Scrapped vehicle intelligent evaluation method and system based on image recognition
By using an image recognition-based intelligent valuation method, which utilizes YOLO and XGBoost models for vehicle identification and feature fusion, the problems of high subjectivity, low efficiency, and poor accuracy in the valuation of scrapped vehicles are solved, achieving accurate, fair, and efficient valuation results.
Patent Information
- Application Number
- CN202511718247.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-02-17
AI Technical Summary
Existing methods for valuing scrapped vehicles suffer from high subjectivity, low efficiency, poor accuracy, and insufficient fairness, failing to achieve a good balance between efficiency, accuracy, and fairness.
An image recognition-based intelligent valuation method is adopted, which uses the YOLO model for vehicle type identification and key component localization, and combines the XGBoost model for feature fusion and residual value regression prediction. Through multi-source feature fusion of image data and non-visual data, accurate valuation is achieved.
It has significantly improved the accuracy and fairness of vehicle valuation, greatly increased efficiency, reduced operating costs, increased transparency, and the system has the ability to continuously evolve, thus forming a competitive advantage.
Smart Images
Figure CN121544291A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of end-of-life vehicle valuation and trading technology, specifically to an intelligent valuation method and system for end-of-life vehicles based on image recognition. Background Technology
[0002] In the scrap vehicle transaction, the traditional valuation method relies on human experience, which heavily depends on the valuer's personal experience. This results in problems such as strong subjectivity, inconsistent standards, and low efficiency. At the same time, this valuation method is opaque and has low credibility. Users find it difficult to understand the valuation process and often question the fairness of the valuation results, thus damaging the industry's reputation.
[0003] Secondly, some existing companies or platforms attempt to use formulaic pricing based on simple parameters. While this avoids the problem of manual labor to some extent, it ignores the individual differences of vehicles and cannot identify brand premiums or high-value components. It only makes valuations based on a very small number of macro parameters, which often seriously underestimates vehicles in good condition and overestimates vehicles in extremely poor condition, which is unfair.
[0004] In conclusion, existing technologies cannot achieve a good balance between efficiency, accuracy, and impartiality. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent valuation method and system for scrapped vehicles based on image recognition, so as to overcome the shortcomings of the prior art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for intelligent valuation of end-of-life vehicles based on image recognition includes the following steps:
[0008] S1: Data Collection and Preprocessing: This includes collecting user vehicle information data, collecting business data, cleaning image data, and data fusion and correlation.
[0009] S2: Image Recognition and Matching: After a user uploads a vehicle photo, the YOLO model is used to identify the vehicle type. The CSPDarknet convolutional network is used as the backbone network to extract features from the image at multiple levels and scales. Then, a shallow network is used to capture detailed features, including edges and textures, to identify scratches and gaps. A deep network is then used to synthesize semantic features, including overall shape and component combinations, to identify specific semantics. Combined with OCR recognition, the license plate number is identified.
[0010] S3: Algorithm Model Calculation: Transform the image recognition results into structured and quantifiable feature variables: Based on market data, perform one-hot encoding on brand model, displacement, and vehicle class. Combine the identified high residual value items with the crawled local market and environment as feature variables. Then, use the XGBOOST machine learning model to learn the complex nonlinear relationship between these variables and vehicle residual value. Combine the backend data and input the model to calculate the valuation result.
[0011] S4: Parameter optimization configuration: By continuously collecting massive amounts of vehicle image data, final transaction data, and market residual value fluctuation data, these newly generated real business data are used as incremental learning samples to regularly retrain the core recognition model and valuation model, thereby improving the self-evolution and iterative upgrade of model performance.
[0012] Furthermore, the method for collecting user vehicle information data in S1 is as follows: high-definition images of the vehicle to be scrapped are collected through the image upload interface deployed on the user's mobile APP or a specific recycling point terminal. The images include: the front and sides of the vehicle, as well as close-ups of key components, including the engine compartment, the chassis catalytic converter, tires, and high residual value information of the wheels. Users can optionally upload vehicle registration information: through vehicle registration image recognition, the vehicle's brand, model, model year, manufacturing date, VIN code, and basic attributes of the standard configuration are obtained.
