Sorting method and system executed by comb sorting robot system

Abstract

The invention relates to the technical field of crossing of artificial intelligence and intelligent equipment, discloses a sorting method and system executed by a comb sorting robot system, and aims to solve the problems of low manual detection efficiency, different standards and high cost in existing comb production. The system comprises a multi-view synchronous imaging module, an image preprocessing and feature extraction module, a comb defect and model joint recognition module, a sorting decision generation module and a six-degree-of-freedom mechanical arm execution module. Defect and model collaborative identification is achieved through multi-view high-precision imaging and a double-branch deep learning network, and full-automatic high-precision sorting is completed in combination with dynamic path planning and visual servo control. According to the method and the device, the sorting throughput of not less than 120 pieces per minute can be realized, the defect detection rate is higher than 99.5%, the model classification accuracy is higher than 99.8%, and the sorting efficiency, the consistency and the intelligent level are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and intelligent equipment, and specifically relates to a sorting method and system executed by a comb sorting robot system. Background Technology

[0002] As the consumer goods manufacturing industry accelerates its transformation towards intelligent and automated production, combs, as typical small plastic or wooden products, have long relied on manual operation for the inspection and sorting of product appearance quality, such as tooth integrity, surface defects, shape regularity, and color consistency. Traditional manual visual inspection methods are limited by human visual discrimination and subjective judgment standards, resulting in low inspection efficiency, difficulty in matching the pace of high-speed production lines, and susceptibility to operator experience, fluctuations in attention, and fatigue, leading to frequent missed inspections and misjudgments, making it difficult to guarantee product quality consistency. At the same time, continuously rising labor costs and increasingly stringent quality control and traceability requirements make the traditional model increasingly unsustainable in large-scale production.

[0003] Machine vision-based automated inspection technology offers a potential solution to these problems. This technology, through image acquisition and pattern recognition algorithms, can achieve objective quantitative analysis of product appearance features. However, existing automated solutions often employ single-view imaging or simple photoelectric sensors, which can only identify regular geometric defects or rough color differences. They struggle to handle the complex structures of combs, the diverse types of defects (such as tiny broken teeth, burrs, internal cracks, and color differences), and the numerous models available. Especially in the absence of multi-view complementary information and a highly robust recognition model, the system is extremely sensitive to interference factors such as changes in lighting, posture shifts, and material reflections, resulting in severely insufficient detection accuracy and generalization ability.

[0004] Existing technologies generally suffer from insufficient image information acquisition, limited feature extraction dimensions, and rigid defect recognition logic, making it impossible to simultaneously achieve high-precision quality assessment and multi-category intelligent sorting. Especially when facing challenges such as densely spaced comb teeth, large variations in surface curvature, and transparent or semi-transparent materials, traditional image processing methods struggle to effectively segment targets, extract key features, and establish reliable classification criteria. Therefore, there is an urgent need for an integrated system that deeply integrates multi-view imaging, advanced image recognition algorithms, and precise actuators to achieve efficient, stable, and intelligent detection and sorting of the full-dimensional appearance quality of comb products. Summary of the Invention

[0005] This invention provides a sorting method and system for a comb sorting robot system, aiming to solve the technical problems of low efficiency, inconsistent standards, and high costs caused by relying on manual appearance inspection and sorting in the existing comb production process. The system achieves fully automated, high-accuracy, and high-throughput online detection and sorting of combs based on multiple dimensions such as tooth integrity, surface defects, shape regularity, and color consistency by constructing a high-precision multi-view imaging device, a deep learning-based defect and model joint recognition model, and a mechanical actuator with dynamic path planning capabilities.

[0006] According to one aspect of the present invention, an intelligent comb sorting robot system integrating image recognition function is provided, comprising: a multi-view synchronous imaging module, an image preprocessing and feature extraction module, a comb defect and model joint recognition module, a sorting decision generation module, and a six-degree-of-freedom robotic arm execution module.

[0007] The multi-view synchronous imaging module is positioned above and on both sides of the comb conveyor line. It is used to simultaneously acquire high-resolution color images of the top, front, and two sides of the comb as it passes through the detection area at a constant speed with the conveyor belt. The module consists of four industrial-grade global shutter cameras. The optical axes of each camera are perpendicular to the top surface of the comb, parallel to the center line of the front of the comb, and at a 45-degree angle to the two sides of the comb, respectively. Each camera is equipped with an independent ring LED light source with a color temperature of 5,500 Kelvin and an illuminance uniformity of no less than 95%. All cameras achieve microsecond-level synchronous exposure through hardware trigger signals, ensuring that images of the same comb from different perspectives strictly correspond to the same physical moment.

