Search

CN-121998893-A - Automatic detection method for image quality of second-hand vehicle based on image recognition

CN121998893ACN 121998893 ACN121998893 ACN 121998893ACN-121998893-A

Abstract

The invention relates to the technical field of secondary handcart photographing quality detection, in particular to a secondary handcart image quality automatic detection method based on image recognition. The method comprises the steps of collecting images of a vehicle to be detected and shooting scene information, generating an original image dataset, preprocessing the original image dataset, extracting four core quality characteristics of definition, exposure degree, angle compliance and shooting completeness, generating a multi-dimensional quality characteristic set, inputting the multi-dimensional quality characteristic set into a pre-trained quality classification model, generating an image quality grade and a defect type identifier, wherein the image quality grade comprises high quality, repairable and unrepairable, calling a corresponding image restoration algorithm for targeted optimization aiming at the image marked as repairable, generating a restored image, re-extracting the quality characteristic of the restored image, checking, confirming whether the quality characteristic meets a preset quality standard or not, and generating a final quality detection result. The invention can greatly improve the credibility of the second-hand vehicle image.

Inventors

  • SU CHANGZHENG

Assignees

  • 北京酷车易美网络科技有限公司

Dates

Publication Date
20260508
Application Date
20251201

Claims (10)

  1. 1. The automatic detection method for the quality of the second-hand vehicle image based on the image recognition is characterized by comprising the following steps of: Collecting a vehicle image to be detected and shooting scene information, and generating an original image data set; preprocessing an original image data set, and generating an adaptive detection image through image noise reduction and format standardization processing; Based on the adaptive detection image, four core quality characteristics of definition, exposure, angle compliance and shooting integrity are extracted, and a multi-dimensional quality characteristic set is generated; inputting a multi-dimensional quality feature set into a pre-trained quality classification model, and generating an image quality grade and a defect type identifier by combining a preset quality evaluation rule, wherein the image quality grade comprises high quality, repairable and unrepairable; Aiming at the image marked as repairable, a corresponding image repairing algorithm is called based on the defect type to carry out targeted optimization, and a repaired image is generated; And re-extracting quality characteristics of the repaired image, checking, and determining whether the quality characteristics meet the preset quality standard or not to generate a final quality detection result.
  2. 2. The automatic detection method for quality of images of a second-hand vehicle based on image recognition according to claim 1, wherein the steps of collecting images of the vehicle to be detected and shooting scene information include the steps of: Receiving a multi-view image of a vehicle to be detected through an image acquisition interface, wherein the multi-view image comprises a vehicle head, a vehicle tail, a vehicle body side surface, an engine cabin and an interior trim, extracting an image pixel matrix, resolution and color space parameters, and generating initial image data; acquiring shooting scene information, including shooting environment illumination intensity, shooting equipment type, shooting distance estimation value and whether flash lamp identification is used or not, and generating scene attribute data; Extracting image shooting time, uploading terminal identification and vehicle identification code association information to generate image metadata; carrying out association analysis on the initial image data, the scene attribute data and the image metadata, calculating the semantic matching degree of the data, and generating a data association coefficient; Weighting and integrating the initial image data, the scene attribute data and the image metadata based on the data association coefficient, and adding a source identifier and an integrity tag to each piece of data to generate an original image data set; And detecting missing values of the original image data set, identifying the non-shot part and the key scene information missing type, marking the data integrity level, and providing basis for subsequent preprocessing.
  3. 3. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 1, wherein the steps of extracting four core quality features of definition, exposure, angle compliance and shooting integrity based on the adaptive detection image comprise the following steps: Extracting edge contour information of the adaptive detection image by adopting an edge gradient analysis algorithm, calculating contour sharpness parameters, and generating a definition characteristic value; Through pixel brightness histogram analysis, counting an image brightness distribution interval and an extremum duty ratio, and generating an exposure characteristic value; calling a vehicle key part detection model to identify a vehicle core part in an adaptation detection image, wherein the vehicle core part comprises a headlight, wheels and a vehicle body contour, calculating the deviation between an actual shooting angle and a standard angle, and generating an angle compliance characteristic value; marking a vehicle coverage area in an image through a semantic segmentation algorithm based on a secondary vehicle shooting part list, comparing the shooting part coverage rate, and generating a shooting integrity characteristic value; Carrying out standardization processing on the four types of characteristic values, eliminating dimension differences and generating standardized characteristic vectors; and carrying out scene adaptation correction on the standardized feature vector by combining the illumination intensity and shooting distance parameters in shooting scene information to generate a multi-dimensional quality feature set.
  4. 4. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 3, wherein the step of extracting edge contour information of the adapted detection image by using an edge gradient analysis algorithm, and calculating a contour sharpness parameter comprises the steps of: Carrying out graying treatment on the adaptive detection image, converting the adaptive detection image into a single-channel gray image, retaining edge contour information, and generating gray image data; Extracting edge pixel points in gray image data by adopting a self-adaptive edge detection algorithm to generate an edge pixel set; Calculating gray difference values of adjacent pixels in the edge pixel set, and counting gray difference value distribution variance to generate edge gradient dispersion parameters; calculating a contour sharpness basic value based on the edge gradient dispersion parameter and combining the image resolution information; invoking a scene illumination correction model, and adjusting a contour sharpness basic value according to illumination intensity parameters in shooting scene information to generate a corrected sharpness value; And comparing the corrected sharpness value with a preset sharpness threshold range to generate a sharpness characteristic value which comprises a sharpness value and a blur mark.