[0013] The method for collecting business data is as follows: using web crawler technology, data is periodically scraped from multiple public channels, including scrap metal market information: the latest market prices of scrap steel, scrap aluminum, and scrap copper; second-hand parts market information: scraping core components from large second-hand parts trading platforms, including the recycling prices of engines, transmissions, generators, starters, and catalytic converters; and historical transaction data: obtaining historical scrap transaction prices and corresponding vehicle condition descriptions for the same model or category of vehicles from the database of scrapping and dismantling companies.
[0014] The image data cleaning method is as follows: Automated quality detection is performed on uploaded images to filter out invalid images that are blurry, too dark, too bright, or have severely distorted angles, and users are prompted to re-upload; qualified images are uniformly scaled to a fixed size and pixel value normalization is performed to accelerate model convergence, while contrast enhancement and noise reduction are performed to improve the accuracy of subsequent image recognition.
[0015] The method of data fusion and association is to associate and integrate images, attributes and market data from different sources in the database to form a complete data profile for each vehicle record.
[0016] Furthermore, S2 also includes the localization of key components and damaged areas: The target detection task is constructed as an end-to-end regression problem using the YOLO model. The input image is passed through a convolutional neural network in one pass, and the spatial location, class probability, and confidence score of the target are directly output. Specifically:
[0017] Suppose the input image is uniformly divided into an S×S grid, and each grid cell is responsible for predicting B bounding boxes and the conditional class probabilities of C categories corresponding to these bounding boxes. The YOLO model is further trained to identify and locate multiple key targets, including: high-value surviving parts such as engine compartment, three-way catalytic converter, and car wheel hub, as well as damage types and areas such as body rust, glass breakage, sheet metal deformation, tire wear, and missing parts.
[0018] For each grid cell, the model predicts an output tensor of dimension (B×5+C), where the prediction for each bounding box contains 5 core parameters:
[0019] (1) Boundary box center coordinates (b x b y ):
[0020] This coordinate is relative to the top-left corner coordinate of its grid cell (c x c y Normalization is performed, and the σ constraint is set between 0 and 1 using the sigmoid function. The calculation formula is as follows:
[0021] b x =σ(t) x )+c x
[0022] b y =σ(t) y )+c y
[0023] (2) Boundary box size (b) w b h ):
[0024] This size is relative to the width p of the entire image. w and height p h Normalization is performed using the predefined dimensions of the prior anchor box (p). w p h ) and the scaling factor (t) of the network prediction w , t h The calculation formula is as follows:
[0025]
[0026] (3) Bounding box confidence score:
[0027] This score represents the probability that the bounding box contains a target object and the accuracy of the predicted box's location. Confidence consists of two parts: the probability Pr(object) of the target object within the bounding box and the Intersection over Union (IoU) between the predicted and ground truth boxes. The formula for calculating this is:
[0028]
[0029] Among them, t c It is the raw confidence score directly predicted by the network;
[0030] (4) Each grid cell also predicts a set of conditional class probabilities Pr(Class i |Object):
[0031] Where i = 1, 2, ..., C, these probabilities are normalized by the softmax function and represent the probability that the target belongs to each class, assuming that the grid cell contains the target.
[0032] (5) Final object detection judgment: For a bounding box, its final score for a specific class i is determined by the class probability and the bounding box confidence:
[0033] Class-Specific Confidence Score=Pr(Class i |Object)×Confidence
[0034] By setting a threshold and applying a nonmaximum suppression algorithm to all candidate boxes, the final target detection results are selected.
[0035] Furthermore, by locating key components and damaged areas, the identification results are quantified into structured feature data, generating a "component existence vector" to record whether each high-value component exists, and generating a "damage degree vector" to record the area and severity level of various types of damage.
[0036] Furthermore, the algorithm model calculation in S3 includes: a multi-feature fusion and residual regression prediction algorithm based on XGBoost.
[0037] Let X be the characteristics of the sample of scrapped vehicles to be valued. i Its corresponding real residual value is y I The model's predicted value for this sample The additive model formula, determined by the outputs of K decision trees, is as follows:
[0038]
[0039] In the formula, The model's predicted output for the i-th sample represents the estimated vehicle residual value; K represents the total number of ensemble trees; f k f represents the k-th independent decision tree, which belongs to the function space F; k (X i ) represents the k-th tree pair of sample X i The predicted score;
[0040] In order to learn the set of functions {f} in the model k Minimize the following objective function with regularization:
[0041]
[0042] In the formula, Used to measure the difference between the predicted value and the actual residual value; ∑ k Ω(f k ) is a regularization term used to control the complexity of the model, specifically defined as:
[0043]
[0044] In the formula, T represents the number of leaf nodes in the tree model; W represents the score vector of the leaf node; γ and λ are configurable hyperparameters used to penalize model complexity and leaf node score, respectively.