[0008] The image preprocessing and feature extraction module receives raw image data from the multi-view synchronous imaging module. First, it performs lens distortion correction on each image, with correction parameters pre-calibrated using the Zhang Zhengyou calibration method. Then, it performs illumination normalization processing, employing an adaptive histogram equalization algorithm to eliminate the influence of ambient light fluctuations. Next, it performs background segmentation on the corrected image, using a combination of color space thresholding and morphological operations to accurately extract the foreground region of the comb. Finally, it performs sub-pixel-level edge detection on the foreground region, generating a closed polygon representation of the comb outline, and calculating its geometric moments, Fourier descriptors, and local binary pattern texture features to form a preliminary structured feature vector.

[0009] The comb defect and model joint recognition module adopts a dual-branch deep convolutional neural network architecture. Its input is a multi-view image and its structured feature vector after image preprocessing and feature extraction. The first branch is a defect detection sub-network, which is based on an improved U-shaped encoder-decoder structure. In the encoding stage, it fuses multi-scale feature maps, and in the decoding stage, it introduces an attention gating mechanism. The output is a pixel-level defect probability map with the same size as the input image. Defect types include missing teeth, broken tooth tips, surface scratches, pits, and color spots. The second branch is a model and color classification sub-network, which adopts a residual network backbone. Its end is connected to two parallel fully connected layers, which output the probability distribution of comb model category and the probability distribution of main color category, respectively. The two sub-networks interact with each other in the middle layer through a feature cross-fusion unit. This unit uses channel attention weighting to inject the spatial saliency information of the defect feature map into the model classification feature, and at the same time guides the model prior knowledge to the defect localization process, thereby realizing the collaborative optimization of defect recognition and model classification.

[0010] The sorting decision generation module receives the output from the comb defect and model joint identification module. First, it sets a threshold based on the defect probability map to determine whether the comb has any unqualified defects. If so, it is marked as a defective product. If not, it selects the model with the highest confidence as the current comb's model identifier based on the model category probability distribution, and determines its color grouping based on the main color category probability distribution. Then, it combines a preset sorting rule library, which defines the sorting exit numbers corresponding to different model and color combinations, to generate a sorting instruction that includes the target exit number, sorting priority, and gripping posture suggestion. The sorting instruction is output in the form of a structured data packet, which includes the comb's unique serial number, model code, color code, defect status flag, target exit number, and expected gripping posture parameters.

[0011] The six-degree-of-freedom robotic arm execution module includes a six-axis articulated industrial robot, an end effector, and a control system. The end effector is an adaptive pneumatic gripper with a flexible silicone pad covering its gripping surface and an integrated pressure sensor that can provide real-time feedback on the gripping force. The control system receives sorting instructions from the sorting decision generation module, parses the target exit number and desired gripping pose parameters, combines them with the real-time speed information of the conveyor belt, calculates the target angles of each joint of the robotic arm using an inverse kinematics solver, and generates a smooth joint spatial motion trajectory using a time-optimal trajectory planning algorithm. Before executing the gripping action, the system performs secondary precise positioning of the comb's current position using a high-speed visual servo to compensate for conveyor belt vibration and positioning errors, ensuring gripping accuracy within ±0.2 mm. After gripping, the robotic arm transports the comb along the planned trajectory to the designated sorting exit and releases it into the corresponding collection container.

[0012] As one embodiment of the present invention, the industrial-grade global shutter camera in the multi-view synchronous imaging module has a resolution of 2,448 x 2,048 pixels, a frame rate of not less than 60 frames per second, a lens focal length of 16 mm, and a working distance of 350 mm.

[0013] In one embodiment of the present invention, the dual-branch deep convolutional neural network in the comb defect and model joint recognition module adopts a multi-task loss function during the training phase. This loss function is composed of a weighted sum of the Dice loss for defect segmentation, the cross-entropy loss for model classification, and the cross-entropy loss for color classification, with weight coefficients of 0.6, 0.3, and 0.1, respectively.

[0014] In one embodiment of the present invention, the sorting rule base is stored in non-volatile memory in the form of key-value pairs, where the key is a combination string of model code and color code, and the value is the target exit number; the rule base supports dynamic updates through a human-computer interaction interface, and the update operation requires dual authentication.

