Method and system for recycling and approving waste electric bicycles
By using a dual-channel detection model and large-scale model technology, the problems of low efficiency, large errors, and poor robustness in the traditional manual approval process have been solved, achieving efficient, accurate, and intelligent approval of used electric bicycles.
Patent Information
- Application Number
- CN202510995892.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-10-31
AI Technical Summary
The traditional approval process for recycling used electric bicycles relies on manual verification, which suffers from low efficiency, large subjective errors in classification, rigid thinking, and poor robustness, and cannot complete multiple tasks simultaneously.
Employing a dual-channel detection model and large-scale model technology, it receives standardized views and vehicle information, identifies brand information, outputs damage scores, and combines these with a decision-making model for intelligent approval.
It has enabled efficient and accurate classification, identification, and intelligent approval of waste electric bicycles, improving approval efficiency and ensuring the objectivity of approval results.
Smart Images
Figure CN120875858A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of computer vision and circular economy technology, and in particular to a method and system for approving the recycling of used electric bicycles. Background Technology
[0002] Currently, with the strong support and promotion of the national trade-in policy, the corresponding application and approval system is also facing efficiency bottlenecks.
[0003] Traditional approval processes for the recycling of used electric bicycles rely on manual verification. In practice, this involves manual review of uploaded images to identify, classify, and assess the damage of the bicycles, then combining this information with other data reported by the user to provide an approval opinion. However, this manual approval method exposes several problems, primarily in the following aspects: Excessive reliance on manual labor leads to low processing efficiency, especially when dealing with massive volumes of approvals, where the inefficiency and error-proneness of manual operation become significant bottlenecks. Large subjective errors exist in classification; for manual approvals, aspects such as damage assessment of used electric bicycles are highly subjective, potentially affecting the accuracy of the final approval opinion. The method of providing approval opinions manually can be prone to rigid thinking, failing to accurately assess certain characteristics of recycled electric bicycles.
[0004] Therefore, manual approval processes suffer from three major bottlenecks: low efficiency, significant subjective errors in classification, and rigid policy adaptation. However, for approval scenarios involving the recycling and scrapping of abandoned transportation vehicles, especially used electric bicycles, the large volume of approvals and relatively fixed types of related documents make it entirely possible to automate the approval process by incorporating machine learning technology, thereby improving efficiency.
[0005] However, most current target classification and detection algorithms can only be applied to a single classification task and cannot simultaneously complete multiple tasks in the waste electric bicycle recycling business scenario. They are not closely integrated with the old-for-new business scenario. For example, they cannot simultaneously complete tasks such as damage assessment. The robustness of the algorithm may be poor, and improvements need to be made based on the current algorithm. Summary of the Invention
[0006] This application provides a method and system for approving the recycling of used electric bicycles, which is used to solve at least one of the above-mentioned technical problems.
[0007] The technical solution adopted in this application is as follows:
[0008] On the one hand, this application provides a method for approving the recycling of used electric bicycles. The method includes: receiving a standardized view and vehicle information uploaded by a user for the used electric bicycle; inputting the standardized view into a dual-channel detection model to identify the brand information of the used electric bicycle and output a damage score of the used electric bicycle; encapsulating the vehicle information, the brand information, and the damage score into a preset data format and inputting it into a decision model to obtain the recycling approval result corresponding to the used electric bicycle. The decision model uses subsidy rules as a knowledge base to participate in the decision-making process.
[0009] In one possible implementation of this application, after receiving a standardized view uploaded by a user for a discarded electric bicycle, the method further includes: verifying the standardized view, specifically, determining whether the proportion of the area corresponding to the brand nameplate of the discarded electric bicycle in the front image of the standardized view meets a first minimum proportion constraint, and determining whether the proportion of the area corresponding to the body of the discarded electric bicycle in the side and back images of the standardized view meets a second minimum proportion constraint.
[0010] In one possible implementation of this application, the dual-channel detection model includes a brand classification network and a damage detection network; the brand classification network is used to identify the frontal image in the input standardized view and output the brand information and confidence result corresponding to the waste electric bicycle; the damage detection network is used to identify the frontal image, side image and back image in the input standardized view and output the damage score corresponding to the waste electric bicycle.
