A component visual recognition automatic sorting system

Abstract

The application provides a component visual recognition automatic sorting system, and relates to the technical field of data processing, and comprises the following steps: based on the sub-pixel level spatial position vector and the target sorting category of all components, a global sorting path optimization is performed by using an analog annealing heuristic algorithm to generate a component sorting sequence; based on the component sorting sequence, a vacuum adsorption gradient control instruction is generated according to the aspect ratio parameter of each component; a three-axis linkage position control is performed according to the sub-pixel level spatial position vector of the component, and the vacuum adsorption gradient control instruction is loaded on the negative pressure adsorption end to capture the target component; and the component is transferred to the corresponding classified bin according to the target sorting category. The application improves the sorting accuracy and the production efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a component visual recognition automatic sorting system. BACKGROUND

[0002] With the development of electronic components towards miniaturization and high integration, the sorting of multi-component in bulk distribution needs to meet the requirements of high speed, high precision and high reliability. Some traditional sorting systems rely on manual design features and classical algorithms, which are low in efficiency and high in labor intensity. Limited by the algorithm principle and hardware performance, the sorting performance of the traditional system in complex scenes cannot meet the industrial sorting requirements.

[0003] Some traditional sorting systems have limitations in category judgment mechanism. The physical feature template library is mostly based on single-dimensional parameter construction, and lacks the fusion index capability of multi-dimensional features, which is prone to misjudgment in the category matching of similar-shaped components. The updating and maintenance of the template library rely on manual intervention, which has poor scalability and cannot quickly adapt to new component categories. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a component visual recognition automatic sorting system, which improves the sorting accuracy and improves the production efficiency.

[0005] To solve the above technical problems, the technical scheme of the present application is as follows:

[0006] In a first aspect, a component visual recognition automatic sorting system comprises:

[0007] A processing module is configured to perform synchronous optical scanning on multi-component in bulk distribution to generate an original optical signal data set, and to perform preprocessing on the original optical signal data set to generate an optimized binary data matrix.

[0008] An analysis module is configured to extract component geometric feature based on the optimized binary data matrix through a multi-order Gaussian differential operator, and to parse surface identification information using a local texture feature encoder to construct a structured feature tensor.

[0009] A calculation module is configured to input the structured feature tensor into a pre-trained YOLOV2 deep convolution network, and to obtain a sub-pixel level spatial position vector and a bounding box tensor of the component through forward propagation calculation of convolution kernel weight parameters.

[0010] An execution module is configured to calculate the profile curvature distribution function and the aspect ratio parameter of the component based on the bounding box tensor to generate a spatial geometric feature descriptor, and to input the spatial geometric feature descriptor into a pre-constructed component physical feature template library to obtain a target sorting category through the Hamming distance minimization principle.

[0011] The sorting module is used for global sorting path optimization based on the sub-pixel level spatial position vector of all components and the target sorting category by using an analog annealing heuristic algorithm, and generates a component sorting sequence.

[0012] The control module is used for generating vacuum adsorption gradient control instructions according to the aspect ratio parameter of each component in sequence based on the component sorting sequence; performing three-axis linkage position control according to the sub-pixel level spatial position vector of the component, and controlling the negative pressure adsorption end to load the vacuum adsorption gradient control instructions to capture the target component; and moving the component to the corresponding classified bin according to the target sorting category.

[0013] Further, the multiple components in a scattered state are synchronously optically scanned to generate an original optical signal data set; the original optical signal data set is preprocessed to generate an optimized binary data matrix, including:

[0014] The components in a scattered state on the conveying belt are synchronously optically scanned by a linear array CCD sensor and a ring-shaped LED light source to obtain the gray value of each pixel point, and an original optical signal data set is generated; the original optical signal data set is subjected to median filtering processing to generate a denoised optical signal data set;

[0015] The denoised optical signal data set is subjected to Gaussian low-pass filtering to generate a low-frequency background component, and the denoised optical signal data set and the low-frequency background component are subjected to pixel-by-pixel difference operation to generate a background-suppressed optical signal data set;

[0016] The background-suppressed optical signal data set is subjected to local contrast limited histogram equalization processing to obtain an enhanced optical signal data set; the enhanced optical signal data set is subjected to adaptive threshold segmentation processing to generate an optimized binary data matrix.

[0017] Further, based on the optimized binary data matrix, the geometric topographic features of the components are extracted by a multi-order Gaussian differential operator, and the surface identification information is analyzed by using a local texture feature encoder to construct a structured feature tensor, including:

[0018] The optimized binary data matrix is subjected to second-order Gaussian differential convolution processing, a first-order Gaussian differential operator is used to perform X-axis gradient convolution operation to generate a gradient amplitude distribution matrix; and a second-order Gaussian differential operator is used to perform Laplace convolution operation to generate a curvature feature distribution matrix;

[0019] Based on the gradient amplitude distribution matrix and the curvature feature distribution matrix, spatial feature fusion processing is performed, a spatial coordinate mapping relationship is established by detecting the local maximum points of the gradient amplitude distribution matrix, extreme value features in the curvature feature distribution matrix of the corresponding coordinate region are extracted, and a geometric topographic feature matrix is generated;

[0020] Performing local texture feature coding operation on the geometric topography feature matrix, calculating the contrast statistic and entropy statistic of the gray level co-occurrence matrix in a predetermined analysis window centered on the feature point, and jointly coding the two as a texture feature description vector;

[0021] Stacking the geometric topography feature matrix and the texture feature description vector in multiple channels to construct a structured feature tensor.

[0022] Further, input the structured feature tensor into the pre-trained YOLOV2 deep convolutional network, and calculate the sub-pixel level spatial position vector and the bounding box tensor of the component through the forward propagation of the convolution kernel weight parameters, including:

[0023] Input the structured feature tensor into the pre-trained YOLOV2 deep convolutional network, and generate a primary feature representation through convolution kernel weight parameters and calculation; perform down-sampling convolution operation on the primary feature representation to generate a secondary feature representation;

[0024] Calculate the center coordinate offset and the width-height scaling factor of the bounding box through the convolution kernel weight of the regression prediction layer on the secondary feature representation; determine the reference integer coordinate position according to the center coordinate offset of the bounding box; establish a local neighborhood analysis region centered on the reference integer coordinate position; perform weight distribution calculation in the local neighborhood analysis region, and perform weighted average calculation to finally generate a sub-pixel level offset vector;

[0025] Fuse the sub-pixel level offset vector and the width-height scaling factor to generate a sub-pixel level spatial position vector of the component; fuse the center coordinate offset of the bounding box and the width-height scaling factor to generate a bounding box tensor.

[0026] Further, based on the bounding box tensor, calculate the profile curvature distribution function and the aspect ratio parameter of the component to generate a spatial geometric feature descriptor; input the spatial geometric feature descriptor into the pre-constructed component physical feature template library, and obtain the target sorting class through the Hamming distance minimization principle, including:

[0027] Based on the bounding box tensor, extract the circumscribed rectangular region of the component, and perform a contour point sequence extraction operation in the circumscribed rectangular region to obtain an ordered contour point coordinate sequence; select three-point analysis units composed of contour points and adjacent points before and after in the contour point coordinate sequence; calculate the curvature estimation value based on the unit, and generate a curvature value set by traversing all contour points;

[0028] Perform frequency distribution statistics operation based on the contour point curvature value set to construct a discrete curvature distribution function; calculate the aspect ratio parameter according to the width and height parameters in the bounding box tensor;

[0029] The discrete curvature distribution function is spliced with a length-width ratio parameter to generate a spatial geometric feature descriptor; the spatial geometric feature descriptor is input into a pre-constructed physical feature template library of the component, and a target sorting category is obtained through a Hamming distance minimization matching operation.