[0015] As one embodiment of the present invention, the control system of the six-degree-of-freedom robotic arm execution module has a built-in safety monitoring unit. This unit monitors the joint torque of the robotic arm, the clamping force of the end effector, and the status of the safety light curtain in the surrounding area in real time. Once an abnormality is detected, an emergency stop procedure is immediately triggered to brake the robotic arm to its current position.

[0016] As one embodiment of the present invention, the system also includes a data management and traceability module, which records the original image, recognition result, sorting instructions and final destination of each comb to form a complete quality traceability file, and supports querying and statistical analysis by time, batch, model or defect type.

[0017] Compared with the prior art, the advantages of the present invention are as follows: This invention completely replaces manual visual inspection and sorting processes by constructing an integrated system of multi-view synchronous imaging, a deep learning-driven joint recognition model, and a high-precision mechanical actuator. The multi-view imaging design overcomes the problem of missed detections caused by comb tooth occlusion and surface reflection under a single viewpoint; the dual-branch neural network, through feature cross-fusion, achieves mutual enhancement of defect detection and model classification, significantly improving the accuracy and robustness of recognition under complex appearance features; dynamic grasping control based on real-time visual servoing solves the problem of workpiece positioning drift on high-speed conveyor lines, ensuring the reliability and cycle stability of sorting actions. The entire system can achieve a throughput of no less than 120 combs per minute, with a defect detection rate of over 99.5% and a model classification accuracy of over 99.8%, effectively solving the core pain points of low efficiency, inconsistent standards, and high cost of manual inspection, and providing a complete data foundation for product quality traceability and process optimization. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical architecture of the system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the comb defect and model joint identification module in this invention.

[0019] Figure reference numerals: 1. Multi-view synchronous imaging module; 2. Image preprocessing and feature extraction module; 3. Comb defect and model joint recognition module; 4. Sorting decision generation module; 5. Six-degree-of-freedom robotic arm execution module. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly or indirectly attached to that other component. When a component is referred to as being "connected to" another component, it can be directly or indirectly connected to that other component.

[0022] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0024] This invention provides a sorting method and system performed by a comb sorting robot system, see [link to relevant documentation]. Figure 1 and Figure 2 Its core lies in the coordinated operation of multi-view synchronous imaging, a deep learning-driven joint recognition model, and a high-precision mechanical actuator to achieve fully automated, high-accuracy, and high-throughput online detection and sorting of comb teeth integrity, surface defects, shape regularity, and color consistency. The following will describe in detail the specific implementation of this invention, combining the various functional modules of the system and their internal processing flow.

[0025] The system includes a multi-view synchronous imaging module 1, an image preprocessing and feature extraction module 2, a comb defect and model joint recognition module 3, a sorting decision generation module 4, and a six-degree-of-freedom robotic arm execution module 5. All modules work collaboratively under unified timing control to ensure that the entire process from the comb entering the detection area to the completion of sorting is completed within milliseconds, meeting the high-speed cycle requirements of industrial production lines.

[0026] The multi-view synchronous imaging module 1 is positioned above and on both sides of the comb conveyor line. It is used to simultaneously acquire high-resolution color images of the top, front, and two sides of the comb as it passes through the detection area at a constant speed with the conveyor belt. This module consists of four industrial-grade global shutter cameras. The optical axes of each camera are perpendicular to the top surface of the comb, parallel to the center line of the front surface, and at a 45-degree angle to the two sides of the comb, respectively. Each camera is equipped with an independent ring-shaped LED light source with a color temperature of 5500 Kelvin and an illuminance uniformity of no less than 95%, ensuring high-contrast, low-noise image data can be obtained on comb surfaces of different materials and colors. All cameras achieve microsecond-level synchronous exposure via hardware trigger signals, ensuring that images of the same comb from different viewpoints strictly correspond to the same physical moment, avoiding image misalignment or blurring caused by conveyor belt movement. As one embodiment of the present invention, the industrial-grade global shutter camera has a resolution of 2,448 x 2,048 pixels, a frame rate of no less than 60 frames per second, a lens focal length of 16 mm, and a working distance of 350 mm, which can cover the complete outer dimensions of a standard comb product while ensuring depth of field.