[0011] In one possible implementation of this application, the damage detection network processes the frontal image, side image, and back image respectively to obtain damage scores corresponding to the frontal image, side image, and back image respectively, and determines the damage score by weighted averaging of the damage scores.
[0012] In one possible implementation of this application, the dual-channel detection model includes a brand classification network and a damage detection network. The training process of the brand classification network and / or the damage detection network includes: acquiring a frontal image of an electric bicycle as training data for the brand classification network; acquiring a frontal image, a side image, and a rear image of an electric bicycle as training data for the damage detection network; the side image includes images of the left and right sides of the electric bicycle; labeling the training data; inputting the labeled training data into the network to be trained in batches for network training; calculating the loss function value according to the loss function definition of the network to be trained; using the Adam optimizer to perform backpropagation based on the loss function value to update the network parameters of the network to be trained; performing a network evaluation after each preset number of training iterations; and ending network training when the accuracy difference between two network evaluations is less than 1%.
[0013] In one possible implementation of this application, the network to be trained includes a Backbone network module, a Neck network module, and convolutional layers: the Backbone network module adopts a CSP structure for feature extraction; the Neck network module adopts an SPPF structure and a PAN structure for feature fusion; and the convolutional layers are used to output detection results.
[0014] In one possible implementation of this application, in the network to be trained, the CSP structure is used to divide the input feature map into two parts, one part is processed by a convolutional sub-network, and the other part is directly fed into the next layer for processing. Then, the two parts of the feature map are concatenated as the input of the next layer. The SPPF structure is used to perform pooling operations at different scales on the input feature map and introduce a feature fusion module to concatenate the pooling results at different scales as the output of the SPPF structure. The PAN structure is used to perform multi-level feature fusion.
[0015] In one possible implementation of this application, the dual-channel detection model uses a cross-entropy loss function and a leaky ReLU activation function during training.
[0016] In one possible implementation of this application, the recycling approval result corresponding to the waste electric bicycle obtained includes at least one of the following: specific amount, disposal requirements and policy basis.
[0017] On the other hand, this application also provides a recycling approval system for used electric bicycles. The system includes: an information receiving module that receives standardized views and vehicle information uploaded by users for used electric bicycles; a vehicle detection module that inputs the standardized views into a dual-channel detection model to identify the brand information of the used electric bicycle and output a damage score for the used electric bicycle; and an approval decision module that encapsulates the vehicle information, the brand information, and the damage score into a preset data format and inputs it into a decision model to obtain the recycling approval result corresponding to the used electric bicycle. The decision model uses subsidy rules as a knowledge base to participate in the decision-making process.
[0018] The recycling approval method and system for used electric bicycles provided in this application have the following beneficial effects:
[0019] The proposed method for approving the recycling of used electric bicycles combines an improved target detection and classification algorithm with large model technology. It can accurately extract image features and, in the business scenario of recycling and scrapping used electric bicycles, accurately and efficiently classify and identify used electric bicycles, including but not limited to specific features such as brand. Furthermore, it can utilize large model technology to input the extracted image features for real-time and efficient intelligent approval. This not only improves the robustness of the intelligent approval algorithm but also enhances the efficiency of intelligent approval and ensures the objectivity of the approval results. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:
[0021] Figure 1 A flowchart illustrating the approval process for the recycling of used electric bicycles provided in this application;
[0022] Figure 2 This is a schematic diagram of the brand classification network in Example 1;
[0023] Figure 3 This is a schematic diagram of the damage detection network in Example 1;
[0024] Figure 4 This is a flowchart of Example 2;
[0025] Figure 5 This application provides an architecture diagram of a recycling approval system for used electric bicycles. Detailed Implementation
[0026] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.
[0027] In practical scenarios of waste electric bicycle recycling approval, existing technologies for waste electric bicycle recycling and scrapping approval have exposed multiple problems, mainly in the following aspects: ① Excessive reliance on manual labor: Approval requires a large amount of human intervention, leading to low business processing efficiency. Especially when facing massive approval tasks, the inefficiency and error-proneness of manual operation become significant bottlenecks. ② Large subjective errors in classification: For manual approval, the determination of damage to waste electric bicycles and other aspects are highly subjective, which may affect the accuracy of the final approval opinion. ③ The way manual approval opinions are given may suffer from rigid thinking, making it impossible to accurately judge some characteristics of waste electric bicycles to be recycled. ④ Most existing target detection algorithms cannot simultaneously complete multiple tasks in waste electric bicycle recycling scenarios.