[0030] Further, based on the sub-pixel level spatial position vector of all components and the target sorting category, a simulated annealing heuristic algorithm is used for global sorting path optimization to generate a component sorting sequence, including:

[0031] An initial sorting path sequence is constructed based on the sub-pixel level spatial position vector of all components; a category aggregation degree weight parameter is calculated based on the target sorting category; and an energy function is defined based on the category aggregation degree weight parameter and the sub-pixel level spatial position vector.

[0032] The energy value of the path state is calculated based on the initial sorting path sequence and the energy function, and a simulated annealing initial temperature parameter is set; a neighborhood solution generation operation is performed based on the path state, a new sorting sequence is generated by randomly exchanging the positions of two components in the sorting sequence; and a quantitative evaluation value is generated by calculating the new sorting sequence based on the energy function.

[0033] A path update decision result is generated based on the difference between the temperature parameter, the path state energy value and the quantitative evaluation value; the current path state is updated according to the path update decision result, and the neighborhood solution generation and state update operations are repeatedly executed until the temperature balance condition is met, and the temperature decay operation is performed, and the optimization cycle is repeated until the convergence condition is met, and finally the component sorting sequence is generated.

[0034] Further, based on the component sorting sequence, vacuum adsorption gradient control instructions are generated in sequence according to the length-width ratio parameter of each component; three-axis linkage position control is performed according to the sub-pixel level spatial position vector of the component, and the vacuum adsorption gradient control instructions are loaded into the negative pressure adsorption end to capture the target component, and the component is transferred to the corresponding classification bin according to the target sorting category, including:

[0035] The length-width ratio parameter is obtained according to the component sorting sequence; the vacuum adsorption gradient control instructions are generated based on the length-width ratio parameter; and the three-axis linkage displacement parameters are calculated based on the sub-pixel level spatial position vector to generate three-axis linkage displacement parameters.

[0036] The three-axis linkage position control is performed according to the three-axis linkage displacement parameters, so that the negative pressure adsorption end is accurately positioned above the target component; and the vacuum adsorption gradient control instructions are loaded into the negative pressure adsorption end to capture the target component in the accurate positioning state.

[0037] The spatial coordinates of the corresponding classification bin are determined based on the target sorting category; and the three-axis linkage position control is performed based on the spatial coordinates of the classification bin to transfer the component to the target position.

[0038] Further, the structured feature tensor is input into a pre-trained YOLOV2 deep convolutional network, and a primary feature representation is generated by convolution kernel weight parameters and calculation; a subsampling convolution operation is performed on the primary feature representation to generate a secondary feature representation, including:

[0039] The structured feature tensor is input into the convolution kernel weight parameters of the pre-trained convolution layer to perform a spatial feature extraction operation to generate an initial feature representation; a feature value sign determination process is performed on each feature value in the initial feature representation; a feature value resetting operation is performed according to the sign determination result, and if the feature value is negative, it is reset to zero, and if the feature value is non-negative, it is maintained as the original value; an activated feature representation is generated based on the reset feature value; a spatial resolution reduction operation is performed on the activated feature representation to generate a down-sampling feature representation;

[0040] The down-sampling feature representation is input into the convolution kernel weight parameters of the secondary convolution layer to perform a feature enhancement operation to generate a primary feature representation; a convolution operation of reducing the size of a feature map is performed on the primary feature representation to generate a secondary feature representation.

[0041] The second aspect is a computing device, including:

[0042] One or more processors;

[0043] A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the system.

[0044] The third aspect is a computer readable storage medium, the computer readable storage medium stores a program, the program is executed by a processor to implement the system.

[0045] The above-mentioned scheme of the present application at least includes the following beneficial effects:

[0046] By adopting the synchronous optical scanning technology, the morphological features of the randomly distributed components can be completely captured, and the feature loss problem caused by stacking shielding and surface reflection can be effectively reduced. The optimized binary data matrix is preprocessed by material self-adaptation, which reduces the noise interference and artifact influence and reduces the edge fracture phenomenon, providing high-quality data support for subsequent feature extraction.

[0047] By using the spatial geometric feature descriptor and the Hamming distance matching mechanism, the category judgment logic is optimized, the classification accuracy of similar appearance components is improved, the sorting module adopts the simulated annealing heuristic algorithm, the local optimization problem of traditional path planning is improved, the rationality of the global path is enhanced, and the sorting process continuity is improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1A schematic diagram of a component visual recognition automatic sorting system is provided by an embodiment of the present application.

[0049] Figure 2 A target sorting category flow schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0050] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0051] As Figure 1 shown, an embodiment of the present application proposes a component visual recognition automatic sorting system, comprising:

[0052] A processing module is configured to perform synchronous optical scanning on the plurality of components in a bulk distribution to generate a raw optical signal dataset, and to perform preprocessing on the raw optical signal dataset to generate an optimized binary data matrix.

[0053] An analysis module is configured to extract component geometric topographic features by a multi-order Gaussian differential operator based on the optimized binary data matrix, and to parse surface identification information using a local texture feature encoder to construct a structured feature tensor.

[0054] A calculation module is configured to input the structured feature tensor into a pre-trained YOLOV2 deep convolutional network, and to obtain a sub-pixel level spatial position vector and a bounding box tensor of the component via forward propagation calculation of convolution kernel weight parameters.

[0055] An execution module is configured to calculate a profile curvature distribution function and an aspect ratio parameter of the component based on the bounding box tensor to generate a spatial geometric feature descriptor, and to input the spatial geometric feature descriptor into a pre-constructed component physical feature template library to obtain a target sorting category based on a Hamming distance minimization principle.

[0056] A sorting module is configured to perform global sorting path optimization using a simulated annealing heuristic algorithm based on the sub-pixel level spatial position vector of all components and the target sorting category to generate a component sorting sequence.

[0057] A control module is configured to generate a vacuum suction gradient control instruction according to the aspect ratio parameter of each component based on the component sorting sequence, to perform three-axis linkage position control according to the sub-pixel level spatial position vector of the component, and to control a negative pressure suction end to load the vacuum suction gradient control instruction to capture a target component, and to move the component to a corresponding classification bin according to the target sorting category.

[0058] In the embodiment of the present application, by adopting synchronous optical scanning technology, the morphological characteristics of the randomly distributed components can be completely captured, and the feature loss problem caused by stacking shielding and surface reflection can be effectively reduced. The optimized binary data matrix is preprocessed by material self-adaptive preprocessing, which reduces the influence of noise interference and artifacts and reduces the edge fracture phenomenon, providing high-quality data support for subsequent feature extraction.

[0059] By using spatial geometric feature descriptor and Hamming distance matching mechanism, the classification accuracy of similar shaped components is improved by optimizing the category judgment logic. The sorting module uses simulated annealing heuristic algorithm to improve the local optimization problem of traditional path planning and enhance the rationality of the global path to improve the continuity of the sorting process.