[0027] The image preprocessing and feature extraction module 2 receives raw image data from the multi-view synchronous imaging module 1. First, it performs lens distortion correction on each image. The correction parameters are pre-calibrated using the Zhang Zhengyou calibration method. This correction process employs a bilinear interpolation algorithm to remap pixel coordinates, eliminating barrel or pincushion distortion introduced by lens optical characteristics. Next, it performs illumination normalization processing using an adaptive histogram equalization algorithm. This algorithm divides the image into several local regions, independently calculates the cumulative distribution function within each region, and performs grayscale mapping, effectively eliminating the impact of ambient light fluctuations and local reflections on image quality. Then, it performs background segmentation on the corrected image, using a combination of color space thresholding and morphological operations: first, the image is converted from the RGB color space to the HSV color space, and threshold ranges for hue, saturation, and brightness are set to initially separate the comb foreground; then, opening operations are performed on the binarized result to remove isolated noise points, followed by closing operations to fill holes inside the foreground; finally, connected component analysis is used to extract the largest area region as the comb foreground. Finally, subpixel-level edge detection is performed on the foreground region. The Canny operator is combined with nonmaximum suppression and hysteresis thresholding to generate a high-precision edge point set, which is then fitted into a closed polygon representation using the Douglas-Peucker algorithm. Based on this, its geometric moments, including zero-order moments, first-order moments, second-order central moments, and Fourier descriptors, are calculated. The top twenty low-frequency coefficients and local binary pattern texture features are sampled using an eight-neighbor circular sampling pattern with a radius of one to form a preliminary structured feature vector containing shape, contour, and texture information. This vector has a dimension of 128 and serves as an auxiliary input for the subsequent deep recognition model.

[0028] The comb defect and model joint recognition module 3 adopts a dual-branch deep convolutional neural network architecture. Its input is a four-view image processed by the image preprocessing and feature extraction module 2 and its corresponding structured feature vector. The first branch is a defect detection sub-network, which is based on an improved U-shaped encoder-decoder structure. The encoder part uses ResNet34 as the backbone network, downsampling layer by layer to extract multi-scale feature maps with downsampling factors of 2, 4, 8, and 16. At each skip connection, the feature map of the corresponding level of the encoder is concatenated with the upsampled feature map of the decoder, and an attention gating mechanism is introduced: this mechanism generates a spatial weight map of the decoder features through a small convolutional layer, and then multiplies it element-wise with the encoder features, thereby dynamically suppressing irrelevant background areas and enhancing the response of defect areas. The decoder part uses transposed convolution for upsampling, and the final output is a pixel-level defect probability map with the same size as the input image. The defect types include five categories: missing teeth, broken tooth tips, surface scratches, pits, and discoloration, each corresponding to an independent probability channel. The second branch is the model and color classification sub-network. This sub-network also uses ResNet34 as its backbone, but after the last layer of global average pooling, it connects two parallel fully connected layers: the output dimension of the first fully connected layer is equal to the preset number of model categories, for example, thirty, and it uses the Softmax activation function to generate the model category probability distribution; the output dimension of the second fully connected layer is equal to the number of dominant color categories, for example, eight, and it also uses the Softmax activation function to generate the dominant color category probability distribution. The two sub-networks interact with each other in the middle layer through a feature cross-fusion unit, which is located at the third stage output of ResNet34. Specifically, the feature map of this layer in the defect detection sub-network is compressed into a channel description vector through global average pooling, and then a single-layer perceptron generates channel attention weights, which are applied to the corresponding feature map of the model classification sub-network; at the same time, the global feature vector of the model classification sub-network is expanded into a spatial attention map through linear projection, guiding the defect detection sub-network to focus on the region related to the typical structure of the current model. This cross-fusion mechanism allows prior knowledge of the model to constrain the defect localization range. For example, if the spacing between the teeth of a comb model is fixed, areas deviating from this spacing are more likely to be defects. The spatial saliency of defects can also assist in model identification; for instance, if a break appears in a decorative pattern unique to a specific model, it strengthens the confidence in identifying that model. During the training phase, this dual-branch network employs a multi-task loss function, which is a weighted sum of the Dice loss for defect segmentation, the cross-entropy loss for model classification, and the cross-entropy loss for color classification. The weight coefficients are 0.6, 0.3, and 0.1, respectively, to balance the learning intensity of each task. The Dice loss is defined as follows:

[0029] Where pi is the defect probability predicted by the model, gi is the true defect mask, and ϵ is a smoothing term with a value of one. This loss function is more sensitive to small target defects and effectively alleviates the problem of extreme imbalance between positive and negative samples.

