Intelligent identification method and system for high-value components of waste circuit board

Abstract

The invention belongs to the technical field of electronic waste recovery processing, and discloses a waste circuit board high-value component intelligent identification method and system, and the method comprises the steps: obtaining an original image, and carrying out the preprocessing, and obtaining a final image; constructing and training a target detection model, and detecting the final image to obtain an initial detection result; obtaining a final detection result based on the final image and the initial detection result in combination with threshold filtering and a non-maximum suppression algorithm; identifying the final detection result by using an OCR (Optical Character Recognition) algorithm to obtain an identification result; constructing a component database, matching the recognition results in combination with a similarity algorithm to obtain a matching result, and obtaining an evaluation result by using a time sequence prediction model and a graph neural network model; and integrating and outputting the detection result, the identification result and the evaluation result. According to the invention, when the waste circuit board is recycled, high-value components in the waste circuit board can be accurately identified and evaluated, and the value of the whole waste circuit board is evaluated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic waste recycling, and particularly relates to a method and system for intelligently identifying high-value components of waste circuit boards. BACKGROUND

[0002] With the rapid growth of electronic waste, the pollution and waste of waste circuit boards are becoming increasingly serious. In order to maximize the value of resources, it is necessary to accurately identify and evaluate high-value components in waste circuit boards during the fine recycling of waste circuit boards.

[0003] However, the traditional intelligent identification scheme for high-value components in waste circuit boards has the following problems:

[0004] 1. The pixel proportion of micro components in the image is small, and the detailed features are not obvious. Traditional methods are difficult to capture the key features of micro components, such as pin shape and size, resulting in reduced accuracy of identification.

[0005] 2. The layout of components on the circuit board is complex, and the components are staggered in height, with mutual occlusion or small spacing. Traditional methods are difficult to accurately separate and extract the features of each component, which can easily cause confusion of adjacent component features, and thus affect the accuracy and stability of identification.

[0006] 3. Traditional methods are usually trained based on specific data sets, and have poor generalization ability when facing different types and specifications of circuit boards or newly emerging component types, making it difficult to accurately identify.

[0007] 4. The circuit board substrate may have various textures, colors and other interference factors. Traditional methods have difficulty in distinguishing components and background, and are prone to misidentifying background as components or ignoring some components similar to the substrate color.

[0008] 5. Traditional methods have high computational complexity when processing images, resulting in slow identification speed and difficulty in meeting real-time requirements in high real-time scenarios.

[0009] Therefore, how to provide a method and system for intelligently identifying high-value components of waste circuit boards is a problem to be solved at present. SUMMARY

[0010] The embodiments of the present application provide a method for intelligently identifying high-value components of waste circuit boards, to solve the problems of low identification accuracy and poor stability when facing micro circuit board components in the prior art.

[0011] The following presents a simplified summary of some aspects of the disclosure in order to provide a basic understanding. This summary is not an extensive overview of the disclosure and is not intended to identify key / critical elements or to delineate the scope of the disclosure. Its sole purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is presented later.

[0012] According to a first aspect of embodiments of the present application, a waste circuit board high-value component intelligent identification method is provided.

[0013] In one embodiment, a waste circuit board high-value component intelligent identification method includes: obtaining an original image and pre-processing the original image to obtain a final image;

[0014] A target detection model is constructed and trained, and the final image is detected by the target detection model to obtain an initial detection result;

[0015] Based on the final image and the initial detection result, and combined with threshold filtering and non-maximum suppression algorithm, a final detection result is obtained;

[0016] An OCR recognition algorithm is used to identify each component image in the final detection result to obtain an identification result;

[0017] A component database is constructed, and the identification result is matched by combining a similarity algorithm to obtain a matching result; based on the matching result, a time series prediction model and a graph neural network model are used to obtain an evaluation result;

[0018] The final detection result, the identification result and the evaluation result are integrated for visual output.

[0019] In one embodiment, obtaining an original image and pre-processing the original image to obtain a final image includes the following steps:

[0020] The original image is obtained, and a lossy compression algorithm is used to compress the original image to obtain a first image;

[0021] A threshold segmentation algorithm is used to segment the first image to obtain a second image;

[0022] Based on a preset size, the second image is block cropped to obtain a third image;

[0023] According to a histogram equalization algorithm, the third image is color optimized to obtain the final image.

[0024] In one embodiment, a target detection model is constructed and trained, and the final image is detected by the target detection model to obtain an initial detection result, including the following steps:

[0025] A Darknet53 network optimized across stages is used as a backbone network to build a target detection model;

[0026] A dynamic positive sample distribution algorithm and a loss function algorithm are used to train the target detection model;

[0027] Based on the final image, the target detection model is used to detect the final image to obtain an initial detection result.

[0028] In one embodiment, the dynamic positive sample distribution algorithm and the loss function algorithm are used to train the target detection model, including the following steps:

[0029] A training image is obtained and preprocessed to obtain a standard image;

[0030] The backbone network of the target detection model is used to extract the standard image to obtain a feature map;

[0031] The neck network of the target detection model is used to perform feature fusion on the feature map to obtain a feature fusion map;

[0032] The decoupled detection head is used in combination with the dynamic positive sample distribution algorithm and the loss function algorithm to optimize the parameters of the target detection model.

[0033] In one embodiment, the decoupled detection head is used in combination with the dynamic positive sample distribution algorithm and the loss function algorithm to optimize the parameters of the target detection model, including the following steps:

[0034] The decoupled detection head is used to perform position and category prediction on the feature fusion map to obtain classification scores and regression box parameters;

[0035] Based on the dynamic positive sample distribution algorithm, the classification scores and regression box parameters are dynamically distributed to obtain positive and negative sample labels and matching relationships;

[0036] Based on the classification scores, regression box parameters, dynamically distributed classification scores and regression box parameters, and using the loss function algorithm, the parameters of the target detection model are optimized.

[0037] In one embodiment, based on the final image and the initial detection result, and in combination with a threshold filtering and a non-maximum suppression algorithm, a final detection result is obtained, including the following steps:

[0038] Based on the initial detection result, a confidence threshold is used to filter to obtain a preliminary prediction box;

[0039] A non-maximum suppression algorithm is used to merge the preliminary prediction box to obtain a detection box;

[0040] Based on the detection box, a ResNet model is used to verify the detection box to obtain a final detection box;

[0041] Superimpose the final detection frame into the final image to obtain a final detection result.

[0042] In an embodiment, an OCR recognition algorithm is used to recognize each component image in the final detection result to obtain a recognition result, including the following steps:

[0043] Each component image in the final detection result is preprocessed to obtain a plurality of standard detection images.