[0060] In a preferred embodiment of the present application, the randomly distributed multi-component devices are synchronously optically scanned to generate an original optical signal data set; the original optical signal data set is preprocessed to generate an optimized binary data matrix, comprising:

[0061] The line array CCD sensor and the ring LED light source are used to synchronously optically scan the randomly distributed components on the conveying belt to obtain the gray value of each pixel point and generate an original optical signal data set; the original optical signal data set is subjected to median filter processing to generate a denoising optical signal data set;

[0062] The denoising optical signal data set is subjected to Gaussian low-pass filtering to generate a low-frequency background component, and the denoising optical signal data set and the low-frequency background component are subjected to pixel-by-pixel difference operation to generate a background-suppressed optical signal data set;

[0063] The background-suppressed optical signal data set is subjected to local contrast limited histogram equalization processing to obtain an enhanced optical signal data set; the enhanced optical signal data set is subjected to adaptive threshold segmentation processing to generate an optimized binary data matrix.

[0064] In the embodiment of the present application, the line array CCD sensor and the ring LED light source are combined to realize synchronous optical scanning: the line array CCD sensor selects an industrial grade device with a resolution of 1024-4096 pixels, and the scanning speed matches the running speed of the conveying belt (usually 500-2000 lines / s), ensuring that each component in the random distribution is covered by at least 3-5 scanning lines; the ring LED light source uses a composite spectrum with a wavelength of 450-650 nm, and the brightness can be adaptively changed in the range of 500-3000 cd / m2 through feedback adjustment (dynamic adjustment for different material components such as metal, ceramic and plastic). During the scanning process, the CCD sensor acquires the optical signal of the components on the conveying belt line by line, converts the light intensity information into gray values in the range of 0-255, and finally generates an original optical signal data set containing spatial coordinates and gray information.

[0065] A square filter window of 3x3 to 7x7 is selected (dynamically selected according to noise intensity, a larger window is selected when there is more noise), the gray values of all pixels in the window are sorted, the middle value is taken to replace the original gray value of the center pixel of the window, and the process is repeated until the entire data set is traversed to generate a denoised light signal data set; the median filter can effectively remove isolated noise points while preserving the sharp features of the edges of the components.

[0066] Gaussian low-pass filtering generates a low-frequency background component: Gaussian low-pass filtering is used to extract the image background, the Gaussian kernel window size is set to 15x15 to 31x31, and the standard deviation σ is set to 0.5-2.0 (adjust according to the background smoothness, the smaller the σ value, the more uniform the background), the denoised light signal data set is smoothed by Gaussian weighting to obtain a low-frequency background component reflecting overall illumination changes.

[0067] Pixel-by-pixel difference operation: Perform pixel-by-pixel gray value subtraction between the denoised light signal data set and the low-frequency background component, i.e., the gray value of each pixel is equal to the gray value of the denoised data minus the gray value of the low-frequency background component at the corresponding position, to generate a background-suppressed light signal data set, thereby eliminating the interference of uniform illumination changes or large-area background on the features of the components.

[0068] Perform CLAHE processing on the background-suppressed light signal data set: Divide the image into non-overlapping rectangular blocks of 8x8 to 32x32 (the block size is selected according to the size of the components, smaller components use smaller blocks), set a contrast limit threshold for the histogram of each block (usually 2-10 to avoid local over-enhancement), and redistribute the histogram frequency that exceeds the threshold to stretch the local gray range, and finally obtain an enhanced light signal data set to improve the clarity of the edges and surface details of the components.

[0069] Adaptive threshold segmentation technique is used to generate a binary matrix: Select a local window of 3x3 to 15x15, and use the local mean or Gaussian weighted mean of the pixels in the window as the threshold (select the mean type dynamically for different materials, and use the Gaussian weighted mean for metal components), compare the gray value of each pixel in the enhanced light signal data set with the threshold value of the corresponding window, and mark the pixels with a gray value higher than the threshold as foreground (assigned value 1) and the pixels with a gray value lower than the threshold as background (assigned value 0) to generate an optimized binary data matrix; through adaptive threshold adjustment, the foreground and background of the components can be accurately separated in uneven lighting scenes, reducing edge breaks or artifacts.

[0070] The combination of median filtering and Gaussian low-pass filtering reduces the interference of salt and pepper noise and illumination changes on the signal, providing a clean data basis for subsequent feature extraction. Through dynamic window selection and adaptive parameter adjustment, edge blurring or detail loss is avoided in the denoising and background suppression process, ensuring the integrity of key geometric features such as component outlines, corners, etc. The combination of annular light source spectrum optimization and adaptive threshold segmentation technology enables the system to stably process different material components such as metal reflective, ceramic matte, and plastic translucent, improving data quality in multi-category mixed scenes.

[0071] In a preferred embodiment of the present application, based on the optimized binary data matrix, the component geometric feature is extracted by a multi-order Gaussian differential operator, and the surface identification information is analyzed by a local texture feature encoder to construct a structured feature tensor, including:

[0072] A second-order Gaussian differential convolution is applied to the optimized binary data matrix, a first-order Gaussian differential operator is used to perform X-axis gradient convolution operation to generate a gradient amplitude distribution matrix; and a second-order Gaussian differential operator is used to perform Laplace convolution operation to generate a curvature feature distribution matrix;

[0073] Based on the gradient amplitude distribution matrix and the curvature feature distribution matrix, spatial feature fusion processing is performed, the spatial coordinate mapping relationship is established by detecting the local maximum points of the gradient amplitude distribution matrix, the extreme value features of the corresponding coordinate region in the curvature feature distribution matrix are extracted, and a geometric feature matrix is generated;

[0074] The local texture feature encoding operation is performed on the geometric feature matrix, the contrast statistics and entropy statistics of the gray level co-occurrence matrix are calculated in the predetermined analysis window centered on the feature point, and the two are jointly encoded into a texture feature description vector;

[0075] The geometric feature matrix and the texture feature description vector are stacked in multiple channels to construct a structured feature tensor.

[0076] In the embodiment of the present application, a second-order Gaussian differential convolution is applied to the optimized binary data matrix, a 5x5 to 11x11 two-dimensional Gaussian kernel is selected (the kernel is dynamically selected according to the size of the component, and a smaller kernel is used for small-size components), and the scale parameter k of the kernel function is set to 0.8-2.5 (the k value increases with the complexity of the component features); During convolution, the Gaussian kernel traverses the entire matrix in the form of a sliding window, the pixel values in each window are multiplied by the kernel elements one by one and then summed to generate a preliminary edge enhancement matrix, highlighting the outline edges and corner features of the components.

[0077] X-axis gradient convolution operation: the first-order Gaussian differential operator is used to perform X-axis gradient extraction, and a 3x3 or 5x5 horizontal direction Gaussian kernel (kernel elements are symmetrically distributed along the X-axis) is selected; during convolution, the kernel window starts from the top left corner of the matrix, slides row by row in pixel units, and the weighted sum of the pixels in the window and the kernel elements is calculated every time the window is slid, generating an X-direction gradient amplitude distribution matrix, which can clearly reflect the change intensity of the component edge in the horizontal direction.

[0078] Laplacian convolution operation:

[0079] The second-order Gaussian differential operator (Laplacian operator) is used to perform curvature feature extraction, and a 3x3 to 7x7 Laplace Gaussian kernel (LoG kernel) is selected, with the scale parameter k set to 1.0-3.0 (larger k value is selected for components with rich surface features); during convolution, the weighted sum of the second-order derivatives of the local region is calculated through point-by-point convolution operation of the kernel window and the matrix, generating a curvature feature distribution matrix, which is used to describe the bending degree and concave-convex features of the component surface.