[0045] Furthermore, the structured feature vectors generated in the YOLO stage are concatenated and fused with non-visual data obtained from external systems. The XGBoost regression model inputs the fused complete feature vectors into the pre-trained XGBoost regression model. Through its powerful ensemble learning capabilities, the XGBoost model performs weighted calculations and decision tree ensembles on hundreds of features, automatically learning the influence weights of each feature on the final residual value.
[0046] Furthermore, parameter optimization in S4 utilizes a machine learning model self-tuning algorithm, including:
[0047] (1) Incremental data collection and preprocessing: During operation, the system continuously collects two types of key data:
[0048] High-quality valuation samples are the complete feature data corresponding to the valuation records of successful transactions and their final transaction prices. The transaction price is regarded as a strong label for supervised learning.
[0049] User feedback data includes user behavior logs of "acceptance" or "rejection" of valuation results, as well as correction suggestions provided during manual review;
[0050] (2) Model performance monitoring and triggering mechanism: Collect data and monitor performance indicators, prediction accuracy, and data distribution drift to determine whether the model may be about to "expire";
[0051] The model retraining process is triggered when the model's prediction accuracy is below a preset threshold for a continuous period of time or when a significant data distribution shift is detected.
[0052] (3) Automated incremental training and hyperparameter optimization, including:
[0053] Incremental learning: The incremental learning algorithm is used. The new training process uses the existing pre-trained model as the basis and continues to train it using newly collected incremental data.
[0054] Automatic hyperparameter optimization: During each retraining, the model will start a grid search algorithm to optimize hyperparameters, searching for the parameter combination that achieves the best performance on the new validation set within a preset hyperparameter space.
[0055] This invention provides another technical solution: an intelligent valuation system for scrapped vehicles based on image recognition, comprising a data input layer, a visual perception layer, a feature fusion layer, a value decision layer, and a result output layer. The data input layer stores non-visual vehicle data, including year, mileage, and market price, as well as vehicle image data. The non-visual vehicle data is input to the feature fusion layer, and the vehicle image data is input to the visual perception layer. The visual perception layer is implemented based on the YOLO model and includes image preprocessing and grid partitioning, feature extraction and bounding box prediction, target classification and confidence scoring, and structured feature output. The structured feature output data is transmitted to the feature fusion layer, which concatenates multi-source feature vectors from the visual perception layer's output data and the vehicle's non-visual data before outputting it to the value decision layer. The value decision layer is implemented based on the XGBOOST model and includes feature importance weighted calculation and ensemble tree model regression prediction. The ensemble tree model regression prediction result is input to the result output layer, which generates an accurate valuation amount and an interpretable assessment report.
[0056] Compared with the prior art, the beneficial effects of the present invention are:
[0057] 1. The intelligent valuation method and system for scrapped vehicles based on image recognition of the present invention achieves a qualitative leap in valuation accuracy and fairness: By objectively extracting visual evidence through the YOLO model, the bias of subjective human judgment is completely eliminated. Combined with XGBoost's precise calculation of dozens or even hundreds of feature factors, the valuation result can accurately reflect the true residual value of the vehicle, avoiding the unfair phenomenon of "selling good cars cheaply and buying bad cars expensively," and greatly improving the scientificity and credibility of the valuation.
[0058] 2. The intelligent valuation method and system for scrapped vehicles based on image recognition of the present invention greatly improves valuation efficiency and business scale: the entire valuation process can be completed automatically within tens of seconds, achieving "second-level response". Compared with the traditional manual evaluation mode of several hours or even several days, the efficiency is improved by hundreds of times, enabling recycling companies to easily handle massive valuation requests from individual car owners, providing a solid technical foundation for large-scale, online expansion of business.
[0059] 3. The image recognition-based intelligent valuation method and system for end-of-life vehicles of this invention significantly reduces operating costs and yields substantial economic benefits. It greatly reduces reliance on experienced appraisers, thereby lowering enterprises' labor and management costs. Simultaneously, through accurate valuation, enterprises can more scientifically formulate acquisition and dismantling strategies, maximizing profit margins and avoiding economic losses due to misjudgments.