[0044] An OCR recognition algorithm is used to perform OCR recognition on the plurality of standard detection images to obtain a plurality of initial recognition results.

[0045] Based on a preset text length threshold, the plurality of initial recognition results are filtered to obtain a plurality of standard recognition results.

[0046] The plurality of standard recognition results are screened according to text similarity and confidence factors to obtain a recognition result.

[0047] In an embodiment, preprocessing each component image in the final detection result to obtain a plurality of standard detection images includes the following steps:

[0048] Each component image in the final detection result is subjected to size screening to obtain a first detection image.

[0049] Based on a predefined rotation matrix, the first detection image is subjected to angle transformation to obtain second detection images at different angles.

[0050] The second detection images at different angles are subjected to vertical stacking using a vertical stacking function to obtain a plurality of standard detection images.

[0051] In an embodiment, a component database is constructed, and a similarity algorithm is combined to match the recognition result to obtain a matching result; based on the matching result, a time series prediction model and a graph neural network model are used to obtain an evaluation result, including the following steps:

[0052] A component database is constructed.

[0053] Based on a similarity algorithm, the similarity of the recognition result to component models in the component database is calculated to screen a matching result.

[0054] A time series prediction model is used to perform time series prediction on the matching result to obtain a prediction result.

[0055] Based on a graph neural network model and in combination with the prediction result, the matching result is evaluated to obtain an evaluation result.

[0056] According to a second aspect of the embodiments of the present application, a waste circuit board high-value component intelligent identification system is provided.

[0057] In one embodiment, a waste circuit board high-value component intelligent identification system comprises:

[0058] An acquisition and preprocessing module is configured to acquire an original image and pre-process the original image to obtain a final image.

[0059] A detection module is configured to construct and train a target detection model, detect the final image by using the target detection model, and obtain an initial detection result.

[0060] A post-processing module is configured to obtain a final detection result based on the final image and the initial detection result, and in combination with threshold filtering and non-maximum suppression algorithm.

[0061] An identification module is configured to identify each component image in the final detection result by using an OCR identification algorithm to obtain an identification result.

[0062] An evaluation module is configured to construct a component database, match the identification result by using a similarity algorithm to obtain a matching result, and obtain an evaluation result by using a time series prediction model and a graph neural network model based on the matching result.

[0063] An output module is configured to integrate the final detection result, the identification result, and the evaluation result, and perform visual output.

[0064] According to a third aspect of the embodiments of the present application, a computer device is provided.

[0065] In some embodiments, the computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0066] According to a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided.

[0067] In one embodiment, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0068] The technical solutions provided by the embodiments of the present application can include the following beneficial effects:

[0069] 1. The present application can intelligently identify waste circuit board high-value components. The present application uses a YOLOx target detection model and an anti-interference OCR algorithm, and the component positioning accuracy in complex scenes reaches 95%, and the character recognition rate is over 85%.

[0070] 2、The application can accurately identify and evaluate high-value components in waste circuit boards during recycling, and evaluate the value of the whole waste circuit board. The application integrates global component transaction data, including the purchase value of new components, secondary utilization value, and material recycling value, builds a multi-dimensional dynamic database, and matches the identification results with the database to estimate the value of the waste circuit board.

[0071] 3、The application reduces the pollution of toxic and harmful substances to recycled materials through identification and classification of various components, and improves the resource recovery rate.

[0072] 4、The application supports the access of industrial cameras, mobile terminals and other devices, and meets the real-time processing needs of complex scenes such as disassembly workshops and recycling sites.

[0073] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the application. BRIEF DESCRIPTION OF DRAWINGS

[0074] The drawings incorporated into the specification and forming part of the specification, show embodiments consistent with the application, and together with the specification, serve to explain the principles of the application.

[0075] Figure 1 is a flow chart of a waste circuit board high-value component intelligent identification method according to an exemplary embodiment;

[0076] Figure 2 is a principle block diagram of a waste circuit board high-value component intelligent identification system according to an exemplary embodiment;

[0077] Figure 3 is a flow chart of the post-processing stage of a waste circuit board high-value component intelligent identification method according to an exemplary embodiment;

[0078] Figure 4 is a structural schematic diagram of a computer device according to an exemplary embodiment. DETAILED DESCRIPTION

[0079] The following description and drawings are illustrative of specific embodiments thereof and are not intended to limit the scope of the embodiments. Parts and features of some embodiments can be included or substituted in or for parts and features of other embodiments. The scope of the embodiments encompassed herein includes the whole scope of the claims together with all available equivalents of the claims. In this document, the terms "first", "second", etc. are used merely to distinguish one element from another, and do not require or imply any actual relationship or order between the elements. In fact, the first element could be called the second element, and vice versa. Also, the terms "comprises", "comprising", or any other variations thereof are intended to cover a non-exclusive inclusion, such that a structure, device or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such structure, device or apparatus. An element proceeded by "comprises... " does not, without more constraints, exclude the existence of additional identical elements in the structure, device or apparatus that includes the element. Various embodiments are described in progressive stages, each of which focuses on the differences from other embodiments, and the same or similar parts between various embodiments can be referred to each other.

[0080] The terms "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like as used herein to indicate orientation or positional relationships based on the orientations or positional relationships shown in the drawings, are only for convenience of description and simplification of description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description herein, unless otherwise specified and limited, the terms "mount", "connect", "connection" should be interpreted broadly, for example, it can be a mechanical connection or an electrical connection, it can be a communication between two elements inside, it can be a direct connection or an indirect connection through an intermediate medium, and the specific meaning of the above terms can be understood by the person skilled in the art according to the specific circumstances.

[0081] In this document, the term "multiple" means two or more, unless otherwise specified.

[0082] In this document, the character " / " represents an "or" relationship between the objects before and after it. For example, A / B means: A or B.

[0083] In this document, the term "and / or" is a description of the relationship between the objects, which means that there can be three relationships. For example, A and / or B means: A or B, or, A and B, the three relationships.

[0084] It should be understood that although the steps in the flowchart are shown in a sequential order following the arrows, the steps are not necessarily executed in the order shown by the arrows. Unless otherwise explicitly stated herein, the execution of the steps is not necessarily limited to the order shown in the figure, and the steps can be executed in other orders. Moreover, at least some of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the sub-steps or stages is not necessarily sequential, but can be round-robin or alternating with at least some of the other steps or sub-steps or stages of the other steps.

[0085] The various modules in the device or system of the present application can be implemented wholly or partially by software, hardware, and combinations thereof. The various modules described above can be embedded in the processor in the computer device in hardware form or independent of the processor in the computer device, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the various modules.

[0086] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0087] Figure 1 An embodiment of a method for intelligently identifying high-value components and devices of waste circuit boards is shown.