[0080] In the gradient amplitude distribution matrix, a 3x3 window is used to traverse each pixel, and the current pixel value is compared with the values of the surrounding 8 neighboring pixels. If the current value is the maximum value in the window and exceeds a predetermined threshold (the threshold is 15%-25% of the maximum gradient value of the matrix), it is marked as an edge feature point, and a spatial coordinate mapping relationship of the feature point is established.

[0081] According to the spatial coordinates of the feature points, the corresponding region in the curvature feature distribution matrix is located, a 5x5 window is taken with the feature point as the center, the maximum value, the minimum value and the mean value of the curvature values in the window are calculated, the extreme value features are arranged in coordinate order, and a geometric topographic feature matrix is generated. The matrix dimension matches the number of feature points, and the geometric information such as the contour curvature and edge gradient of the component is completely retained.

[0082] Taking the feature points in the geometric topographic feature matrix as the center, a predetermined analysis window of 7x7 to 15x15 is taken (the window size is adjusted according to the size of the surface mark), and a gray level co-occurrence matrix (GLCM) is calculated in the window; the pixel spacing is set to 1-3 pixels, the angle is selected from 0°, 45°, 90° and 135°, and the co-occurrence frequency of different gray level pixel pairs is counted.

[0083] From the gray level co-occurrence matrix, the contrast statistic (reflecting the texture clarity) and the entropy statistic (reflecting the texture complexity) are extracted, the contrast value ranges from 0 to 2000, and the entropy value ranges from 0 to 8 (the larger the value, the more complex the texture); after normalization processing of the two statistics, they are combined and encoded as a fixed-length texture feature description vector according to the "contrast-entropy" structure, and each feature point corresponds to a vector.

[0084] The geometric topographic feature matrix and the texture feature description vector are stacked in multiple channels: the geometric topographic feature matrix is taken as a basic channel (spatial coordinate information is reserved), the texture feature description vector is mapped to a texture channel according to the corresponding coordinates, and the two are fused into a three-dimensional structured feature tensor through channel dimension expansion; the height and width of the tensor are consistent with the size of the original matrix, the number of channels includes the geometric feature channel and the texture feature channel, and the joint representation of the geometric topography and the surface identification information is realized.

[0085] The hierarchical application of the multi-order Gaussian differential operator can accurately capture multi-level geometric features from edge gradients to curvature changes, and avoid feature omission of complex topography caused by traditional single operators. The encoding mode of the gray level co-occurrence matrix combined with multi-direction statistics has anti-interference ability to the light change and slight wear of the surface identification, and improves the texture recognition stability. The structured feature tensor realizes the spatial alignment and fusion of geometric and texture features, and solves the information fragmentation problem caused by traditional feature separation processing.

[0086] In a preferred embodiment of the present application, the structured feature tensor is input into the pre-trained YOLOV2 deep convolutional network, and the sub-pixel level spatial position vector and the boundary box tensor of the component are obtained through the forward propagation calculation of the convolution kernel weight parameters, including:

[0087] The structured feature tensor is input into the pre-trained YOLOV2 deep convolutional network, and the primary feature representation is generated by convolution kernel weight parameters and calculation; the secondary feature representation is generated by performing down-sampling convolution operation on the primary feature representation;

[0088] The secondary feature representation is calculated by the convolution kernel weight of the regression prediction layer to obtain the boundary box center coordinate offset and the width-height scaling factor; the reference integer coordinate position is determined according to the boundary box center coordinate offset; the local neighborhood analysis area is established with the reference integer coordinate position as the center; the weight distribution calculation is performed in the local neighborhood analysis area, and the weighted average calculation is performed, and finally the sub-pixel level offset vector is generated;

[0089] The sub-pixel level offset vector and the width-height scaling factor are fused to generate the sub-pixel level spatial position vector of the component; the boundary box tensor is generated based on the fusion of the boundary box center coordinate offset and the width-height scaling factor.

[0090] In the embodiment of the application, the network structure is constructed: a classical architecture of 22 convolutional layers + 2 fully connected layers is adopted, the input feature tensor size is set to 416x416x3 (adapted to the channel dimension of the structured feature tensor); the first 13 layers use the convolution kernel parameters (5x5, 3x3 convolution kernel is used alternately, step 1 or 2) of the pre-trained Darknet-19 network, the last 9 layers are self-defined convolutional layers (3x3 convolution kernel is mainly used, and the number of channels is gradually reduced from 1024 to 125); a LeakyReLU activation function is configured after all convolutional layers, and the negative slope a is in the range of 0.01-0.1 (dynamically adjusted according to the feature dimension, 0.01 is taken for high-dimensional feature layers to retain details, and 0.1 is taken for low-dimensional feature layers to enhance gradient flow); maximum pooling (2x2 pooling kernel, step 2) is used to realize down-sampling, and a total of 5 down-sampling operations are performed, and the final feature map size is reduced to 13x13.

[0091] The output layer structure: the regression prediction layer adopts a 1x1 convolution kernel, and the number of output channels corresponds to 5 anchor box parameters (center coordinate offset, width and height scaling factor, confidence) and 20 target probability in each cell of the 13x13 grid.

[0092] Model training process:

[0093] A component dataset containing 100,000+ samples is constructed, covering 20 common components such as resistors, capacitors, and inductors, and each sample is labeled with bounding box coordinates, class labels, and sub-pixel level position true values (collected by high-precision visual measuring instruments); the Darknet-19 weights pre-trained using the ImageNet dataset are used to initialize the first 13 convolutional layers, and the remaining layers are randomly initialized using He normal distribution (mean 0, standard deviation 0.01).

[0094] Training parameter settings: batch size is set to 32, initial learning rate is 0.001 (polynomial decay strategy is adopted, and it is decayed to 0.0001 after 100,000 iterations); momentum parameter 0.9, weight decay coefficient 0.0005.

[0095] The weighted sum of position error, confidence error, and class error is adopted, and the weight ratio is 5:1:1, focusing on optimizing small target position loss; the pre-trained layer weights are frozen for the first 20,000 iterations, and only the self-defined layers are trained; all layers are unfrozen for the last 80,000 iterations, and data augmentation (random rotation ±15°, scaling 0.8-1.2 times, brightness and contrast disturbance) is used to improve the generalization ability.

[0096] Model implementation: based on NVIDIA Tesla V100 GPU acceleration training, with 16GB of video memory; in the inference stage, an industrial-grade edge computing device (such as NVIDIA Jetson AGXXavier) is adapted; the inference time is compressed to 50ms / frame; batch inference mode is enabled, supporting simultaneous processing of 4-8 frames of feature tensors.

[0097] After the structured feature tensor input network, the primary feature representation is calculated by the first 10 convolutional layers (including 5 times of down-sampling), and the feature map size is 52x52x256, which retains the low-order geometric and texture features of the components; through the last 3 convolutional layers, the down-sampling convolution operation (3x3 convolution kernel, step 2) is performed to compress the primary feature representation to the secondary feature representation of 13x13x1024, and the high-order semantic features (such as the overall contour and category attributes) are strengthened.