[0028] The above problems stem from multiple limitations in the design philosophy and technical implementation of existing technologies: ① The current approval system for the recycling of used electric vehicles has only just begun large-scale deployment with policy implementation and has not yet been integrated with machine learning algorithms and large-scale model technologies. ② Integrating artificial intelligence technologies may increase the cost of the approval system. ③ Currently, manual approval may still meet most business needs, but for large-scale approval tasks at the provincial level, manual approval will face inefficiency issues. ④ Current object detection algorithms do not consider object recognition and classification issues specific to the business scenario of recycling used electric vehicles, and the robustness of the algorithms may be poor.
[0029] This application provides a method and system for approving the disposal of used electric bicycles. Based on a target detection and classification algorithm, it identifies and classifies abandoned vehicles by detecting their features. Based on large model technology, it combines the extracted features with other information to provide approval opinions. This enables the identification and classification of used electric bicycles in business scenarios such as trade-in and scrapping, and intelligent approval based on the identification and classification results combined with the model.
[0030] The method in this application will be described in detail below with reference to the accompanying drawings.
[0031] Figure 1 A flowchart of the approval method for the recycling of used electric bicycles provided in this application is shown below. Figure 1As shown, the approval method for the recycling of used electric bicycles in this application includes at least the following steps:
[0032] Step 101: Receive the standardized view and vehicle information uploaded by the user for the discarded electric bicycle.
[0033] When users want to recycle their old electric bicycles, they need to provide a standardized view of the bicycle and fill in the required vehicle information. In one example, the standardized view includes at least a front, side, and rear image of the bicycle, with the side image referring to both the left and right sides. The vehicle information to be filled in includes the vehicle's age, intended disposal method, and location.
[0034] In one possible implementation of this application, to ensure that the standardized view of the discarded electric bicycle uploaded by the user meets the requirements, i.e., can be used by the subsequent dual-channel detection model, after receiving the standardized view of the discarded electric bicycle uploaded by the user, the standardized view is judged or verified. Specifically, for the front image of the discarded electric bicycle, it is determined whether the area occupied by the electric vehicle nameplate satisfies the first minimum proportion constraint. Preferably, the first minimum proportion constraint is 15%. Only when the proportion of the nameplate area meets the constraint can the subsequent dual-channel detection model accurately identify the vehicle's brand information. It should be noted that the vehicle's brand information is preferably set on the vehicle nameplate. For the side and rear images of the discarded electric bicycle, it is determined whether the proportion of the corresponding area of the vehicle body satisfies the second minimum proportion constraint. Preferably, the second minimum proportion constraint is 70%. Only when the proportion of the vehicle body area meets the constraint can the subsequent dual-channel detection model detect the damage of the discarded electric bicycle based on the vehicle body area. In one example, the damage of the discarded electric bicycle includes at least one of the following: vehicle body structural deformation, battery integrity, and tire wear value.
[0035] Once the standardized view of the abandoned electric bicycle uploaded by the user passes the above judgment or verification, the process of receiving the vehicle information of the abandoned electric bicycle is completed. Subsequent detection processes are all based on the information received in this step.
[0036] Step 102: Input the standardized view into the dual-channel detection model to identify the brand information of the discarded electric bicycles and output the damage score of the discarded electric bicycles.
[0037] The standardized view of the verified discarded electric bicycle is input into a pre-trained dual-channel detection model. The dual-channel detection model is used to identify, detect and classify the standardized view, and output the brand information and damage score of the discarded electric bicycle.
[0038] In one possible implementation of this application, the dual-channel detection model consists of two networks that process and identify standardized views in parallel. Specifically, the two networks are a brand classification network and a damage detection network. The brand classification network is used to identify the frontal image in the input standardized view, specifically identifying the nameplate area in the frontal image to output the brand information and confidence result corresponding to the discarded electric bicycle. The damage detection network is used to identify the frontal, side, and rear images in the input standardized view, specifically identifying the vehicle body area in all standardized views, i.e., the four images, to identify the damage to the discarded electric bicycle. For each image, a corresponding damage score is obtained, and finally, the weighted average of these four damage scores is output, which is the damage score corresponding to the discarded electric bicycle.