[0088] In this optional embodiment, the method for intelligently identifying high-value components and devices of waste circuit boards comprises:

[0089] S101, an original image is acquired, and the original image is preprocessed to obtain a final image.

[0090] In this optional embodiment, acquiring the original image and preprocessing the original image to obtain the final image comprises the following steps:

[0091] The original image is acquired, and a lossy compression algorithm is used to compress the original image to obtain a first image;

[0092] A threshold segmentation algorithm is used to segment the first image to obtain a second image;

[0093] The second image is divided and cropped based on a preset size to obtain a third image;

[0094] The third image is color-optimized based on a histogram equalization algorithm to obtain the final image.

[0095] It should be explained that acquiring the image needs to collect the original image of the waste circuit board from the corresponding device (line scanning camera), so as to ensure that the image can clearly reflect the overall appearance, component and device layout, and various identification details of the circuit board.

[0096] Under the premise of ensuring that the key features are not lost, the original image file volume is compressed to 10-30% of the original size by using lossy compression algorithms such as JPEG / WEBP, reducing storage and transmission costs. Through threshold segmentation and contour detection technology, the non-circuit board area in the image is automatically cut off, and the effective area ratio is ≥95%. The specific steps are as follows:

[0097] 1. Convert the image to grayscale and apply a Gaussian blur with a kernel size of 5x5 to reduce noise interference.

[0098] 2. The image is divided into two parts (foreground and background), and the Otsu algorithm (also known as the maximum inter-class variance method) is used to automatically determine the optimal threshold value, generating a binary image that effectively separates the background and foreground. The expression formula of Otsu algorithm is:

[0099] U = W0*U0 + W1*U1;

[0100] g = W0*(U0-U) 2 +W1*(U1-U) 2 ;

[0101] Where W0 represents the proportion of foreground pixels in the image; U0 represents the mean value of W0; W1 represents the proportion of background pixels in the image; U1 represents the mean value of W1; U represents the overall mean value of the image; g represents the inter-class variance under different segmentation thresholds, and when g is maximum, the corresponding threshold is the optimal threshold.

[0102] 3. Connect adjacent regions and fill holes by closing operation (first expansion and then corrosion) with a kernel size of 30x30 to ensure the integrity of the circuit board area.

[0103] 4. Find all contours in the image and select the largest contour as the circuit board area.

[0104] 5. Expand the bounding box by 1% of the original image size in pixels to ensure complete retention of edge information.

[0105] 6. Crop according to the cropped bounding box to finally retain at least 95% of the actual circuit board area.

[0106] The expression formula of the effective area ratio is:

[0107]

[0108] Considering the complexity of large circuit board images in processing and the high demand for computing resources, the large images after preliminary processing are finely cropped into several small images according to specific size rules, supporting custom block size, which can automatically adjust the block number according to the circuit board size, and is conducive to improving the recognition accuracy of local complex component layout and identification. The specific steps are as follows:

[0109] 1、Get the original image information

[0110] Analyze the collected original circuit board image to obtain its image size and resolution information, understand the approximate distribution of components on the circuit board, and preliminarily judge the complexity of the image, high-density component area, etc. The collected image data can be transmitted to the subsequent processing flow in a suitable format and stable transmission mode, ensuring the coherence and efficiency of the entire system operation.

[0111] 2、Custom block size

[0112] Users can customize the size of the block according to the clarity and pixel requirements. The fewer elements each image contains, the higher the accuracy of recognition, but the number of images to be processed will increase. Therefore, users can choose the appropriate block size to meet their needs for efficiency and accuracy.

[0113] 3、Block cropping

[0114] During the cropping process, ensure that there is a certain overlap between the blocks to avoid missing components or features, and ensure that each component is complete on at least one image. Therefore, the size of the overlapping area between two images when blocking needs to be determined based on the size of the largest component in the processed material. The size of the overlapping area can be adjusted at any time according to the specific situation of the circuit board and the requirements of the recognition algorithm.

[0115] 4、Save the block image

[0116] Save each small image obtained by cropping. The saving format can be selected according to the needs of subsequent processing, such as common JPEG, PNG, etc. When saving, you can add appropriate identifiers to each small image to accurately restore its position and related information in the original circuit board image after processing.

[0117] The collected original image may have color deviation, insufficient contrast, etc. Based on the gray world assumption algorithm, automatically correct the color deviation under different lighting conditions, use CLAHE (Contrast Limited Adaptive Histogram Equalization) technology to improve the local contrast of the image by 20-40%, making the key elements such as components and identifiers on the circuit board more prominent in color, facilitating intelligent algorithms to extract features. The specific steps are as follows:

[0118] 1、Divide the image into multiple small regions (windows).

[0119] 2、For each window, count the number of pixels for each gray level, with a gray level range of [0, L-1] (e.g. 8-bit image L = 256), and the number of pixels corresponding to gray value l is n l , get the histogram.

[0120] 3、Limit the number of pixels of each gray scale to not more than the threshold (self-setting), and evenly distribute the excess pixel number to all gray scales, so as to place the noise too large.

[0121] 4、The histogram is normalized to a probability distribution, and the cumulative distribution function (CDF) is calculated, and the expression formula of the cumulative distribution function is:

[0122]

[0123] In the formula, N represents the total number of pixels of the window; P l represents the probability of the occurrence of the gray value; k represents the gray value; the value range of CDF is [0, 1], which represents the proportion of pixels with a gray value less than or equal to k.

[0124] 5、Using CDF to map the original gray value k to the new intensity value S k , realize contrast stretching, the purpose is to stretch the distribution range of CDF to the full gray range (0 to L-1), enhance the local contrast, and the mapping expression formula is:

[0125] S k =(L-1)×CDF(k).

[0126] S102、Construct and train the target detection model, and detect the final image through the target detection model to obtain an initial detection result.

[0127] In this optional embodiment, constructing and training the target detection model, and detecting the final image through the target detection model to obtain an initial detection result includes the following steps:

[0128] Darknet53 optimized by the cross-stage partial network is used as a backbone network to construct the target detection model;

[0129] The dynamic positive sample allocation algorithm is used in combination with the loss function algorithm to train the target detection model;

[0130] Based on the final image, the final image is detected in combination with the target detection model to obtain an initial detection result.

[0131] In this optional embodiment, the dynamic positive sample allocation algorithm is used in combination with the loss function algorithm to train the target detection model, including the following steps:

[0132] Obtain a training image, and pre-process the training image to obtain a standard image;

[0133] The backbone network of the target detection model is used to extract the standard image to obtain a feature map;

[0134] The neck network based on the target detection model performs feature fusion on the feature map to obtain a feature fusion map;

[0135] The parameters of the target detection model are optimized by using the decoupled detection head, in combination with a dynamic positive sample distribution algorithm and a loss function algorithm.