[0098] The regression prediction layer is calculated by a 1x1 convolution kernel on the secondary feature representation, and outputs the center coordinate offset of each anchor box (normalized offset relative to the grid cell, range [-0.5, 0.5]), the width-height scaling factor (scaling ratio relative to the preset anchor box, range [0.1, 5.0]) and the confidence score.

[0099] According to the center coordinate offset, the corresponding 13x13 grid cell is located to obtain the reference integer coordinate position (pixel-level accuracy); a 3x3 or 5x5 local neighborhood analysis region is established with the reference integer coordinate as the center (5x5 for small-size components to increase the sampling points), which contains 9-25 pixel feature information in the region; the weight is assigned according to the feature response intensity (convolution output value) of the pixels in the neighborhood, and the higher the response, the greater the weight (the total weight is normalized to 1.0), and the edge pixel weight decay coefficient is set to 0.5-0.8 (to suppress noise interference); the weighted average is performed on the pixel coordinates and corresponding weights in the neighborhood to generate a sub-pixel level offset vector (accuracy up to 0.1 pixel level), and the offset range is limited within [-0.5, 0.5] pixels to ensure reasonableness.

[0100] The sub-pixel level offset vector and the reference integer coordinate are fused to convert into a three-dimensional position vector (X, Y, Z axes, Z axis calculated by focal length and pixel offset) in the physical coordinate system; the center coordinate offset, the width-height scaling factor and the sub-pixel offset are fused to generate a bounding box tensor containing the upper-left corner and the lower-right corner coordinates, and each bounding box corresponds to a unique component target.

[0101] The sub-pixel level offset calculation breaks through the limitation of pixel-level positioning, and the component position measurement accuracy meets the requirements of industrial-level precision grabbing. The combination of multi-level convolution and activation function strengthens the semantic expression of structured features, and the bounding box parameters can still be stably output for components with surface reflection and partial occlusion. The parameter interpretability is enhanced: the separation calculation of anchor box parameters and sub-pixel offset makes the bounding box generation process traceable, which is convenient for subsequent geometric feature analysis and path planning of the execution module.

[0102] As Figure 2As shown, based on the bounding box tensor, the contour curvature distribution function and the aspect ratio parameter of the component are calculated, and the spatial geometric feature descriptor is generated; the spatial geometric feature descriptor is input into the pre-constructed component physical feature template library, and the target sorting category is obtained through the Hamming distance minimization principle, including:

[0103] The bounding box tensor is used to extract the circumscribed rectangular region of the component, and the ordered contour point coordinate sequence is obtained by performing the contour point sequence extraction operation in the circumscribed rectangular region; the three-point analysis unit is formed by selecting the contour point and the adjacent points before and after in the contour point coordinate sequence; the curvature estimation value is generated based on the unit calculation of the adjacent line segment direction change, and the curvature value set is generated by traversing all the contour points;

[0104] The frequency distribution statistics operation is performed based on the contour point curvature value set to construct the discrete curvature distribution function; the aspect ratio parameter is calculated according to the width and height parameters in the bounding box tensor;

[0105] The discrete curvature distribution function and the aspect ratio parameter are subjected to feature splicing operation to generate the spatial geometric feature descriptor; the spatial geometric feature descriptor is input into the pre-constructed component physical feature template library, and the target sorting category is obtained through the Hamming distance minimization matching operation.

[0106] In the embodiment of the application, the circumscribed rectangular range of the component is determined based on the bounding box tensor, and the corresponding region in the optimized binary data matrix is intercepted with the left upper corner (wmin, tmin) and the right lower corner (wmax, tmax) coordinates of the bounding box as the boundary, so as to ensure that the component contour is completely contained (the bounding box is expanded outward by 1-2 pixels to avoid edge truncation).

[0107] An 8-direction chain code tracking algorithm is used to extract the contour point: starting from the first foreground edge point in the circumscribed rectangle, adjacent foreground pixels are searched along the upper, lower, left, right and four diagonal directions, the tracking step length is 1 pixel (for small size components) or 2 pixels (for large size components) each time, the coordinates (x, y) of each tracking point are recorded, and the closed contour is formed until the starting point is returned; the sequence length is dynamically adjusted according to the component size to ensure the integrity of the contour details (30 points for small components and 100 points for large components).

[0108] In the ordered contour point sequence, for each contour point Pi (i = 1, 2,..., N), its pre-order point Pi-1 and post-order point Pi+1 are selected to form a three-point analysis unit (when i = 1, the last point is taken as the pre-order point, and when i = N, the first point is taken as the post-order point to ensure sequence continuity); the direction change amount (angle range 0°-180°) of line segments Pi-1-Pi and Pi-Pi+1 is calculated, and the curvature estimation value is generated according to the ratio of the direction change amount to the length of the line segment (the curvature value is smaller at the smooth contour, and the curvature value is larger at the corner, the value range is 0-5, and the larger the value is, the more significant the bending is); all contour points are traversed to generate a complete curvature value set.

[0109] The specific calculation process of the discrete curvature distribution function is as follows:

[0110] Based on the geometric characteristics of the component contour, the preset curvature value effective interval is [-5, 5] (unit: 1 / pixel), covering the convex and concave contour characteristics of common electronic components (resistors, capacitors, chips, etc.); wherein the negative value represents the contour inwardly concave, the positive value represents the outward convex, and the 0 value represents the approximate straight line segment; the extreme curvature values (usually <3% of the total) outside the range are uniformly merged into the interval endpoints (i.e. <-5 is recorded as -5, and >5 is recorded as 5), avoiding the interference of abnormal values on the distribution statistics.

[0111] The [-5, 5] interval is evenly divided into 20 equal-width subintervals (bins), and each subinterval has a width of 0.5 (i.e. [-5, -4.5), [-4.5, -4)…[4.5, 5]); the contour point curvature value set is traversed, and each curvature value is judged to belong to which subinterval, and the curvature value number (frequency) of each interval is accumulated, for example, the curvature value 1.3 belongs to the [1.0, 1.5) interval, and the interval frequency is increased by 1; the curvature value -2.7 belongs to the [-3.0, -2.5) interval, and the corresponding frequency is increased by 1.

[0112] The total number of contour points (usually 50-200 contour points for a single component, dynamically adjusted according to the size) is calculated, the frequency of each subinterval is divided by the total number of points to obtain the curvature frequency of each interval (value range 0-1); the sum of the frequencies of all normalized intervals is 1, forming the core data of the discrete curvature distribution function, for example, the [0, 0.5) interval frequency of a certain resistor contour is 0.6 (indicating that 60% of the contour points are close to a straight line), and the [2.5, 3.0) interval frequency is 0.2 (indicating that 20% of the contour points have moderate convexity).

[0113] Intervals with a frequency lower than 5% are smoothed, and their frequencies are merged into adjacent intervals to reduce noise interference; intervals with a frequency greater than 10% are retained as main feature intervals, and are given higher weights in subsequent feature descriptors to highlight the typical curvature characteristics of the contour (such as the right-angled convexity of chip pins and the circular arc edge of capacitors).

[0114] According to the width (o = wmax-tmin) and height (h = wmax-tmin) parameters in the boundary box tensor, the aspect ratio parameter r = o / h (value range 1-10, circular or square components r≈1, long strip components r>3) is calculated. Before calculation, the size of the boundary box is calibrated to remove 1-2 pixel edge redundancy and ensure size accuracy.