[0039] It should be noted that the training process of the above dual-channel detection network can be found in the relevant description below, and will not be repeated here.
[0040] Step 103: After encapsulating the vehicle information, brand information, and damage score into a preset data format, input them into the decision model to obtain the recycling approval result for the waste electric bicycle.
[0041] The brand information and damage scores of the discarded electric bicycles obtained in the above steps, along with the vehicle information filled in by the user, are packaged together into JSON format data and then input into the decision model. Here, the decision model is primarily a large model decision engine, and the subsidy rules of the location of the discarded electric bicycle are used as a knowledge base to participate in the decision-making process. The output includes decision information containing specific amounts, disposal requirements, and policy basis, which is the recycling approval result.
[0042] Example 1
[0043] This embodiment describes the training process of a dual-channel detection model.
[0044] Specifically, before training, training data needs to be constructed, including data collection and data labeling phases. In the data collection phase, images of the front (with brand nameplate) of discarded electric bicycles (hereinafter referred to as electric bikes) are collected for training the brand classification network; images of the front, left and right sides, and rear of the electric bikes are used for training the damage detection network. The collected images cover mainstream electric bike models. Front images are taken at a distance of 1.5 meters at eye level, ensuring that the brand nameplate area accounts for more than 15%, while for side and rear images, the non-vehicle area should be less than 30%. The collected images are then augmented: through affine transformation, preferably rotating ±15°, and adjusting brightness preferably ±30%, fogging is used to simulate and generate enhanced sample training data to ensure the generalization of the training network. The collected image data is then labeled. The labels for the front images include the location information of the brand nameplate and brand classification information; the labels for the damage detection images include the degree of damage to the electric bike. Based on the appearance of the electric bike's components, the degree of damage can be manually divided into nine levels from 1 to 9, with 9 representing the most severe damage.
[0045] After data collection and labeling are completed, training data is obtained and can be used for network training. The network training process mainly includes the following explanations:
[0046] 1) Network Structure
[0047] The dual-channel detection model in this application requires training two networks: a brand classification network and a damage detection network, both trained based on an improved YOLO network architecture, hereinafter collectively referred to as the network to be trained. The main structure of the network to be trained consists of two network modules: Backbone and Neck.
[0048] like Figure 2-3As shown, the Backbone network uses a Cross-Stage Partial (CSP) structure. The CSP structure divides the input feature map into two parts: one part is processed by a small convolutional network (called a sub-network), while the other part is directly processed by the next layer. The two feature maps are then concatenated as the input to the next layer. This process effectively reduces network parameters and computational cost, while improving feature extraction efficiency and ensuring the backbone network maintains strong feature extraction capabilities and computational efficiency. The Neck network mainly includes SPPF and PAN structures. SPPF is a pyramid pooling structure that can pool feature maps of different sizes, thereby enhancing the model's ability to perceive targets at different scales. Specifically, the SPPF structure performs 1x1, 2x2, and 3x3 pooling operations on the input feature map and introduces a feature fusion module to concatenate the pooling results at different scales as the output of the SPPF structure. The PAN structure aims to improve the model's ability to perceive targets at different scales through multi-level feature fusion. These two structures enable the Neck network to perform detection at different feature map levels, which can improve the performance of object detection and fuse information from different feature map levels.
[0049] Both networks ultimately output prediction results through a convolutional layer. The network used for brand classification has an output size of 80×80×(4×number of categories) and is used to detect small targets such as car logos; the network used for damage detection has an output size of 1×1×9 and is used to predict the probability value of each damage level.
[0050] 2) Loss function, activation function, and optimizer
[0051] Loss function: The cross-entropy loss function is used in the network training process. This function can alleviate the problem of class imbalance in object detection and improve the performance of the final trained dual-channel detection model.
[0052] Activation function: The network training process uses the leaky ReLU activation function instead of the ReLU activation function to improve the performance of the final trained dual-channel detection model.
[0053] Optimizer: The Adam optimizer is selected as the optimizer for updating network parameters during network training.