[0136] In this optional embodiment, the optimization of the parameters of the target detection model by using the decoupled detection head, in combination with the dynamic positive sample distribution algorithm and the loss function algorithm, includes the following steps:

[0137] The decoupled detection head is used to perform position and category prediction on the feature fusion map to obtain classification scores and regression box parameters;

[0138] Based on the dynamic positive sample distribution algorithm, the classification scores and the regression box parameters are dynamically distributed to obtain positive and negative sample labels and matching relationships;

[0139] Based on the classification scores, the regression box parameters, the dynamically distributed classification scores and the regression box parameters, and by using the loss function algorithm, the parameters of the target detection model are optimized.

[0140] It should be explained that the improved YOLOx target detection model improves the small target detection capability, the rectangular frame positioning accuracy and the image processing time.

[0141] The YOLOx target detection model adopts an Anchor-free manner, compared with the Anchor-based manner commonly used in other algorithms in the YOLO series. This method reduces the amount of calculation and does not involve IoU calculation, and the number of predicted frames produced will also be less. The predicted frame of this method is only one-third of the Anchor-based method. Since most of the predicted frames are negative samples, the Anchor-free method can reduce the number of negative samples and further alleviate the imbalance between positive and negative samples.

[0142] The main network of the YOLOx target detection model mainly includes an input end, a backbone network, a neck network, and a detection head. The input end part adopts random horizontal flip, color jitter, multi-scale enhancement, and other data enhancement methods for image preprocessing. The backbone part adopts Darknet53 to extract picture features. The neck part adopts a path aggregation feature pyramid (PAFPN) for feature fusion. The head part performs result prediction, mainly adopting a decoupled head, an anchor-free method, and a multi-positive method.

[0143] Anchor-free means anchor-free, which is a method adopted by a target detection algorithm, and does not rely on a pre-set anchor box to predict the target position, but directly predicts the boundary or key point of the target, simplifies the calculation process, and improves the detection flexibility.

[0144] Anchor-based means based on anchor box, which is a method adopted by a traditional target detection algorithm. Different sizes and proportions of anchor boxes are pre-set, and whether the anchor boxes contain targets is judged by sliding in the image. The algorithm performs regression and classification according to the matching degree of the predicted anchor box and the real target.

[0145] IoU means Intersection over Union, which is an index used to measure the overlapping degree between the predicted box and the real box in target detection.

[0146] Darknet53 is a deep convolutional neural network structure composed of 53 convolutional layers.

[0147] The training method of the YOLOx target detection model is as follows:

[0148] 1. Data preprocessing is performed on the input image. The input image of any size is scaled to the preset size of the model (the preset size of the model is 640*640 pixels) in proportion, the aspect ratio is maintained, and the image is filled with gray pixels to avoid deformation. Random flipping, multi-image splicing, cropping, color adjustment, and other methods can also be used to increase the diversity of data and improve the generalization ability of the model (this process can be determined by the model itself).

[0149] 2、Backbone network feature extraction, Darknet53 optimized by CrossStagePartial Network (CSPNet) is used as the backbone network, and the calculation redundancy is reduced by cross-stage partial connection, and multi-scale feature maps are output, which correspond to small, medium and large target detection levels respectively.

[0150] 3、Path aggregation feature pyramid (PAFPN) fusion, the deep low-resolution features are up-sampled from top to bottom and fused with shallow high-resolution features to transfer semantics; the shallow detail features are down-sampled from bottom to top and spliced with deep features to enhance spatial perception ability; the channel numbers are aligned by 1*1 convolution in the horizontal direction to preserve the original feature information.

[0151] 4、Decoupling detection head prediction results, different branches are used to predict target position and category information respectively, the classification branch is used to predict category information, the regression branch is used to output the position information of bbox (bounding box), and the IoU branch is added in the regression branch, and the two parts are predicted separately, which can improve the performance of the model.

[0152] 5、Dynamic positive sample assignment, based on SimOTA, the positive samples (predicted boxes) corresponding to each real box are dynamically assigned, the matching cost is calculated by combining the classification score, IoU and center point distance, and the cost matrix is constructed, and the expression formula of the cost matrix is:

[0153] Cost i，j =λ cls ·(1-p i，j )+λ IoU ·(1-IoU i，j )+λ center ·d center (i,j);

[0154] In the formula, p i,j represents the classification probability of the predicted box i to the real box j; Cost i,j represents the cost matrix; IoU i,j represents the intersection over union of the predicted box and the real box; d center represents the normalized Euclidean distance of the center point; λ cls represents the weight coefficient of classification; λ IoU represents the weight coefficient of IoU; λ center represents the weight coefficient of center distance.

[0155] The number of positive samples for each real box is dynamically determined, the first m candidate predicted boxes with the minimum cost are selected for each real box, the mean and standard deviation of the IoU of these candidate predicted boxes are calculated, and the expression formula of dynamic positive sample assignment is:

[0156]

[0157] IoU = max(0, min(IoU, 1)) i μ = 1 / m å IoU IoU σ = å(IoU - μ)2 / (m - 1) IoU k = σ -1 j For each real box, the first k optimal candidate prediction boxes with the minimum cost are selected as positive samples, and the calculation formula of the number of positive samples is: j

[0158]

[0159] wherein, a represents an adjustment factor; ∈ represents a minimum value, preventing division by zero.

[0160] 6、Loss function calculation, the loss function of this model is composed of three parts: classification loss (Classification Loss), regression loss (Bounding Box Regression Loss) and target existence probability loss (ObjectnessLoss), and the expression formula of the loss function calculation formula is:

[0161]

[0162] wherein, λ cls , λ reg , λ obj represent weight coefficients, used to balance the contribution of each part (by default λ cls = 1, λ reg = 5, λ obj = 1); N pos represents the total number of positive samples; N all represents the total number of all prediction boxes; p i represents the prediction probability; y i represents the real label (y i = 1 represents a positive sample, and y i = 0 represents a negative sample).

[0163] S103、Based on the final image and the initial detection result, and combined with the threshold filtering and non-maximum suppression algorithm, the final detection result is obtained.

[0164] In this optional embodiment, based on the final image and the initial detection result, and combined with the threshold filtering and non-maximum suppression algorithm, the final detection result includes the following steps:

[0165] Based on the initial detection result, a preliminary prediction box is obtained by using a confidence threshold filtering.​

[0166] The preliminary prediction frame is merged by using a non-maximum suppression algorithm to obtain a detection frame.