[0115] The discrete curvature distribution function (5-15 dimensions) is spliced with the aspect ratio parameter (1 dimension) to form a spatial geometric feature descriptor with a dimension of 6-16. Before splicing, the curvature distribution function is normalized to make the features of different size components comparable.

[0116] The pre-constructed template library contains standard features of common component categories (resistor, capacitor, chip, connector, etc.): 100-500 samples are collected for each component, and the spatial geometric feature descriptor is extracted to take the mean value as the standard template of the category, which is stored in the library (the template library supports dynamic updating, and the standard template is automatically generated when a new category is added).

[0117] The spatial geometric feature descriptor of the component to be sorted is calculated with the Hamming distance (measuring the difference between feature vectors, value range 0-1) of all standard templates in the template library one by one, and the category corresponding to the smallest distance template is selected as the target sorting category; set the matching threshold (usually 0.3-0.5), when the minimum distance is less than the threshold, it is determined that the matching is successful, otherwise it is marked as an unknown category and needs manual intervention.

[0118] In a preferred embodiment of the present application, based on the sub-pixel level spatial position vector of all components and the target sorting category, a simulated annealing heuristic algorithm is used for global sorting path optimization to generate a component sorting sequence, including:

[0119] Based on the sub-pixel level spatial position vector of all components, an initial sorting path sequence is constructed; based on the target sorting category, a category aggregation degree weight parameter is calculated; based on the category aggregation degree weight parameter and the sub-pixel level spatial position vector, an energy function is defined;

[0120] Based on the initial sorting path sequence and the energy function, the energy value of the path state is calculated, and the initial temperature parameter of simulated annealing is set; based on the path state, a neighborhood solution generation operation is performed, and a new sorting sequence is generated by randomly exchanging the positions of two components in the sorting sequence; based on the energy function, a quantitative evaluation value is generated by calculating the new sorting sequence;

[0121] The path update decision result is generated based on the temperature parameter, the path state energy value and the difference of the quantitative evaluation value; the current path state is updated according to the path update decision result, the neighborhood solution generation is repeatedly executed until the temperature attenuation operation is executed after the state update operation meets the temperature balance condition, and the optimization cycle is repeatedly executed until the convergence condition is met, and finally the component sorting sequence is generated.

[0122] Based on the sub-pixel level spatial position vector of all components (including X and Y coordinate information), an initial sequence is constructed: a "category priority + position proximity" hybrid strategy is adopted, components are grouped according to the target sorting category (continuous arrangement in the same category), and the X coordinate is sorted from small to large within the group (when the X coordinate is similar, the Y coordinate is sorted), and an initial sorting path sequence is generated; if there are N components, the sequence length is N, and each element corresponds to a unique identifier of a component.

[0123] The proportion of the number of components in the same category is counted, and the aggregation degree weight ω (value range 0.3-0.7) is defined, the higher the proportion of the number of components in the same category, the greater the value of ω (such as ω=0.7 when the proportion is >50%, and ω=0.3 when the proportion is <20%); the weight is used to balance the priority of "category aggregation" and "position proximity" in path optimization.

[0124] The energy function is used to quantify the advantages and disadvantages of the path, and the spatial distance and the category aggregation degree are considered comprehensively.

[0125] The spatial distance term: the Euclidean distance sum of adjacent components in the sequence is calculated, the greater the distance, the higher the energy; the category penalty term: if adjacent components belong to different categories, a penalty value is added (the penalty coefficient is positively correlated with ω); the energy function value is the weighted sum of the two terms, and the lower the value, the better the path.

[0126] Initial temperature L0: set according to the energy value E0 of the initial path, L0=b×E0 (b is a proportion coefficient, value 10-20, the more the number of components, the greater the value of b), the initial temperature range is usually 100-1000;

[0127] Temperature attenuation coefficient Ʊ: value 0.85-0.95 (default 0.9), controls the temperature drop speed;

[0128] Balancing iteration number F: the iteration step number at each temperature, set as N×5 (N is the number of components), to ensure sufficient search of the neighborhood at the current temperature;

[0129] Convergence threshold Le: stop optimization when the temperature drops to Le (value 1e-3-1e-2).

[0130] Generate a new solution by random disturbance to the current path sequence: adopt the "random exchange" strategy, randomly select two different index positions q and j (q≠j) in the sequence, exchange the positions of the corresponding components to generate a new sorting sequence; each neighborhood operation only exchanges a pair of elements, ensuring that the new solution has a small difference from the current solution and maintaining the stability of the search.

[0131] Calculate the energy value U_new of the new sorting sequence based on the energy function, compare it with the current sequence energy value U_current, and get the energy difference ΔU=U_new-U_current; ΔU<0 indicates that the new path is better, and ΔU≥0 indicates that the new path is worse.

[0132] If ΔU<0: directly accept the new path, update the current sequence to the new sequence, and update the current energy value to E_new; if ΔE≥0: calculate the acceptance probability V (positively related to temperature L and negatively related to ΔU), generate a random number d between 0 and 1, if d<V accept the new path, otherwise keep the current path; the higher the temperature, the larger V, allowing to accept worse solutions at high temperature to escape local optimum.

[0133] Repeat the "neighborhood solution generation → evaluation → decision" operation at each temperature L to complete temperature balance; after balancing, perform temperature decay: L_new=ʊ×L, update the temperature to L_new.

[0134] Repeat the "neighborhood search → update decision → temperature balance → temperature decay" cycle until one of the following convergence conditions is met: temperature T≤T e ; the energy value change rate of the last 5 temperature stages is less than 1% (the energy tends to be stable); the total number of iterations reaches the upper limit (set to N×20 to avoid excessive calculation); the final output energy value is the lowest path sequence as the component sorting sequence.

[0135] Through the random search at high temperature and the local fine-tuning at low temperature, the problem of traditional greedy algorithm falling into local optimum is effectively avoided, ensuring the generation of globally near-optimal sorting path and reducing the empty travel distance of the mechanical arm. The introduction of the category aggregation weight parameter makes the path planning preferentially gather components of the same category, reduces the cross-category transfer frequency, and improves the efficiency of loading and unloading of classification bins, especially suitable for multi-category mixed sorting scenarios. Through reasonable setting of temperature decay coefficient and balance iteration number, the search accuracy is guaranteed while the calculation time is controlled, which can adapt to the sorting needs of components of different orders of magnitude (tens to hundreds), and meet the real-time requirements of production lines.

[0136] In a preferred embodiment of the application, based on the component sorting sequence, vacuum adsorption gradient control instructions are generated in sequence according to the aspect ratio parameter of each component; three-axis linkage position control is performed according to the sub-pixel level spatial position vector of the component, and the vacuum adsorption end is controlled to load the vacuum adsorption gradient control instructions to capture the target component; and the component is transferred to the corresponding classification bin according to the target sorting category, including:

[0137] The aspect ratio parameter is obtained according to the component sorting sequence; the vacuum adsorption gradient control instructions are generated based on the aspect ratio parameter; and the three-axis linkage displacement parameter calculation is performed based on the sub-pixel level spatial position vector to generate the three-axis linkage displacement parameter;

[0138] The three-axis linkage position control is performed according to the three-axis linkage displacement parameter, so that the vacuum adsorption end is accurately positioned above the target component; and the vacuum adsorption end is controlled to load the vacuum adsorption gradient control instructions to capture the target component in the accurate positioning state;

[0139] The spatial coordinates of the corresponding classification bin are determined based on the target sorting category; and the three-axis linkage position control is performed based on the spatial coordinates of the classification bin to transfer the component to the target position.