[0030] The sorting decision generation module 4 receives the output from the comb defect and model joint identification module 3. First, it sets a threshold of 0.5 based on the defect probability map. If the defect probability of any pixel exceeds this threshold and the area of ​​the connected region is greater than ten pixels, the comb is determined to have a defect and marked as scrap. If no defect exists, the model with the highest confidence level is selected as the current comb's model identifier based on the model category probability distribution, and its color group is determined based on the main color category probability distribution. Subsequently, it combines a preset sorting rule library, which is stored in non-volatile memory in key-value pairs. The key is a string combining the model code and color code, such as "M05-C3," and the value is the target exit number, such as "OUT_07." This rule library supports dynamic updates via a human-machine interface. Update operations require dual authentication, including the operator's employee ID password and the supervisor's authorization code, to prevent misconfiguration. Based on this, the sorting decision generation module 4 generates sorting instructions containing the target exit number, sorting priority, and suggested gripping posture. The sorting instructions are output in the form of structured data packets, including a unique comb serial number generated by combining a system timestamp and a serial number, a model code, a color code, a defect status flag, the target exit number, and the desired gripping pose parameters, including three-dimensional position coordinates and Euler angles. This data packet is transmitted in real time to the control system of the six-degree-of-freedom robotic arm execution module 5 via a high-speed industrial Ethernet.

[0031] The six-degree-of-freedom robotic arm execution module 5 includes a six-axis articulated industrial robot, an end effector, and a control system. The end effector is an adaptive pneumatic gripper with a gripping surface covered by a flexible silicone pad with a Shore A hardness of 40 degrees to accommodate comb handles of varying thicknesses and curvatures. The gripper integrates a miniature piezoresistive pressure sensor with a range of 0 to 50 Newtons and an accuracy of 0.5 Newtons, providing real-time feedback of the gripping force to the control system. The control system receives sorting instructions from the sorting decision generation module 4, analyzes the target exit number and desired gripping pose parameters, and combines this with real-time conveyor belt speed information fed back by an encoder with an accuracy of 0.1 mm / s. The system then calculates the target angles of each joint of the robotic arm using an inverse kinematics solver. The inverse kinematics solution employs a numerical iterative method, with the initial solution obtained by backward kinematics calculation. The iteration terminates when the end-effector pose error is less than 0.1 mm and 0.1 degree. Subsequently, a time-optimal trajectory planning algorithm is used to generate a smooth joint space motion trajectory. This algorithm minimizes the motion time while satisfying joint velocity, acceleration, and jerk constraints, ensuring that the grasping and placement actions are completed within the conveyor belt cycle. Before executing the grasping action, the system performs secondary precise positioning of the comb's current position using high-speed visual servoing: the control system triggers the top-side camera in the multi-view synchronous imaging module 1 to take a single snapshot, uses the extracted comb outline polygon to perform template matching with the desired grasping pose, calculates the offset between the actual position and the expected position, and compensates it to the target pose of the robotic arm in real time, ensuring that the grasping accuracy is within ±0.2 mm. After grasping, the robotic arm transports the comb along the planned trajectory to the designated sorting exit. This exit is a collection container with a buffer chute, and a photoelectric counter is installed at the bottom of the container to count the sorted quantity. Before releasing the comb, the control system instructs the pneumatic gripper to slowly depressurize, so that the gripping force linearly decreases to five Newtons before fully opening, preventing the comb from bouncing or rolling due to sudden release. The control system has a built-in safety monitoring unit that monitors the joint torque of the robotic arm in real time. It calculates the clamping force of the end effector and the status of the safety light curtain in the surrounding area by converting the motor current. Once an abnormality is detected, such as a sudden increase in torque exceeding the threshold, abnormal fluctuation in clamping force, or obstruction of the safety light curtain, an emergency stop procedure is immediately triggered to brake the robotic arm to its current position and send a fault code to the central monitoring system.

[0032] In addition, the system includes a data management and traceability module, which runs on an embedded industrial computer. This module records the original four-view images and recognition results for each comb, including defect probability maps, model and color probability distributions, sorting instructions, and final destination. Feedback from the exit photoelectric counter confirms the data, forming a complete quality traceability file. All data is timestamped and stored on a solid-state drive, supporting queries and statistical analysis by time, batch, model, or defect type. For example, it can generate a defect distribution heatmap for a specific comb model within a specific shift, or statistically analyze the pass rate trend for a specific color batch, providing data support for process optimization.