[0167] Based on the detection frame, the detection frame is verified by using a ResNet model to obtain a final detection frame.

[0168] The final detection frame is superimposed into the final image to obtain a final detection result.

[0169] It needs to be explained that the detection result outputs a bounding box (Bounding Box, Bbox) pixel coordinate with a confidence, and the detected component is cropped according to the final Bbox coordinate, realizing seamless connection from the detection result to the OCR (Optical Character Recognition) input, and providing standardized image data for subsequent model identification.

[0170] The prediction frame is filtered and de-duplicated. First, the prediction frame with a classification score lower than 0.5 is filtered out by a confidence threshold (0.5), and then the non-maximum suppression (NMS) algorithm is used to delete the frame with an IoU higher than the IoU threshold with the highest score frame. The expression formula of non-maximum suppression is:

[0171] IoU(B i ,B max )≤θ NMS ;

[0172] In the formula, B i represents any prediction frame; B max represents the prediction frame with the highest classification score; and θ NMS represents a threshold value (default is 0.5, adjustable).

[0173] Finally, the detection results of different block images are merged to obtain the optimal detection frame of all target components.

[0174] For each detection result, a fine-tuned ResNet model is used for secondary classification verification, and the prediction results of the two models are compared. Only when the categories are consistent and the consistency score reaches the threshold, the verification is considered to be passed, effectively reducing the false detection situation. The specific steps are as follows:

[0175] 1. Extract the RoI (Region of Interest) for each detection result. Extract the bounding box coordinates of each detection target from the YOLOx detection result, and crop the corresponding region of interest according to the coordinates.

[0176] 2. Verify the RoI using the classification model. First, perform image preprocessing operations such as scaling and normalization on the RoI to ensure that all RoIs meet the model input requirements, i.e., ensure that all RoI sizes are uniform at 224*224 and that the color is 3 channels and the order is RGB. Then, input the RoI into the fine-tuned ResNet model to obtain the classification result and confidence.

[0177] 3. Compare the two classification results. Compare whether the class predicted by YOLOx and the class predicted by the classification model are the same, and combine the confidence of the two model results to calculate the final consistency score. The calculation formula is: consistency score = YOLOx confidence x ResNet confidence.

[0178] Only when both models are confident in their predictions will the final consistency score be high. If the confidence of either model is low, the consistency score will be low, reducing the risk of false detection.

[0179] 4. Output the final detection result. If the consistency score is greater than the set threshold, the target detection result passes the verification, and the detection result, bounding box, and consistency score are output.

[0180] Overlay the detection box on the final image and perform fusion. Support RGB / CMYK multi-color mode and generate IPEG / PNG images with annotations. At the same time, based on the optimal Bbox coordinates, extract the RoI (Region of Interest) of the component from the original image, and automatically adjust the component image according to the Bbox on the original image.

[0181] S104. Use an OCR recognition algorithm to recognize each component image in the final detection result to obtain a recognition result.

[0182] In this optional embodiment, using an OCR recognition algorithm to recognize each component image in the final detection result to obtain a recognition result includes the following steps:

[0183] Preprocess each component image in the final detection result to obtain a plurality of standard detection images;

[0184] Use an OCR recognition algorithm to perform OCR recognition on the plurality of standard detection images to obtain a plurality of initial recognition results;

[0185] Filter the plurality of initial recognition results based on a preset text length threshold to obtain a plurality of standard recognition results;

[0186] According to the text similarity and confidence factors, screen the plurality of standard recognition results to obtain a recognition result.

[0187] In this optional embodiment, the pre-processing of each component image in the final detection result obtains a plurality of standard detection images, including the following steps:

[0188] The size screening of each component image in the final detection result obtains a first detection image;

[0189] Based on the predefined rotation matrix, the angle transformation of the first detection image obtains a second detection image at different angles;

[0190] The vertical stacking of the second detection image at different angles is performed by using the vertical stacking function to obtain a plurality of standard detection images.

[0191] It needs to be explained that the OCR recognition algorithm recognition step is:

[0192] 1. Size check is performed on the input component image to determine whether it meets the minimum size requirement (ocrminimum size = 81). If the image size is too small, the text extraction is not considered, and the image that does not meet the requirement is screened out in advance to avoid invalid calculation.

[0193] 2. Four angle images of 0°, 90°, 180° and 270° are generated to comprehensively cover all possible directions of component text, so that the surface text of the component can be recognized in a positive form regardless of the angle of the component in the captured image. The cv2.rotate function of OpenCV is used to transform the image according to the predefined rotation matrix to generate images at different angles. The expression formula of the rotation matrix is:

[0194]

[0195] 3. The np.vstack function is used to vertically stack the images of 0° and 180°, and 90° and 270° to reduce the number of images to be processed, reduce the calculation amount, and improve the recognition efficiency without losing text recognition information.

[0196] 4. For each angle image, the ocr method of PaddleOCR is called for single-image OCR recognition. The input image is subjected to text detection and recognition at the same time, and the classifier is disabled during the recognition process, focusing only on text content and position recognition. After the recognition is completed, the results containing text content and confidence are returned.

[0197] 5. The recognition results are filtered according to the set text length threshold (length threshold = 2) to screen out short texts and improve the accuracy of the final results

[0198] 6. Merge and select the best results. Taking into account factors such as text similarity (text similarity threshold default 0.5) and confidence (confidence threshold default 0.6), the system merges similar results, processes similar text, and sorts the results in descending order of confidence to select the most reliable component text recognition content from the multi-angle recognition results.

[0199] like Figure 3 As shown, the post-processing stage workflow is as follows: For newly recognized text, a confidence level is determined. If the confidence level of the newly recognized text is ≥0.6, it is further determined whether similar text exists; otherwise, the newly recognized text is discarded. For newly recognized text, a similar text is determined. If similar text exists, the confidence level of the newly recognized text is further determined; if no similar text exists, the newly recognized text is added as a new result. For newly recognized text, a confidence level is determined again. If the confidence level of the newly recognized text is ≤0.5, the longer text is selected as the text recognition content; otherwise, the high-confidence text is selected as the text recognition content.

[0200] S105. Construct a component database and combine it with a similarity algorithm to match the identification results to obtain matching results; based on the matching results, use a time-series prediction model and a graph neural network model to obtain evaluation results.

[0201] In this optional embodiment, a component database is constructed, and a similarity algorithm is used to match the identification results to obtain matching results. Based on the matching results, an evaluation result is obtained using a time-series prediction model and a graph neural network model, including the following steps:

[0202] Build a component database;

[0203] Based on the similarity algorithm, the similarity between the recognition result and the component model in the component database is calculated, and the matching result is obtained by filtering.