[0054] 3) Training Process
[0055] To enhance the generalization ability of the final trained dual-channel detection model and prevent gradient explosion and vanishing phenomena, batch training is used during the training process. This means that multiple sets of images and label data are input at once, and the model is trained for multiple rounds. The main steps of each training round are as follows:
[0056] Input the training image data into the network to be trained in batches;
[0057] Calculate the value of the loss function according to its definition;
[0058] Backpropagation is performed using the Adam optimizer to update the network parameters of the network to be trained.
[0059] During network training, a network evaluation is performed every 10 training rounds. The network weight parameters are not updated during the network evaluation. The network training ends when the accuracy difference between two evaluations is less than 1%, resulting in a dual-channel detection model.
[0060] Example 2
[0061] This example describes the approval process for the recycling of used electric bicycles, which is enabled by a large model.
[0062] like Figure 4 As shown, in the business process of approving the recycling of used electric bicycles, users are required to upload four standardized views: a front view (including brand logo), left and right side views, and a rear view. These images are input into a dual-channel detection model that processes them in parallel. The brand recognition channel in the dual-channel model only inputs the front view into a trained brand classification network, outputting the classification results and confidence values of mainstream brands to obtain the brand information of the used electric bicycle. The damage assessment channel in the dual-channel model simultaneously inputs all four views into a trained damage detection network, using neural networks to identify indicators such as vehicle body structural deformation, battery integrity, and tire wear. Finally, the weighted average score of the damage values from the four views is taken as the overall vehicle damage score, resulting in a damage rating for the used electric bicycle. Preferably, the weight of the side view is set to 40%, the weight of the rear view to 25%, and the weight of the front view to 35%.
[0063] The output data obtained through the aforementioned visual analysis—brand classification and damage score—will be combined with the vehicle information submitted by the user, including age, disposal method, and location, and packaged into structured data in JSON format before being input into the large model's decision engine. Local subsidy rules will be used as knowledge base information to generate a recycling approval opinion containing specific amounts, disposal requirements, and policy basis.
[0064] In this embodiment, by combining an improved target detection and classification algorithm with large model technology, the image features of the four views of the input waste electric bicycle can be accurately extracted. In the business scenario of waste electric bicycle recycling and scrapping, waste electric bicycles can be accurately and efficiently classified and identified, including but not limited to specific features such as brand. Furthermore, by utilizing large model technology, the extracted image features can be input for real-time and efficient intelligent approval.
[0065] Based on the same inventive concept, this application also provides a recycling approval system for used electric bicycles, the structure of which is as follows: Figure 5 As shown.
[0066] Figure 5 An architecture diagram of a recycling approval system for used electric bicycles provided in this application is shown below. Figure 5 As shown, the waste electric bicycle recycling approval system 500 in this application includes at least:
[0067] The information receiving module 501 receives standardized views and vehicle information uploaded by users for discarded electric bicycles.
[0068] The vehicle detection module 502 inputs the standardized view into the dual-channel detection model to identify the brand information of the discarded electric bicycle and output the damage score of the discarded electric bicycle.
[0069] The approval decision module 503 encapsulates the vehicle information, brand information, and damage score into a preset data format and inputs it into the decision model to obtain the recycling approval result corresponding to the waste electric bicycle. The decision model uses subsidy rules as a knowledge base to participate in the decision-making process.
[0070] This application proposes a method and system for approving the recycling of used electric bicycles. While using used electric bicycles as an example, the method is not limited to this scenario. It can also be applied to other used or scrapped vehicles, such as electric tricycles, bicycles, and even cars. The only difference is that the data used in the scheme is adjusted to correspond to the images of the waste recycling objects. The specific implementation process is the same as or similar to the process described above, and those skilled in the art can directly apply it. This application will not elaborate further. In other words, object detection and classification algorithms can be used to identify and classify the features of waste vehicles and other objects. Based on large-scale model technology combined with extracted features and other information, approval opinions are given, achieving intelligent approval in the business process scenarios of trade-in and scrapping. This improves the efficiency of recycling approval and ensures the objectivity of the results.
[0071] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0072] The equipment and method provided in this application are one-to-one correspondences. Therefore, the equipment also has similar beneficial technical effects as its corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the equipment will not be repeated here.