[0060] 4. The image recognition-based intelligent valuation method and system for scrapped vehicles of this invention significantly enhances transparency and user trust: by generating a visualized valuation report, it clearly displays the identified components, damage, and their specific impact on the price, transforming the valuation process from a "black box" to a "white box." This unprecedented transparency effectively dispels car owners' doubts, builds brand trust, and promotes the healthy development of the industry.
[0061] 5. The intelligent valuation method and system for scrapped vehicles based on image recognition of the present invention has the ability to continuously evolve and has a long technology life cycle: its self-learning mechanism ensures that the system will not become outdated over time, but will become more and more intelligent with use, forming a technical barrier and data barrier that is difficult for competitors to overcome, and bringing users a lasting competitive advantage in the market. Attached Figure Description
[0062] Figure 1 This is a flowchart of the valuation system of the present invention. Detailed Implementation
[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0064] To address the technical challenge of achieving a good balance between efficiency, accuracy, and fairness in end-of-life vehicle transactions using existing technologies, this invention provides an intelligent valuation method for end-of-life vehicles based on image recognition, comprising the following steps:
[0065] S1: Data Collection and Preprocessing: This includes collecting user vehicle information data, collecting business data, image data cleaning, and data fusion and correlation. The method for collecting user vehicle information data is as follows: High-resolution images of vehicles to be scrapped are collected via image upload interfaces deployed on user mobile apps or specific recycling point terminals. These images include the front and sides of the vehicle, as well as close-ups of key components such as the engine compartment, chassis catalytic converter, tires, and wheels—high-value parts. Users can optionally upload vehicle registration information: Through vehicle registration image recognition, basic attributes such as brand, model, vehicle year, manufacturing date, VIN code, and standard configuration (e.g., engine displacement, transmission type, drive type) are obtained, as shown in Table 1 below.
[0066] Table 1 Basic Vehicle Information
[0067]
[0068]
[0069] The method for collecting business data is as follows: using web crawler technology, data is periodically scraped from multiple public channels, including scrap metal market prices: the latest market prices of scrap steel, scrap aluminum, scrap copper, etc.; second-hand parts market prices: scraping the recycling prices of core components (such as engines, transmissions, generators, starters, catalytic converters, etc.) from large second-hand parts trading platforms; historical transaction data: obtaining the historical scrap transaction prices and corresponding vehicle condition descriptions of the same model or category of vehicles from the database of the scrap dismantling company.
[0070] The image data cleaning method is as follows: Uploaded images undergo automated quality checks, filtering out blurry, overly dark, overly bright, or severely distorted images, prompting users to re-upload. For qualified images, they are uniformly scaled to a fixed size, and pixel value normalization is performed to accelerate model convergence. Simultaneously, contrast enhancement and noise reduction are performed to improve the accuracy of subsequent image recognition, as shown in Table 2 below.
[0071] Table 2 Vehicle Image Storage Table
[0072]
[0073]
[0074] The method of data fusion and association is to associate and integrate images, attributes and market data from different sources in the database to form a complete data profile for each vehicle record.
[0075] S2: Image Recognition and Matching: After a user uploads a vehicle photo, the YOLO model is used to identify the vehicle type. The CSPDarknet convolutional network is used as the backbone network to perform multi-level, multi-scale feature extraction on the image. Shallow networks are used to capture detailed features (such as edges and textures, used to identify scratches and gaps), while deep networks are used to synthesize semantic features (such as overall shape and component assembly, used to identify "this is a car door" or "this is an aluminum wheel"), combined with OCR recognition to identify the license plate number, as shown in Table 3 below.
[0076] Table 3 Image Recognition Results
[0077]
[0078]
[0079] S3: Algorithm Model Calculation: The image recognition results are transformed into structured, quantifiable feature variables. Based on market data, brand model, engine displacement, vehicle class, etc., are encoded using one-hot encoding. High-residual-value items (complete doors, wheels, catalytic converters, etc.) are identified and combined with crawled local market and environmental data (government subsidies, scrap steel prices, regional coefficients, market factors) as feature variables. Then, the XGBOOST machine learning model is used to learn the complex nonlinear relationship between these variables and vehicle residual value. Combined with backend data, the model is input to calculate the valuation results, as shown in Table 4 below.