[0204] Using a time-series prediction model, the matching results are predicted in time series to obtain the prediction results;

[0205] Based on the graph neural network model and combined with the prediction results, the matching results are evaluated to obtain the evaluation results.

[0206] It needs to be explained that in the field of fine recycling of waste circuit boards, in order to achieve the goal of maximizing resource value, a real-time price database of components based on dynamic market conditions needs to be built. Through dynamic tracking of market conditions, the real-time recycling value of components in waste circuit boards is accurately assessed, avoiding resource waste caused by price lag or information asymmetry. Through multi-source data collection and machine learning algorithm integration of real-time quotes, supply and demand fluctuations, and material composition data of global electronic component trading platforms, combined with component intelligent identification and OCR technology for accurate identification and classification of circuit board components, dynamic optimization of disassembly strategy and matching of the highest value recycling path, while supporting historical price trend analysis and future value prediction, ultimately realizing multi-dimensional evaluation of recycling benefits through data visualization tools, thus promoting the efficient development of electronic waste recycling economy.

[0207] Building a multi-dimensional dynamic database of components includes the following steps:

[0208] 1. Multi-source data collection: Through API interface, web crawler, data cooperation and other compliance methods, collect electronic component transaction quotes, supply and demand fluctuations, material composition and other data from trading platforms, industry information websites, social media, metal trading markets and other channels, and use distributed node deployment and real-time stream processing technology to ensure the timeliness and comprehensiveness of the data.

[0209] 2. Data cleaning and preprocessing: Clean up transaction quotes and other data through missing value filling, outlier removal, and duplicate value removal operations, and use format unification, text vectorization and other methods to standardize different types of data, laying a foundation for subsequent analysis.

[0210] Establish a stable connection with the component recycling price database to ensure quick and accurate access to data in the database, support multiple database types; calculate the similarity between OCR recognized text and component model in the database, filter out possible matching results, sort matching results by similarity, retain top 3 most likely results, filter low-quality matches through similarity threshold and confidence threshold to ensure result reliability. The specific matching process is as follows:

[0211] 1. Traverse the components and check if each component has OCR text.

[0212] 2. Traverse the OCR results, compare each OCR result (including confidence) with all component models in the database one by one.

[0213] 3. Calculate similarity, use character similarity to check the similarity of two strings, first standardize the text, then use Levenshtein distance to calculate similarity and comprehensive confidence, the expression formula of similarity and comprehensive confidence is:

[0214]

[0215] Comprehensive confidence = similarity * OCR confidence.

[0216] 4. Threshold filtering, only keeping results with similarity and comprehensive confidence exceeding threshold, where similarity threshold = 0.8, comprehensive confidence threshold = 0.7.

[0217] 5. Result sorting, ranking by similarity in descending order, taking top 3 with highest similarity to store in component database matching results.

[0218] In the process of value analysis and prediction of waste circuit boards, the time series prediction model Prophet algorithm is selected to capture the time series rules of the multi-dimensional prices of components and analyze the trends of different price variables over time. The price trends and prediction values obtained by the Prophet algorithm are used as new features, combined with multi-source feature data such as transactions, supply and demand, and materials, and then input into the Random Forest (RF) ensemble learning algorithm. The Random Forest integrates multi-dimensional information to output the future price trend prediction value of components.

[0219] The historical data of component prices y at different times t are input into the Prophet algorithm, which automatically fits the trend, seasonality, and holiday model parameters. By adjusting the assignment of the core smoothing parameter change point prior scale (default change point prior scale = 0.05) in the algorithm, the trend change speed is controlled. The smaller the value (0.01-0.1), the smoother the trend, and the larger the value (0.5-1), the more sensitive the model to changes, and the more tortuous the trend. Finally, the trend and prediction value of the price in the future time interval are output. The expression formula of the Prophet algorithm is:

[0220] y(t) = g(t) + s(t) + h(t) + τ;

[0221] In the formula, y(t) represents the time series of component prices; g(t) represents the long-term trend of prices; s(t) represents periodic fluctuations; h(t) represents the influence of special events such as holidays; and τ represents the error term, i.e., unexplained random fluctuations.

[0222] Random Forest is composed of multiple decision trees, which are trained on price trends, prediction values, and multi-source feature data such as transactions, supply and demand, and materials. Each decision tree is trained by randomly sampling part of the data and features from the original data. After training, the prediction results of the sample are obtained by inputting the prediction sample into the trained decision tree. The prediction results of multiple decision trees are combined to obtain the final prediction result of the sample. The average is taken to obtain the final prediction value, and the expression formula of the prediction formula is:

[0223]

[0224] In the formula, U represents the number of decision trees.

[0225] A graph neural network model (GNN) is constructed, regarding components on the circuit board as nodes and regarding model, price and other information as feature vectors of the nodes; regarding transaction relationship, physical connection, functional dependence and the like as edges, a correlation graph structure is constructed, and a feature vector of an xth layer node a is calculated; through a message passing mechanism of the graph neural network, the node exchanges feature information with neighbor nodes, the feature of a final layer node is summed, and is mapped to a waste circuit board value evaluation result through a fully connected layer. The calculation expression formula of the feature vector of the xth layer node a is:

[0226]

[0227] In the formula, represents the feature vector of node a at the x+1th layer; N(a) represents a neighbor node set of a; c am represents a normalization constant for balancing the number of neighbors of different nodes; W (x) represents a weight matrix of the xth layer; b (x) represents a bias vector of the xth layer; ReLU represents an activation function.

[0228] S106, integrate the final detection result, the recognition result and the evaluation result, and perform visual output.

[0229] It needs to be explained that the original image with labeled results, the surface model information recognized by OCR and the matching result with the database are integrated, the output result is customized according to the user's needs, the content and format of the output are selected, and the user's experience is improved.

[0230] The component detection result image is a picture of the waste circuit board after a series of processing and recognition, which visually marks the positions of the detected components on the picture, and directly displays the distribution of the components through a marked box with a distinctive color, and supports output in multiple common picture formats, so that the user can select a suitable picture format for saving and sharing according to the actual situation.

[0231] The JSON file of the detection result has good structure and readability, and can conveniently record detailed information of each component, such as the position coordinates, model, matching confidence and the like of the component, and with the continuous improvement of the system function and the change of the business demand, the content of the JSON file can be conveniently adjusted and expanded.

[0232] The OCR recognition matching result is output clearly and accurately, and the component surface model string recognized by the OCR recognition can be directly displayed on the system interface in the form of text or saved as a text file for the user to view and use.

[0233] Figure 2 An embodiment of the waste circuit board high-value component intelligent identification system of the application is shown.