[0140] According to the component sorting sequence, the aspect ratio parameter (defined as the ratio of width to height) of each component is extracted in sequence, and the ratio is divided into three categories according to the ratio range: small precision type (aspect ratio 0.5-1.5), medium conventional type (1.5-3.0), and large special-shaped type (3.0-5.0); the classification threshold can be dynamically calibrated through the physical feature template library to adapt to the structural characteristics of different types of components.

[0141] In an embodiment of the application, the vacuum adsorption gradient control instructions are generated based on the classification results:

[0142] Small precision type: low negative pressure gradient is adopted, the initial negative pressure value is set to -20kPa to -40kPa, and the pressure fluctuation is controlled within ±5kPa in the adsorption maintenance stage to avoid deformation of the component caused by excessive adsorption.

[0143] Medium conventional type: medium negative pressure gradient is adopted, the initial negative pressure value is set to -40kPa to -70kPa, and dynamic fine adjustment is made according to the material difference (metal / plastic), the upper limit is taken for metal parts, and the lower limit is taken for plastic parts.

[0144] Large special-shaped type: hierarchical pressure boosting gradient is adopted, pre-adsorption is performed at an initial negative pressure of -50kPa, linear pressure boosting is performed to -70kPa to -90kPa within 0.5 seconds to ensure the stability of large-area contact.

[0145] The instruction format includes three core parameters: negative pressure starting value, pressure boosting rate (5kPa / ms-10kPa / ms), and maintenance pressure value, which are sent to the negative pressure control unit through the RS485 bus.

[0146] Convert the sub-pixel level spatial position vector of the component (unit: pixel) to the Cartesian coordinate three-axis physical coordinates (unit: mm): calculate the X and Y axis plane coordinates through the preset pixel-physical size conversion coefficient (such as 0.01 mm / pixel), and estimate the initial height of the Z axis (5-15 mm away from the surface of the conveying belt, and a small value is taken for small components) in combination with the height parameter of the bounding box tensor.

[0147] X / Y axis: adopt a trapezoidal velocity curve to plan the displacement parameters, the acceleration is limited within 500 mm / s2 in the acceleration stage, the speed is dynamically adjusted according to the distance in the uniform speed stage (near distance ≤300 mm / s, far distance ≤500 mm / s), and the deceleration stage starts to decelerate 5 mm in advance to ensure that there is no impact during positioning.

[0148] Z axis: independently plan the lifting displacement, the descending speed is ≤200 mm / s from the initial height to the adsorption height (1-3 mm away from the surface of the component), and the speed is reduced to ≤50 mm / s at the moment of contact to avoid crushing the component.

[0149] The displacement parameters include three types of parameters such as target coordinates, speed threshold and acceleration threshold, and the pulse output control is realized through the motion control card.

[0150] After receiving the displacement parameters, the three-axis motion mechanism performs linkage control: the X / Y axis is first moved to the target plane coordinate in linkage, and the positioning error is controlled within ±0.02 mm; then the Z axis is lowered to the adsorption height, which is calibrated in real time (sampling frequency 1 kHz) through the laser displacement sensor to ensure that the parallelism deviation between the end and the surface of the component is ≤0.5°.

[0151] After positioning is completed, the negative pressure control unit loads the vacuum adsorption gradient control instruction: first output 50% of the preset negative pressure for pre-adsorption (for 100 ms), detect the adsorption state through the pressure sensor (feedback value deviation ≤10% is considered as effective adsorption); confirm the effectiveness and load the full negative pressure to maintain the adsorption state until it is moved to the classification bin. If adsorption failure (feedback value anomaly for 200 ms) is detected, the retry mechanism (maximum of 3 retries) is triggered.

[0152] Path planning: based on the spatial coordinates of the target classification bin (pre-stored in the system database with an accuracy of ±0.1 mm), the shortest path from the current position to the classification bin is planned to avoid other components and mechanical interference areas.

[0153] Moving process: the Z axis is first raised to a safe height (≥20 mm), the X / Y axis is moved to above the classification bin in linkage, the Z axis is lowered to a placement height (3-5 mm away from the bottom surface of the classification bin), the negative pressure unit is reduced by gradient (10 kPa / 100 ms) to release the component at normal pressure to avoid displacement caused by air flow impact.

[0154] Through the length-width ratio adaptive negative pressure gradient control, the adsorption success rate of different size components is improved, the damage rate of small precision parts is reduced, and the leakage rate of large special-shaped parts is reduced. Pressure feedback detection and retry mechanism reduces abnormal failure, classified warehouse pre-storage coordinates and path obstacle avoidance design reduces mechanical interference risk, and running stability is improved. Gradient parameters can be dynamically updated through template library, without hardware modification to adapt to new component categories, reducing system iteration cost.

[0155] Embodiments of the present application also provide a computing device, comprising: a processor, a memory storing a computer program, when the computer program is run by the processor, the system as described above is executed. All implementation manners in the above system embodiment are applicable to this embodiment, and the same technical effects can also be achieved.

[0156] Embodiments of the present application also provide a computer readable storage medium, storing instructions, when the instructions are run on a computer, the computer executes the system as described above. All implementation manners in the above system embodiment are applicable to this embodiment, and the same technical effects can also be achieved.

[0157] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A component visual recognition automatic sorting system, characterized in that, The method comprises the following steps: a processing module is used to synchronously scan the multi-component devices in a scattered state to generate an original optical signal data set; the original optical signal data set is preprocessed to generate an optimized binary data matrix; a parsing module is used to extract the geometric and topographic features of the component devices based on the optimized binary data matrix by using a multi-order Gaussian differential operator, and to parse the surface identification information by using a local texture feature encoder to construct a structured feature tensor; a calculation module is used to input the structured feature tensor into a pre-trained YOLOV2 deep convolutional network, and to obtain the sub-pixel level spatial position vector and the boundary box tensor of the component devices through the forward propagation calculation of the convolution kernel weight parameters; an execution module is used to calculate the profile curvature distribution function and the aspect ratio parameter of the component devices based on the boundary box tensor, and to generate a spatial geometric feature descriptor; the spatial geometric feature descriptor is input into a pre-constructed component device physical feature template library to obtain the target sorting category by the Hamming distance minimization principle; a sorting module is used to perform global sorting path optimization by using an simulated annealing heuristic algorithm based on the sub-pixel level spatial position vector and the target sorting category of all component devices to generate a component device sorting sequence; a control module is used to generate a vacuum adsorption gradient control instruction according to the aspect ratio parameter of each component device based on the component device sorting sequence; to perform three-axis linkage position control according to the sub-pixel level spatial position vector of the component devices, and to control the negative pressure adsorption end to load the vacuum adsorption gradient control instruction to capture the target component device, and to move the component device to the corresponding classification bin according to the target sorting category.

2. The component visual recognition automatic sorting system according to claim 1, characterized in that, Synchronously scanning the multi-component devices in a scattered state to generate an original optical signal data set; Preprocessing the original optical signal data set to generate an optimized binary data matrix, comprising: synchronously scanning the scattered component devices on the conveying belt by using a linear array CCD sensor and a ring-shaped LED light source to obtain the gray value of each pixel point and generate an original optical signal data set; performing median filter processing on the original optical signal data set to generate a denoised optical signal data set; performing pixel-by-pixel difference operation on the denoised optical signal data set and the low-frequency background component to generate a background-suppressed optical signal data set; performing local contrast limited histogram equalization processing on the background-suppressed optical signal data set to obtain an enhanced optical signal data set; and performing adaptive threshold segmentation processing on the enhanced optical signal data set to generate an optimized binary data matrix.