[0080] Table 4 Intelligent Valuation Record Table
[0081]
[0082]
[0083] S4: Parameter Optimization Configuration: By continuously collecting massive amounts of vehicle image data, final transaction data, and market residual value fluctuation data, this newly generated real business data is used as incremental learning samples to periodically retrain the core recognition and valuation models. Through this continuous model self-learning and parameter optimization configuration, the system can constantly adapt to market changes and accurately capture subtle value-influencing factors, thereby making the valuation results increasingly accurate and intelligent, achieving self-evolution and iterative upgrades of system performance. Simultaneously, business units can also adjust relevant parameter information, including key component price information and the weights of various feature variables, according to their own business strategies and market conditions.
[0084] In the above embodiments, the specific algorithm portion of the present invention includes:
[0085] 1. Algorithm for locating key components and damaged areas: This invention uses the YOLO model to construct its target detection task as an end-to-end regression problem. Its core lies in passing the input image through a convolutional neural network in one pass and directly outputting the target's spatial location, class probability, and confidence score. Specifically:
[0086] The input image is uniformly divided into an S×S grid. Each grid cell is responsible for predicting B bounding boxes and the conditional class probabilities of C categories corresponding to these bounding boxes. The YOLO model is further trained to identify and locate multiple key targets, including: high-value surviving parts such as engine compartments, catalytic converters, and car wheels; damage types and areas such as body rust, broken glass, sheet metal deformation, tire wear, and missing parts.
[0087] For each grid cell, the model predicts an output tensor of dimension (B×5+C), where the prediction for each bounding box contains 5 core parameters:
[0088] (1) Boundary box center coordinates (b x b y ):
[0089] This coordinate is relative to the top-left corner coordinate of its grid cell (c x c y Normalization is performed, and the σ constraint is set between 0 and 1 using the sigmoid function. The calculation formula is as follows:
[0090] b x =σ(t) x )+c x
[0091] b y =σ(t) y )+c y
[0092] (2) Boundary box size (b) w b h ):
[0093] This size is relative to the width p of the entire image. w and height p h Normalization is performed using the predefined dimensions (p) of the prior anchor box. w p h ) and the scaling factor (t) of the network prediction w , t h The calculation formula is as follows:
[0094]
[0095] This design enables the model to more stably predict targets at different scales.
[0096] (3) Bounding box confidence score:
[0097] This score represents the probability that the bounding box contains a target object and the accuracy of the predicted box's location. Confidence consists of two parts: the probability Pr(object) of the target object within the bounding box and the Intersection over Union (IoU) between the predicted and ground truth boxes. The formula for calculating this is:
[0098]
[0099] Among them, t c It is the raw confidence score directly predicted by the network;
[0100] (4) Each grid cell also predicts a set of conditional class probabilities Pr(Class i |Object):
[0101] Where i = 1, 2, ..., C, these probabilities are normalized by the softmax function and represent the probability that the target belongs to each class, assuming that the grid cell contains the target.
[0102] (5) Final object detection judgment: For a bounding box, its final score for a specific class i is determined by the class probability and the bounding box confidence:
[0103] Class-Specific Confidence Score=Pr(Class i |Object)×Confidence
[0104] By setting a threshold and applying the Non-Maximum Suppression (NMS) algorithm to all candidate boxes, the final target detection results are selected.
[0105] During this stage, the output is not the final result, but rather the quantification of the identification results into structured feature data. A "component existence vector" is generated to record whether each high-value component exists (1 or 0), and a "damage severity vector" is generated to record the area and severity level (e.g., none, mild, severe) of various types of damage.
[0106] 2. In the above embodiments, the specific algorithm part of the present invention further includes: a multi-feature fusion and residual regression prediction algorithm based on XGBoost:
[0107] Let X be the characteristics of the sample of scrapped vehicles to be valued.i Its corresponding real residual value is y I The model's predicted value for this sample The additive model formula, determined by the outputs of K decision trees, is as follows:
[0108]
[0109] In the formula, The model's predicted output for the i-th sample represents the estimated vehicle residual value; K represents the total number of ensemble trees; f k f represents the k-th independent decision tree, which belongs to the function space F; k (X i ) represents the k-th tree pair of sample X i The predicted score;
[0110] In order to learn the set of functions {f} in the model k We need to minimize the following objective function with regularization:
[0111]
[0112] In the formula, Used to measure the difference between the predicted value and the actual residual value; ∑ k Ω(f k ) is a regularization term used to control the complexity of the model, specifically defined as:
[0113]
[0114] In the formula, T represents the number of leaf nodes in the tree model; w represents the score vector of the leaf node; γ and λ are configurable hyperparameters used to penalize model complexity and leaf node score, respectively.