[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.
[0074] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0075] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for approving the recycling of used electric bicycles, characterized in that, The method includes: Receive standardized views and vehicle information uploaded by users for discarded electric bicycles; The standardized view is input into the dual-channel detection model to identify the brand information of the discarded electric bicycle and output the damage score of the discarded electric bicycle. After the vehicle information, brand information, and damage score are encapsulated into a preset data format, they are input into the decision model to obtain the recycling approval result corresponding to the waste electric bicycle. The decision model uses subsidy rules as a knowledge base to participate in the decision-making process.
2. The method for approving the recycling of used electric bicycles according to claim 1, characterized in that, After receiving standardized views of discarded electric bicycles uploaded by users, the method further includes: The standardized view is validated by determining whether the proportion of the area corresponding to the brand nameplate of the waste electric bicycle in the front image of the standardized view meets the first minimum proportion constraint, and whether the proportion of the area corresponding to the body of the waste electric bicycle in the side and rear images of the standardized view meets the second minimum proportion constraint.
3. The method for approving the recycling of used electric bicycles according to claim 1, characterized in that, The dual-channel detection model includes a brand classification network and a damage detection network; The brand classification network is used to identify the frontal image in the input standardized view and output the brand information and confidence result corresponding to the waste electric bicycle; The damage detection network is used to identify the front, side, and rear images in the input standardized view and output the damage score corresponding to the waste electric bicycle.
4. The method for approving the recycling of used electric bicycles according to claim 3, characterized in that, The damage detection network processes the frontal image, side image, and back image respectively to obtain damage scores corresponding to the frontal image, side image, and back image, and determines the damage score by weighted averaging of the damage scores.
5. The method for approving the recycling of used electric bicycles according to claim 1, characterized in that, The dual-channel detection model includes a brand classification network and a damage detection network. The training process of the brand classification network and / or the damage detection network includes: A frontal image of an electric bicycle is collected as training data for the brand classification network. A frontal image, a side image, and a rear image of an electric bicycle are collected as training data for the damage detection network. The side images include images of the left and right sides of the electric bicycle. The training data is labeled; The labeled training data is input into the network to be trained in batches for network training. Calculate the loss function value according to the definition of the loss function of the network to be trained; Based on the loss function value, backpropagation is performed using the Adam optimizer to update the network parameters of the network to be trained; After each preset number of training iterations, a network evaluation is performed. When the accuracy difference between two network evaluations is less than 1%, the network training ends.
6. The method for approving the recycling of used electric bicycles according to claim 5, characterized in that, The network to be trained includes a Backbone network module, a Neck network module, and convolutional layers: The Backbone network module adopts a CSP structure for feature extraction; The Neck network module employs SPPF and PAN structures for feature fusion. The convolutional layer is used to output the detection results.
7. The method for approving the recycling of used electric bicycles according to claim 6, characterized in that, In the network to be trained, The CSP structure is used to divide the input feature map into two parts. One part is processed by the convolutional sub-network, and the other part is directly fed into the next layer. Then, the two parts of the feature map are concatenated as the input of the next layer. The SPPF structure is used to perform pooling operations at different scales on the input feature map, and a feature fusion module is introduced to concatenate the pooling results at different scales as the output of the SPPF structure. The PAN structure is used for multi-level feature fusion.
8. The method for approving the recycling of used electric bicycles according to claim 5, characterized in that, The dual-channel detection model uses cross-entropy loss and leaky ReLU activation function during training.
9. The method for approving the recycling of used electric bicycles according to claim 1, characterized in that, The recycling approval results obtained for the aforementioned waste electric bicycles shall include at least one of the following: specific amount, disposal requirements, and policy basis.
10. A recycling approval system for used electric bicycles, characterized in that, The system includes: The information receiving module receives standardized views and vehicle information uploaded by users for discarded electric bicycles. The vehicle detection module inputs the standardized view into the dual-channel detection model to identify the brand information of the discarded electric bicycle and output the damage score of the discarded electric bicycle. The approval decision module encapsulates the vehicle information, brand information, and damage score into a preset data format and inputs them into the decision model to obtain the recycling approval result corresponding to the waste electric bicycle. The decision model uses subsidy rules as a knowledge base to participate in the decision-making process.