[0234] In this optional embodiment, a waste circuit board high-value component intelligent identification system comprises:

[0235] The acquisition and preprocessing module 201 is configured to acquire an original image and pre-process the original image to obtain a final image.

[0236] The detection module 202 is configured to construct and train a target detection model, detect the final image by the target detection model, and obtain an initial detection result.

[0237] The post-processing module 203 is configured to obtain a final detection result based on the final image and the initial detection result, and in combination with threshold filtering and non-maximum suppression algorithm.

[0238] The recognition module 204 is configured to utilize an OCR recognition algorithm to recognize each component image in the final detection result to obtain a recognition result.

[0239] The evaluation module 205 is configured to construct a component database, combine a similarity algorithm to match the recognition result, and obtain a matching result; and based on the matching result, utilize a time series prediction model and a graph neural network model to obtain an evaluation result.

[0240] The output module 206 is configured to integrate the detection result, the recognition result and the evaluation result, and perform visual output.

[0241] It should be explained that a hierarchical architecture design is adopted, which comprises an infrastructure layer, a data processing layer, a core business layer, a service gateway layer and an interaction interface layer, each layer cooperates through standardized protocols to form a modular and scalable technical system.

[0242] Infrastructure layer: the server running environment is built based on the Windows operating system, the structured data is stored by using the MySQL relational database, and the management of massive image data is realized by combining the distributed file system to ensure the stability of the system.

[0243] Data processing layer: a high-efficiency data management mechanism is constructed, which covers data reading and writing, caching and backup function modules, supports transaction processing, guarantees the consistency and integrity of data operation, and is optimized for batch operation to improve the overall data processing efficiency.

[0244] Core business layer: Through the image processing module integrated with the pre-processing algorithm, the image optimization is realized, relying on the intelligent recognition module to fuse feature extraction, pattern matching and deep learning technology to complete component detection, based on the resource management module to realize intelligent matching of component models, and to provide multi-format result export and automatic report generation for the output control module, forming a complete business closed loop covering image acquisition, processing, recognition to output.

[0245] Interactive interface layer: Design responsive web and multi-platform APP, provide visual operation interface and real-time interactive dashboard, provide SDK access scheme for industrial cameras, scanners and other peripherals.

[0246] The recognition system is based on machine vision and deep learning technology, and integrates target detection, OCR character recognition, database matching and other core technologies for electronic waste disassembly scene, realizes the whole process intelligent processing from waste circuit board image acquisition, component detection, model recognition to database matching, breaks the low efficiency mode of traditional manual disassembly. It has high precision recognition and analysis ability, through the optimized target detection algorithm (YOLO series) and OCR technology, the position, type and model of components on the circuit board are recognized with high precision, ensuring that the detection positioning error is controllable and the model recognition accuracy is high, providing reliable data support for subsequent disassembly and recycling. Support multiple image acquisition devices, database interfaces, have flexible input and output compatibility, at the same time, the detection results are standardized (generate standard JSON files and result annotation images), which is convenient for docking with other management systems or disassembly equipment, and adapts to different application scenarios. Covering the complete link of "image acquisition-preprocessing-detection-postprocessing-OCR-model matching-result evaluation-result output", forming a closed loop from data acquisition to decision output, without manual intervention, the intelligent analysis of waste circuit boards can be completed, and the processing efficiency of electronic waste is effectively improved.

[0247] The present application comprises the following interfaces for interfacing with third-party systems:

[0248] 1) Image acquisition device interface

[0249] The interface connects to various image acquisition devices, such as industrial cameras, scanners, etc., including communication of hardware devices and capture cards.

[0250] The interface functions include data transmission, device control and state monitoring.

[0251] Data transmission is to obtain the original waste circuit board image data from the image acquisition device and transmit it to the image acquisition module of the software system. The interface supports multiple image formats and resolutions to ensure that it can adapt to the output of different devices.

[0252] Device control is to send control instructions to image acquisition devices, such as adjusting exposure time, acquisition speed, shooting angle and other parameters, to obtain high-quality images, and also to realize the start, stop, reset and other operations of the device.

[0253] State monitoring is to monitor the working state of image acquisition devices in real time, such as whether the device is running normally, whether there are errors, etc., and feedback the state information to the software system for timely processing.

[0254] 2) Database interface:

[0255] Through the interface, the third-party component information database, price database, etc. are connected, and according to the model of the component identified by the software system, a query request is sent to the third-party database to obtain the price information of the component.

[0256] The database interface connects to the third-party component information database, price database, etc.

[0257] The interface functions include data query, data update and data security.

[0258] Data query is to send a query request to the third-party database according to the component model identified by the software system, and obtain detailed information of the component, such as specification parameters, manufacturers, market prices, etc. The interface has efficient query algorithm to ensure that the query result can be returned in a short time.

[0259] Data update is that the interface supports the synchronization of new component information or price collected by the software system to the third-party database, realizes real-time update of data, and also can obtain the latest data from the third-party database, to ensure the accuracy and timeliness of the data in the software system.

[0260] Data security is to protect the data by encryption technology during data transmission, to ensure the security and confidentiality of the data. At the same time, identity verification and authorization management are also needed to prevent illegal access and data leakage

[0261] 3) Automatic disassembly device interface:

[0262] Through the interface, the waste circuit board automatic disassembly device, such as mechanical arm, cutting machine, etc. is connected, and the position information is transmitted and the working state of the disassembly device is monitored in real time.

[0263] The interface connects to the waste circuit board automatic disassembly device, such as mechanical arm, disassembly robot, etc.

[0264] The interface functions include instruction sending, state feedback and state feedback.

[0265] The instruction sending is sending disassembly instructions such as the disassembly sequence, disassembly force, and position of the components to the automated disassembly equipment according to the results of identification and analysis in the software system. The interface supports multiple communication protocols to ensure communication with different types of disassembly equipment.

[0266] The state feedback is real-time acquisition of the working state of the automated disassembly equipment.

[0267] The state feedback is the cooperative work with the automated disassembly equipment to realize the synchronization and coordination between the software system and the equipment.

[0268] An operation example of the intelligent recognition system for high-value components of waste circuit boards is as follows:

[0269] The operation environment configuration is as follows:

[0270]

[0271] After the user logs in, the main operation panel is entered, which contains operation buttons such as "manual mode", "automatic mode", "connect camera", "task start", and "empty data". In addition, it also contains a result image display area, an output log display area, and a result analysis area.

[0272] The first step of software use needs to determine the software use mode. If the automatic mode is selected, click the "automatic mode" button to select. After selecting the automatic mode, the camera is automatically connected, the background algorithm is ready, and when the material is ready, the collection result can be obtained by triggering the collection signal. If the manual mode is selected, the "manual mode" button needs to be clicked first. After "output log" displays "system initialization success, manual mode selected", click the "connect camera" button. After "output log" displays "camera has been successfully connected, algorithm running …", when the material is ready, click the "task start" button to complete the collection task and output the algorithm analysis result.