[0115] In this stage, the structured feature vectors generated in the YOLO stage are concatenated and fused with non-visual data obtained from external systems. The XGBoost regression model inputs the fused complete feature vectors into a pre-trained XGBoost regression model. Through its powerful ensemble learning capabilities, the XGBoost model performs weighted calculations and decision tree ensembles on hundreds of features. XGBoost can automatically learn the influence weights (i.e., feature importance) of each feature (such as "presence of a three-way catalytic converter" versus "area of rust on the vehicle body") on the final residual value. For example, the model learns that the positive contribution of "a perfectly good three-way catalytic converter" to the price may far outweigh the negative contribution of "minor bumper scratches."
[0116] 3. In the above embodiments, the specific algorithm part of the present invention further includes: parameter optimization configuration using a machine learning model self-tuning algorithm, including:
[0117] (1) Incremental data collection and preprocessing: During operation, the system continuously collects two types of key data:
[0118] High-quality valuation samples are the complete feature data corresponding to the valuation records of successful transactions and their final transaction prices. The transaction price is regarded as a strong label for supervised learning.
[0119] User feedback data – a log of user "acceptance" or "rejection" of valuation results, as well as correction suggestions provided during manual review;
[0120] (2) Model performance monitoring and triggering mechanism: Collect data and monitor and predict accuracy (mean absolute error (MAE) and root mean square error (RMSE) between model valuation and final transaction price) and data distribution drift detection (monitor the difference between the feature distribution of newly collected data and the distribution of the old training set (e.g., through KL divergence)) to determine whether the model may be about to "expire".
[0121] The model retraining process is triggered when the model's prediction accuracy is below a preset threshold for a continuous period of time or when a significant data distribution shift is detected.
[0122] (3) Automated incremental training and hyperparameter optimization, including:
[0123] Incremental learning: The system does not train from scratch every time, but uses an incremental learning algorithm. The new training process uses the existing pre-trained model as a foundation and continues to train it using newly collected incremental data, which greatly saves computing resources and time;
[0124] Automatic hyperparameter optimization: During each retraining, the system will start a grid search algorithm to optimize hyperparameters, searching for the parameter combination that achieves the best performance on the new validation set within a preset hyperparameter space (such as learning rate eta, maximum tree depth max_depth, regularization parameter lambda, etc.).
[0125] To implement the valuation method described in the above embodiments, such as Figure 1As shown, this invention also provides an intelligent valuation system for scrapped vehicles based on image recognition, comprising a data input layer, a visual perception layer, a feature fusion layer, a value decision layer, and a result output layer. The data input layer stores non-visual vehicle data, including year, mileage, and market price, as well as vehicle image data. The non-visual vehicle data is input to the feature fusion layer, and the vehicle image data is input to the visual perception layer. The visual perception layer is implemented based on the YOLO model, including image preprocessing and grid partitioning, feature extraction and bounding box prediction, target classification and confidence scoring, and structured feature output. The structured feature output data is transmitted to the feature fusion layer, which concatenates multi-source feature vectors from the visual perception layer's output data and the vehicle's non-visual data before outputting it to the value decision layer. The value decision layer is implemented based on the XGBOOST model, including feature importance weighted calculation and ensemble tree model regression prediction. The ensemble tree model regression prediction result is input to the result output layer, which generates an accurate valuation amount and an interpretable assessment report.
[0126] In summary, this invention provides an intelligent valuation method and system for scrapped vehicles based on image recognition. It constructs a two-stage intelligent valuation pipeline of "visual perception-value decision-making," achieving refined residual value assessment based on multimodal feature fusion. Abandoning the traditional coarse valuation method that relies solely on macroscopic parameters such as vehicle weight and model, this invention creatively proposes a multi-dimensional, fine-grained vehicle residual value feature system. This invention can continuously collect real transaction data and user feedback, and use this newly generated data as incremental learning samples to periodically fine-tune and retrain the YOLO and XGBoost models, establishing a data-driven model self-optimization closed loop.