[0033] In summary, this embodiment overcomes the problem of single-view occlusion through multi-view synchronous imaging, achieves collaborative identification of defects and models through a dual-branch depth network, and ensures the accuracy and reliability of high-speed sorting through visual servoing and time-optimal trajectory planning. The entire system can achieve a throughput of no less than 120 combs per minute for detection and sorting, with a defect detection rate of over 99.5% and a model classification accuracy of over 99.8%, effectively solving the core pain points of low efficiency, inconsistent standards, and high cost of manual inspection.

[0034] The above embodiments are merely explanations of this application and are not intended to limit it. After reading this specification, those skilled in the art can make modifications to these embodiments without contributing any inventive step, but such modifications are protected by patent law as long as they fall within the scope of the claims of this application.

Claims

1. A method of sorting performed by a comb sorting robot system, characterized by, The application relates to a comb defect and type joint identification system and method. The high-resolution color images are input into an image preprocessing and feature extraction module to perform lens distortion correction, illumination normalization, background segmentation and sub-pixel level edge detection, to generate a closed polygon representation of the comb profile, and to calculate geometric moments, Fourier descriptors and local binary pattern texture features to form a structured feature vector. The preprocessed multi-view images and the corresponding structured feature vectors are input into a comb defect and type joint identification module, which adopts a double-branch deep convolutional neural network architecture, wherein a first branch is a defect detection subnetwork that outputs a pixel-level defect probability map, and defect types include tooth loss, tooth tip breakage, surface scratches, pits and color spots; a second branch is a type and color classification subnetwork that respectively outputs a comb type category probability distribution and a dominant color tone category probability distribution; the two subnetworks interact information through a feature cross fusion unit. The defect probability map, the type category probability distribution and the dominant color tone category probability distribution are input into a sorting decision generation module, which determines whether there is an unqualified defect according to a preset threshold, and if there is no defect, the type and color information are combined to query a sorting rule library to generate a sorting instruction containing a target outlet number, a sorting priority and a grasping posture suggestion. The sorting instruction is transmitted to a six-degree-of-freedom mechanical arm execution module, which analyzes the instruction and combines real-time speed information of the conveying belt to calculate joint target angles through an inverse kinematics solver, generates a motion trajectory by using a time-optimal trajectory planning algorithm, and performs secondary accurate positioning on the comb position through high-speed visual servo before executing the grasping, and after the grasping is completed, the comb is transported to the designated sorting outlet and released. The multi-view synchronous imaging module is composed of four industrial-grade global shutter cameras, the optical axes of the cameras are respectively perpendicular to the top surface of the comb, parallel to the center line of the front surface of the comb, and at a forty-five-degree angle with the two side surfaces of the comb, each camera is equipped with an independent ring LED light source, the color temperature of the light source is five thousand five hundred Kelvin, the illumination uniformity is not less than ninety-five percent, and all the cameras realize microsecond-level synchronous exposure through a hardware trigger signal.

2. The comb picking robot system of claim 1, wherein the method of picking performed by the comb picking robot system is characterized by, When the image preprocessing and feature extraction module performs lens distortion correction, correction parameters obtained by Zhang Zhengyou's calibration method are adopted, and pixel coordinates are remapped through a bilinear interpolation algorithm; adaptive histogram equalization algorithm is adopted for illumination normalization; background segmentation is realized by converting the image to HSV color space and setting a threshold range, and then the maximum area foreground region is extracted through morphological opening and closing operation and connected component analysis; sub-pixel level edge detection is realized by combining Canny operator with non-maximum suppression and hysteresis threshold processing, and the closed polygon is fitted through Douglas-Peucker algorithm.

3. The comb picking robot system performing a picking method according to claim 2, characterized in that, ​ 4. The comb picking robot system performing a picking method according to claim 3, characterized in that, In the comb defect and model joint identification module, the defect detection subnetwork is based on an improved U-shaped encoder-decoder structure, the encoder adopts a ResNet34 backbone network, and the decoder introduces an attention gate mechanism at a jump connection, and a spatial weight map is generated through a small convolutional layer to weight the encoder features; The model and color classification subnetwork also uses ResNet34 as the backbone, and after global average pooling, two parallel fully connected layers are connected to output the model and color probability distribution respectively.