3. The component visual recognition automatic sorting system according to claim 2, wherein, Based on the optimized binary data matrix, the geometric and topographic features of the component devices are extracted by using a multi-order Gaussian differential operator, and the surface identification information is parsed by using a local texture feature encoder to construct a structured feature tensor, comprising: applying a second-order Gaussian differential convolution processing to the optimized binary data matrix, performing X-axis gradient convolution operation by using a first-order Gaussian differential operator to generate a gradient amplitude distribution matrix; and performing Laplace convolution operation by using a second-order Gaussian differential operator to generate a curvature feature distribution matrix; Perform spatial feature fusion processing based on the gradient amplitude distribution matrix and the curvature feature distribution matrix, establish a spatial coordinate mapping relationship by detecting local maximum points of the gradient amplitude distribution matrix, extract extreme value features in the curvature feature distribution matrix corresponding to the coordinate region, and generate a geometric topographic feature matrix; Perform local texture feature coding operation on the geometric topographic feature matrix, calculate the contrast statistics and entropy statistics of the gray level co-occurrence matrix in the predetermined analysis window centered on the feature point, and jointly encode them into a texture feature description vector; Stack the geometric topographic feature matrix and the texture feature description vector in multiple channels to construct a structured feature tensor.

4. The component visual recognition automatic sorting system according to claim 3, characterized in that, Input the structured feature tensor into the pre-trained YOLOV2 deep convolutional network, and calculate the sub-pixel level spatial position vector and the bounding box tensor of the component through the forward propagation of the convolution kernel weight parameters, including: Input the structured feature tensor into the pre-trained YOLOV2 deep convolutional network, and generate a primary feature representation through convolution kernel weight parameters and calculation; perform down-sampling convolution operation on the primary feature representation to generate a secondary feature representation; Calculate the bounding box center coordinate offset and the width-height scaling factor by calculating the secondary feature representation through the convolution kernel weight of the regression prediction layer; determine the reference integer coordinate position according to the bounding box center coordinate offset; establish a local neighborhood analysis region centered on the reference integer coordinate position; perform weight distribution calculation in the local neighborhood analysis region, and perform weighted average calculation to finally generate a sub-pixel level offset vector; Fuse the sub-pixel level offset vector and the width-height scaling factor to generate a sub-pixel level spatial position vector of the component; fuse the bounding box center coordinate offset and the width-height scaling factor to generate a bounding box tensor.

5. The component visual recognition automatic sorting system according to claim 4, characterized in that, Based on the bounding box tensor, calculate the profile curvature distribution function and the aspect ratio parameter of the component to generate a spatial geometric feature descriptor; Input the spatial geometric feature descriptor into the pre-constructed component physical feature template library to obtain the target sorting category through the Hamming distance minimization principle, including: Extract the bounding rectangle region based on the bounding box tensor, and perform profile point sequence extraction operation in the bounding rectangle region to obtain an ordered profile point coordinate sequence; select three-point analysis units composed of profile points and adjacent points before and after in the profile point coordinate sequence; generate curvature estimation values based on the direction change amount of adjacent line segments in the unit, and generate a curvature value set by traversing all profile points; Perform frequency distribution statistics operation based on the profile point curvature value set to construct a discrete curvature distribution function; calculate the aspect ratio parameter according to the width and height parameters in the bounding box tensor; Perform feature splicing operation on the discrete curvature distribution function and the aspect ratio parameter to generate a spatial geometric feature descriptor; input the spatial geometric feature descriptor into the pre-constructed component physical feature template library to obtain the target sorting category through the Hamming distance minimization matching operation.

6. The automatic component visual recognition and sorting system according to claim 5, wherein, Based on the sub-pixel level spatial position vector of all components and the target sorting category, use the simulated annealing heuristic algorithm to perform global sorting path optimization to generate a component sorting sequence, including: constructing an initial sorting path sequence based on the sub-pixel level spatial position vectors of all components; calculating a category aggregation degree weight parameter based on the target sorting category; defining an energy function based on the category aggregation degree weight parameter and the sub-pixel level spatial position vectors; calculating the energy value of the path state based on the initial sorting path sequence and the energy function, and setting an initial temperature parameter of simulated annealing; performing a neighborhood solution generation operation based on the path state, generating a new sorting sequence by randomly exchanging the positions of two components in the sorting sequence; calculating the new sorting sequence based on the energy function to generate a quantitative evaluation value; judging based on the temperature parameter, the energy value of the path state, and the difference of the quantitative evaluation value to generate a path update decision result; updating the current path state according to the path update decision result, repeating the neighborhood solution generation to state update operation until the temperature decay operation is performed after the temperature balance condition is met, and repeating the optimization cycle until the convergence condition is met, finally generating the component sorting sequence.

7. The component visual recognition automatic sorting system according to claim 6, characterized in that, based on the component sorting sequence, generating vacuum adsorption gradient control instructions in turn according to the aspect ratio parameter of each component; performing three-axis linkage position control according to the sub-pixel level spatial position vector of the component, and controlling the negative pressure adsorption end to load the vacuum adsorption gradient control instruction to capture the target component, and moving the component to the corresponding classification bin according to the target sorting category, including: obtaining the aspect ratio parameter according to the component sorting sequence; generating vacuum adsorption gradient control instructions based on the aspect ratio parameter; calculating the three-axis linkage displacement parameter based on the sub-pixel level spatial position vector to generate the three-axis linkage displacement parameter; performing three-axis linkage position control according to the three-axis linkage displacement parameter to accurately position the negative pressure adsorption end above the target component; controlling the negative pressure adsorption end to load the vacuum adsorption gradient control instruction to capture the target component in the accurate positioning state; determining the spatial coordinates of the corresponding classification bin based on the target sorting category; performing three-axis linkage position control based on the spatial coordinates of the classification bin to move the component to the target position. 8.The component visual recognition automatic sorting system according to claim 7, wherein, inputting the structured feature tensor into the pre-trained YOLOV2 deep convolutional network to generate primary feature representation through convolution kernel weight parameters; performing subsampling convolution operation on the primary feature representation to generate secondary feature representation, including: inputting the structured feature tensor into the convolution kernel weight parameters of the pre-trained convolution layer to perform spatial feature extraction operation and generate initial feature representation; performing feature value sign determination processing on each feature value in the initial feature representation; performing feature value reset operation according to the sign determination result, resetting to zero value if the feature value is negative, and maintaining the original value if the feature value is non-negative; generating active feature representation based on the reset feature value; performing spatial resolution reduction operation on the active feature representation to generate downsampled feature representation; inputting the downsampled feature representation into the convolution kernel weight parameters of the secondary convolution layer to perform feature enhancement operation and generate primary feature representation; performing convolution operation with reduced feature map size on the primary feature representation to generate secondary feature representation.

9. A computing device, comprising: including: one or more processors; a storage device for storing one or more programs, when executed by the one or more processors, cause the one or more processors to implement the system of any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program, which is executed by the processor to implement the system of any one of claims 1-8.

Images