[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. An image recognition-based intelligent valuation method for a scrapped vehicle, characterized by, Comprise the following steps: S1: data collection and pretreatment: including collecting user vehicle information data, collecting business data, image data cleaning and data fusion and association; S2: image recognition and matching: after the user uploads the vehicle photo, the YOLO model is used to identify the vehicle type, the CSPDarknet convolutional network is used as the backbone network, the image is extracted in multiple levels and multiple scales, then the shallow network is used to capture the detail features, including edges and textures, which are used to identify scratches and gaps, and the deep network is used to synthesize semantic features, including overall shape and component combination, which are used to identify specific semantics, and the OCR recognition is combined to identify the license plate number; S3: algorithm model calculation: the image recognition result is converted into structured and quantifiable feature variables: according to market data, the brand, model, displacement and vehicle type level are one-hot encoded, the identified high residual value objects are combined with the local market and environment crawled to form feature variables, then the XGBOOST machine learning model is used to learn the complex nonlinear relationship between these variables and vehicle residual value, and the backend data is input into the model to calculate the estimated value result; S4: parameter optimization configuration: by continuously collecting massive vehicle image data, final transaction data and market residual value fluctuation data, these newly generated real business data are used as incremental learning samples, the core recognition model and the valuation model are retrained regularly to improve the self evolution and iterative upgrade of the model performance.
2. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 1, wherein: The method for collecting user vehicle information data in S1 is: through the image upload interface deployed in the user's mobile phone APP or specific recycling point terminal, high-definition images of the vehicle to be scrapped are collected, including: front, side and key component close-up of the vehicle, including engine compartment, chassis three-way catalytic converter, tire and hub high residual value part information; the user can select to upload the driving license information: through driving license image recognition, the basic attributes of the vehicle brand, model, vehicle type year, factory date, VIN code and standard configuration are obtained; The method for collecting business data is: through network crawler technology, data is grabbed from multiple public channels at regular intervals, including: scrap metal market: the latest market prices of scrap steel, scrap aluminum and scrap copper; second-hand parts market: the recycling prices of core parts, including engine, transmission, generator, starter and catalytic converter, are grabbed from large second-hand parts trading platforms; historical transaction data: the historical scrap transaction prices and corresponding vehicle condition descriptions of vehicles of the same type or same category are obtained from the database of scrap dismantling enterprises; The method for image data cleaning is: the uploaded images are automatically quality tested, invalid pictures with blur, too dark, too bright or serious distortion of angle are filtered out, and the user is prompted to upload again; for qualified images, uniform scaling to a fixed size is performed, and pixel value normalization is executed to accelerate model convergence, and contrast enhancement and noise elimination are performed to improve the accuracy of subsequent image recognition; The method for data fusion and association is: the images, attributes and market data from different sources are associated and integrated in the database to form a complete data portrait of each vehicle record.
3. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 1, wherein: S2 also includes positioning of key components and damage areas: the target detection task of YOLO model is constructed as an end-to-end regression problem, the input image is once through a convolutional neural network, and the spatial position, class probability and confidence of the target are directly output. Specifically: The input image is evenly divided into an SxS grid, and each grid cell is responsible for predicting B bounding boxes and the conditional class probability of the C classes corresponding to the bounding boxes. The YOLO model is further trained to identify and locate multiple key targets, including: engine compartment, three-way catalytic converter, high-value residual components of automobile hub, and damage types and areas such as body rust, glass breakage, sheet metal deformation, tire wear, and component missing; For each grid cell, the model predicts a tensor with dimensions (Bx5+C), where the prediction of each bounding box includes 5 core parameters: (1) the boundary box center coordinate (b x , b y ) : The coordinates are normalized with respect to the upper left corner coordinates (c x , c y ) of the grid cell to which they belong and are constrained between 0 and 1 by the sigmoid function σ, whose formula is: b x = σ(t x )+ c x b y = σ(t y )+ c y (2) the size of the bounding box (b w , b h ) The size relative to the width p of the entire image w and height p h is normalized and calculated by the size (p w , p h ) of a predefined prior anchor box AnchorBox and the scaling amount (t w , t h ) predicted by the network, and the calculation formula is: (3) Boundary box confidence score (Confidence): This score represents the possibility of containing a target object in the bounding box and the position accuracy of the prediction box. Confidence is composed of two parts: the probability Pr(object) that there is a target in the bounding box and the intersection over union IoU of the prediction box and the true box, and the calculation formula is: where t c is the original confidence score predicted by the network directly; (4) Each grid cell also predicts a set of conditional class probabilities Pr(Class | Object) i |Object): Where i = 1, 2, …, C, these probabilities are normalized by the softmax function, indicating the probability that the target belongs to each class under the premise that the grid cell contains the target; (5) Final target detection judgment: for a bounding box, the final score for a specific class i is determined by the class probability and the bounding box confidence: Class-Specific Confidence Score = Pr(Class | Object) x Confidence i |Object) x Confidence By setting a threshold and applying a non-maximum suppression algorithm to all candidate boxes, the final target detection result is selected.
4. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 3, wherein: By positioning the key components and damage areas, the recognition results are quantified into structured feature data, generating a "component presence vector" to record the presence or absence of each high-value component, and a "damage degree vector" to record the area and severity level of each type of damage.
5. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 1, wherein: S3 algorithm model calculation includes: XGBoost-based multi-feature fusion and residual value regression prediction algorithm: Let the sample feature of the vehicle to be evaluated be X i , and its corresponding true residual value be y I , the predicted value of the model for the sample is The output of K decision trees jointly determines it, and its additive model formula is as follows: wherein, represents the predicted output of the model for the i-th sample, i.e., the estimated vehicle residual value; K represents the total number of ensemble trees; f k represents the i-th sample; k represents the k-th independent decision tree, which belongs to the function space F; f k (X i ) represents the prediction score of the k-th tree for the sample X i . To learn the function set {f k} in the model, the following objective function with a regularization term is minimized: wherein for measuring the difference between the predicted value and the true residual value;∑ k Ω(f k ) is a regularization term for controlling the complexity of the model, which is defined as: Where T represents the number of leaf nodes in the tree model; W represents the score vector of the leaf node; γ and λ are configurable hyperparameters used to penalize model complexity and leaf node scores, respectively.
6. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 5, wherein: The structured feature vector generated by the YOLO stage is spliced and fused with the non-visual data obtained from external systems, and the XGBoost regression model inputs the complete feature vector after fusion into the pre-trained XGBoost regression model. XGBoost model automatically learns the influence weight of each feature on the final residual value through powerful ensemble learning capability, weighted calculation and decision tree integration of hundreds of features.
7. The image recognition based intelligent valuation method for a scraped vehicle as claimed in claim 1, wherein: S4 parameter optimization configuration uses machine learning model self-tuning algorithm, including: (1) Incremental data collection and preprocessing: the system continuously collects two types of key data during operation: High-quality valuation samples, i.e. complete feature data corresponding to the valuation records of the final successful transactions and the final transaction price, which is regarded as a strong label for supervised learning; User feedback data, i.e. the behavior log of "acceptance" or "rejection" of the valuation results by the user, and the correction opinions provided during manual review; (2) Model performance monitoring and triggering mechanism: collect data and monitor through performance indicators, prediction accuracy, and data distribution drift detection to determine whether the model is likely to be "expired"; The triggering condition of the model retraining process is that the prediction accuracy of the model is lower than the preset threshold for a continuous period of time or significant data distribution drift is detected; (3) Automatic incremental training and hyperparameter optimization, including: Incremental learning: use incremental learning algorithms, and the new training process uses the existing pre-trained model as the basis and continues to train it using the newly collected incremental data; Hyperparameter automatic optimization: in each retraining, the model will start the grid search algorithm for hyperparameter optimization to find the parameter combination that can achieve the best performance on the new validation set in the preset hyperparameter space.
8. An image recognition based intelligent valuation system for a vehicle to be scrapped, for implementing the image recognition based intelligent valuation method for a vehicle to be scrapped according to any one of claims 1-7, characterized in that: The data input layer, visual perception layer, feature fusion layer, value decision layer and result output layer are included; the vehicle non-visual data including year, mileage and market price and vehicle image data are stored in the data input layer; wherein the vehicle non-visual data is used for input to the feature fusion layer, and the vehicle image data is used for input to the visual perception layer; the visual perception layer is realized based on the YOLO model, and includes image preprocessing and grid division, feature extraction and boundary box prediction, target classification and confidence score and structured feature output; wherein the structured feature output data is transmitted to the feature fusion layer, and the feature fusion layer outputs the output data of the visual perception layer and the vehicle non-visual data after multi-source feature vector splicing to the value decision layer; the value decision layer is realized based on the XGBOOST model, and includes feature importance weighting calculation and integrated tree model regression prediction, wherein the integrated tree model regression prediction result is input to the result output layer, and the result output layer is used to generate a precise valuation amount and an explainability evaluation report.