5. The comb picking robot system of claim 4, wherein the method of picking performed by the comb picking robot system is further characterized by, The feature cross fusion unit is located at the output of the third stage of ResNet34, and the feature map of the defect detection subnetwork is subjected to global average pooling to generate channel attention weights acting on the corresponding feature map of the model classification subnetwork, and at the same time, the global feature vector of the model classification subnetwork is linearly projected and expanded into a spatial attention map to guide the defect detection subnetwork to focus on the area related to the typical structure of the current model.

6. The comb picking robot system of claim 5, wherein the method of picking performed by the comb picking robot system is further characterized by, The double-branch deep convolutional neural network adopts a multi-task loss function in the training stage, which is composed of the weighted sum of the Dice loss of defect segmentation, the cross-entropy loss of model classification and the cross-entropy loss of color classification, and the weight coefficients are 0.6, 0.3 and 0.1 respectively, wherein the Dice loss is defined as: where p i is the predicted defect probability, g i is the ground truth defect mask, and e is a smoothing term and takes a value of one.

7. The comb-sorting robot system of claim 6, wherein the method of sorting performed by the comb-sorting robot system is further characterized by, The sorting decision generation module sets the threshold to be 0.5 according to the defect probability map, and if the defect probability of any pixel exceeds the threshold and the connected region area is greater than 10 pixels, it is determined as waste; the sorting rule library is stored in a non-volatile memory, and a combination string of model code and color code is used as a key and a target outlet number is used as a value, and dynamic updating is supported through a human-computer interaction interface, and a double authentication is required for updating operation.

8. The comb picking robot system performing a picking method according to claim 7, characterized in that, The end effector of the six-degree-of-freedom mechanical arm execution module is an adaptive pneumatic gripper, the clamping surface is covered with a forty-degree flexible silicone rubber pad with a Shore A value of forty, and a piezoresistive pressure sensor with a measurement range of zero to fifty Newton is integrated inside; the control system triggers the top surface camera to take a single snapshot before grabbing, calculates the offset between the actual position and the expected position through template matching and compensates in real time to ensure that the grabbing accuracy is within plus or minus 0.2 mm; when releasing the comb, the gripper is slowly depressurized to five Newton and then fully opened.

9. A smart comb sorting robot system integrated with image recognition function, characterized in that, It comprises: A multi-view synchronous imaging module for synchronously collecting high-resolution color images of the top surface, front surface and two side surfaces of the comb when the comb passes through the detection area at a uniform speed with the conveying belt; An image preprocessing and feature extraction module for performing lens distortion correction, illumination normalization, background segmentation and sub-pixel level edge detection on the high-resolution color images, generating a closed polygon representation of the comb profile, and calculating geometric moments, Fourier descriptors and local binary pattern texture features to form a structured feature vector; A comb defect and model joint identification module for receiving the preprocessed multi-view images and their structured feature vectors, adopting a double-branch deep convolutional neural network architecture, wherein the first branch is a defect detection subnetwork outputting a pixel-level defect probability map, and the second branch is a model and color classification subnetwork outputting a comb model class probability distribution and a dominant color class probability distribution respectively, and the two subnetworks interact through a feature cross fusion unit; The sorting decision generation module is used for determining whether there is unqualified defect according to the defect probability map, the model category probability distribution and the main tone category probability distribution, and generating a sorting instruction including a target outlet number, a sorting priority and a grabbing posture suggestion by querying a sorting rule base in combination with the model and color information when there is no defect; The six-degree-of-freedom mechanical arm execution module is used for receiving the sorting instruction, calculating joint target angles in combination with real-time speed information of the conveying belt after analysis, generating a motion track, and performing secondary accurate positioning through high-speed visual servo before grabbing to complete comb grabbing and sorting.

10. The smart comb sorting robot system integrating image recognition functionality of claim 9, wherein, The multi-view synchronous imaging module is composed of four industrial-grade global shutter cameras, the optical axes of the cameras are respectively perpendicular to the top surface of the comb, parallel to the center line of the front surface of the comb and at an angle of 45 degrees with the two side surfaces of the comb, each camera is equipped with an independent ring LED light source, the color temperature of the light source is 5500K, the illumination uniformity is not less than 95%, and all the cameras realize microsecond-level synchronous exposure through a hardware trigger signal.

Images