[0273] In addition to the content displayed on the software interface, detailed data will be saved in the form of a web page. The content mainly includes recognition accuracy statistics, running time statistics, OCR and database matching result analysis, and other information such as the list of recognized components.

[0274] In one embodiment, a computer device, which can be a server, is provided, and an internal structure diagram thereof can be as shown in Figure 3As shown in the figure. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium, an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to realize the steps in the above method embodiments.

[0275] Those skilled in the art can understand that, Figure 4 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0276] In addition, the present application also provides a computer device comprising a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps in the above method embodiments.

[0277] In addition, the present application also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by the processor to realize the steps in the above method embodiments.

[0278] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0279] The present application is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is only limited by the appended claims.

Claims

1. A method for intelligent identification of high-value components on waste circuit boards, characterized in that, The method includes: The original image is acquired and preprocessed to obtain the final image. Build and train an object detection model, and use the object detection model to detect the final image to obtain initial detection results; Based on the final image and the initial detection results, and combined with threshold filtering and non-maximum suppression algorithms, the final detection results are obtained; Using an OCR recognition algorithm, each component image in the final detection result is identified to obtain the recognition result; A component database is constructed, and the identification results are matched using a similarity algorithm to obtain matching results. Based on the matching results, an evaluation result is obtained using a time-series prediction model and a graph neural network model. Integrate the final detection results, identification results, and evaluation results, and output them in a visual format.

2. The intelligent identification method for high-value components on waste circuit boards according to claim 1, characterized in that, The process of acquiring the original image and preprocessing it to obtain the final image includes the following steps: The original image is acquired, and a lossy compression algorithm is used to compress the original image to obtain the first image; The first image is segmented using a threshold segmentation algorithm to obtain the second image; Based on a preset size, the second image is cropped into blocks to obtain the third image; The third image is color-optimized using a histogram equalization algorithm to obtain the final image.

3. The intelligent identification method for high-value components on waste circuit boards according to claim 1, characterized in that, The process of constructing and training the object detection model, and then using the object detection model to detect the final image to obtain initial detection results, includes the following steps: The target detection model is constructed using Darknet53, which is partially optimized across stages, as the backbone network. The target detection model is trained using a dynamic positive sample allocation algorithm combined with a loss function algorithm. Based on the final image, and combined with the object detection model, the initial detection results are obtained.

4. The intelligent identification method for high-value components on waste circuit boards according to claim 3, characterized in that, The process of training the target detection model using a dynamic positive sample allocation algorithm combined with a loss function algorithm includes the following steps: Acquire training images and preprocess them to obtain standard images; The backbone network of the object detection model is used to extract standard images to obtain feature maps; The neck network based on the object detection model performs feature fusion on the feature map to obtain the feature fusion map; By using a decoupled detection head and combining a dynamic positive sample allocation algorithm and a loss function algorithm, the parameters of the target detection model are optimized.

5. The intelligent identification method for high-value components on waste circuit boards according to claim 4, characterized in that, The optimization of the target detection model parameters using a decoupled detection head, combined with a dynamic positive sample allocation algorithm and a loss function algorithm, includes the following steps: Using a decoupled detection head, the location and category of the feature fusion map are predicted to obtain the classification score and regression box parameters; Based on the dynamic positive sample allocation algorithm, the classification score and regression box parameters are dynamically allocated to obtain positive and negative sample labels and matching relationships; The object detection model is optimized by using classification scores, regression box parameters, dynamic allocation of classification scores and regression box parameters, and loss function algorithms.

6. The intelligent identification method for high-value components on waste circuit boards according to claim 1, characterized in that, The process of obtaining the final detection result based on the final image and the initial detection result, combined with threshold filtering and non-maximum suppression algorithms, includes the following steps: Based on the initial detection results, a preliminary prediction box is obtained by filtering using a confidence threshold. The nonmaximum suppression algorithm is used to merge the preliminary predicted boxes to obtain the detection boxes; Based on the detection boxes, the ResNet model is used to verify the detection boxes to obtain the final detection boxes; The final detection bounding box is superimposed onto the final image, and the result is obtained by fusion.

7. The intelligent identification method for high-value components on waste circuit boards according to claim 1, characterized in that, The process of using an OCR recognition algorithm to identify each component image in the final detection result and obtaining the recognition result includes the following steps: Each component image in the final detection result is preprocessed to obtain several standard detection images; Using an OCR recognition algorithm, several standard detection images were subjected to OCR recognition to obtain several initial recognition results. Based on a preset text length threshold, several initial recognition results are filtered to obtain several standard recognition results; Based on text similarity and confidence factors, the recognition results are filtered from several standard recognition results to obtain the recognition results.

8. The intelligent identification method for high-value components on waste circuit boards according to claim 7, characterized in that, The preprocessing of each component image in the final detection result to obtain several standard detection images includes the following steps: The size of each component image in the final detection result is filtered to obtain the first detection image; Based on a predefined rotation matrix, the first detection image is subjected to angle transformation to obtain second detection images at different angles; By using a vertical stacking function, second detection images at different angles are vertically stacked to obtain several standard detection images.

9. The intelligent identification method for high-value components on waste circuit boards according to claim 1, characterized in that, The process of constructing a component database and matching the identification results using a similarity algorithm to obtain matching results, followed by the evaluation results obtained using a time-series prediction model and a graph neural network model, includes the following steps: Build a component database; Based on the similarity algorithm, the similarity between the recognition result and the component model in the component database is calculated, and the matching result is obtained by filtering. Using a time-series prediction model, the matching results are predicted in time series to obtain the prediction results; Based on the graph neural network model and combined with the prediction results, the matching results are evaluated to obtain the evaluation results.

10. A smart identification system for high-value components on waste circuit boards, characterized in that, The system includes: The acquisition and preprocessing module is used to acquire the original image and preprocess the original image to obtain the final image; The detection module is used to build and train an object detection model, and then use the object detection model to detect the final image to obtain initial detection results. The post-processing module is used to obtain the final detection result based on the final image and the initial detection result, combined with threshold filtering and non-maximum suppression algorithms; The recognition module is used to identify each component image in the final detection result using an OCR recognition algorithm to obtain the recognition result; The evaluation module is used to build a component database and, in conjunction with a similarity algorithm, match the identification results to obtain matching results; based on the matching results, an evaluation result is obtained using a time-series prediction model and a graph neural network model. The output module integrates the final detection results, recognition results, and evaluation results for visualization.

Images