelligent picking method of defective red dates provided by the present application, based on the combination of visual detection and machine learning, not only improves the efficiency of sorting, but also greatly improves the accuracy of sorting, and realizes efficient and accurate intelligent screening and picking of red dates. At the same time, this automated system reduces manual intervention and labor costs, and helps to improve the technical effect of the overall automation level of the production line.

[0083] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0084] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. An intelligent picking method for defective red dates, characterized in that: The method comprises: Based on the feeding transmission module, the target material is evenly laid to obtain the initial transmission material; The visual inspection module is used to collect real-time images of the initially transmitted material to obtain images of the transmitted material, wherein the visual inspection module is arranged directly above the feeding transmission module; Obtaining red date defect feature data, and training a defect recognition module based on the red date defect feature data; According to the defect recognition module, material defects are recognized for the transmitted material image to obtain a defect recognition result; The information of red dates to be screened out is obtained through the defect identification results, and the defective red dates are sorted through the sorting control module.

2. The method according to claim 1, based on the feeding transmission module, evenly laying the target material to obtain the initial transmission material, comprises: The feeding transmission module is composed of a transmission device and a vibration device; Obtain basic information of target materials; Based on the basic information of the target material, configure the material conveying rate and vibration frequency; According to the material conveying rate and vibration frequency, the target material is evenly laid to obtain the initial conveying material.

3. The method according to claim 1, characterized in that The initial transmission material is imaged in real time by a visual inspection module to obtain an image of the transmission material, including: The visual inspection module includes a plurality of image sensing devices, which are evenly arranged directly above the feeding transmission module, and the plurality of image sensing devices are provided with position codes; Based on the multiple image sensing devices, real-time image acquisition is performed on the initial transmission material to obtain a transmission material image, wherein the transmission material image includes multiple sensor image sets, and each sensor image set is provided with a position mark.

4. The method according to claim 1, characterized in that The image sensing device also includes a TOF camera; The 3D depth image of red dates is collected based on the TOF camera to capture the depth information of the red dates.

5. The method according to claim 1, characterized in that Obtain the defect feature data of red dates, and train the defect recognition module based on the defect feature data of red dates, including: Based on big data, crawl the defect feature data of red dates; Based on the red date defect feature data, image feature data is extracted for preprocessing to obtain a sample defect image; Based on the sample defect image, a target extraction unit and a defect determination unit are trained, and the target extraction unit and the defect determination unit constitute the defect recognition module.

6. The method according to claim 5, characterized in that Based on the sample defect image, a training target extraction unit is provided, comprising: Pre-labeling the sample defect images to obtain a red date image dataset; Based on the red date image data set, supervised training is performed in combination with a target detection algorithm to obtain the target extraction unit.

7. The method according to claim 6, characterized in that Based on the sample defect image, training a defect determination unit includes: Based on the red date defect feature data, a plurality of defect determination features are obtained; According to the multiple defect determination features, defect annotation is performed on the red date image data set to obtain a sample calibration image set, wherein the sample calibration image set includes positive samples and negative samples; Based on the sample calibration image set, a convolutional neural network is used for training to obtain the defect determination unit.

8. The method according to claim 5, characterized in that According to the defect recognition module, material defects are recognized for the transmitted material image to obtain a defect recognition result, including: Inputting the transmission material image into the target extraction unit of the defect recognition module to obtain a plurality of defect recognition targets; The defect determination unit receives the multiple defect recognition targets, performs defect recognition determination, and obtains a defect recognition result, which includes the defective red date location, defect type, and defect level.

9. The method according to claim 8, characterized in that The information of the red dates to be screened out is obtained through the defect identification result, and the defective red dates are sorted through the sorting control module, including: Through the defect recognition results, the position, type and grade of defective red dates are obtained, and the screening standards are traversed to obtain the information of red dates to be screened out; Based on the information of the red dates to be screened out, generating a sorting instruction; According to the sorting instructions, the sorting control module is activated to pick out defective red dates.

Images