  5. 5. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 4, wherein the step of calling a scene illumination correction model and adjusting the contour sharpness base value according to the illumination intensity parameter in the photographed scene information comprises the steps of: Extracting illumination intensity parameters from shooting scene information, and generating an illumination level identification through an illumination level classification model; Constructing an illumination-sharpness correction coefficient mapping library, and matching corresponding basic correction coefficients based on the illumination level identification; Analyzing and adapting to noise distribution characteristics of the detected image, and calculating a noise interference coefficient, wherein the noise interference coefficient is derived through pixel fluctuation degree of an image smoothing area; Calculating a comprehensive correction coefficient based on the basic correction coefficient and the noise interference coefficient, wherein the formula is that the comprehensive correction coefficient=the basic correction coefficient× (1-noise interference coefficient); performing product operation on the contour sharpness basic value and the comprehensive correction coefficient to generate a corrected sharpness value; and (3) carrying out validity check on the corrected sharpness value, ensuring that the corrected sharpness value is in a reasonable numerical value interval, and if the corrected sharpness value exceeds the range, adjusting the comprehensive correction coefficient to recalculate, so as to generate the final corrected sharpness value.
  6. 6. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 1, wherein the step of inputting the multi-dimensional quality feature set into a pre-trained quality classification model and generating the image quality level and defect type identification by combining a preset quality evaluation rule comprises the following steps: Inputting a multi-dimensional quality feature set into a pre-trained random forest classification model, and outputting a preliminary quality grade and contribution degree weight of each feature, wherein the preliminary quality grade comprises high quality, repairable and unrepairable; extracting abnormal features exceeding a preset threshold value in the multi-dimensional quality feature set, and marking the types of the abnormal features; Based on the abnormal feature type and the primary quality level, invoking a quality evaluation rule base to carry out secondary judgment, wherein the rule base comprises single feature serious exceeding direct judgment logic and multi-feature slight abnormal combination judgment logic; calculating severity parameters of abnormal features, wherein the severity parameters are obtained through deduction of influence weights of deviation degrees of abnormal feature values and standard thresholds on detection results; combining the severity parameter and the contribution degree weight to generate a defect type identifier containing a main defect type and a secondary defect type; And integrating the primary quality grade, the secondary judging result and the defect type identifier to generate a final image quality grade and the defect type identifier.
  7. 7. The automatic detection method for quality of image recognition based on second hand cart of claim 6, wherein calculating severity parameters of abnormal features, generating defect type identification comprises the steps of: classifying and carding the marked abnormal feature types, extracting feature values, standard threshold ranges and influence weight parameters corresponding to each abnormal feature, and generating an abnormal feature detail table; Calculating the deviation degree of each abnormal characteristic value and the standard threshold central value, and generating characteristic deviation parameters; Calculating a severity subparameter of the single abnormal feature based on the feature deviation parameter and the impact weight parameter; Analyzing the association relation among abnormal features, and calculating feature association influence coefficients; calculating a comprehensive severity parameter based on the severity subparameter and the feature-related influence coefficient, the formula being comprehensive severity parameter = Σ (severity subparameter× (1+ feature-related influence coefficient)); and dividing the severity level according to the comprehensive severity level parameters, generating defect type identifiers by combining the abnormal characteristic types, and determining the main defect type and the secondary defect type.
  8. 8. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 7, wherein the analyzing of the association relation between abnormal features, calculating the feature association influence coefficient, comprises the steps of: Constructing a secondary handcart shooting quality abnormal characteristic association knowledge graph which comprises forward/reverse association rules among four types of characteristics including definition, exposure, angle compliance and shooting integrity, and generating an association rule set; Extracting association rules related to the current abnormal feature type from the association rule set, and marking the association feature type and the association direction; Based on the historical detection data, counting the probability and the mutual influence degree of the simultaneous occurrence of the associated feature type and the current abnormal feature type, and generating historical associated parameters; calculating the ratio of the severity subparameter of the current abnormal feature to the reference severity parameter of the associated feature to generate a feature influence proportion; calculating a characteristic association influence coefficient based on the association direction, the history association parameter and the characteristic influence proportion, wherein the positive association takes a positive value and the negative association takes a negative value; and carrying out normalization processing on the characteristic association influence coefficients, ensuring the uniformity of the numerical range, and generating final characteristic association influence coefficients.
  9. 9. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 8, wherein the statistics of the probability and the degree of interaction of the associated feature type and the current abnormal feature type simultaneously occur based on the history detection data comprises the steps of: extracting all detection records containing the current abnormal feature type from a historical detection database to generate a target feature historical data set; Screening records containing associated feature types from the target feature historical data set, and counting the ratio of the number of the screened records to the total number of the target feature historical data to generate co-occurrence probability; calculating the influence degree of the associated feature type on the current abnormal feature type in the abnormal features which occur simultaneously, wherein the degree is deduced through the severity degree sub-parameter change trend of the two types of features; based on the co-occurrence probability and the influence degree, constructing a history association evaluation model, and generating history association basic parameters; Calling a data timeliness correction model, adjusting history associated basic parameters by combining time stamp information of a history record, and giving higher weight to recent data to generate history associated parameters; And carrying out validity check on the history associated parameters, removing abnormal values, and generating final history associated parameters.
  10. 10. The automatic detection method for quality of a second-hand cart based on image recognition according to claim 1, wherein the invoking a corresponding image restoration algorithm based on a defect type for targeted optimization for the image identified as repairable, generating a restored image comprises the following steps: Extracting a main defect type and a secondary defect type from the defect type identifier to generate a repair priority sequence; aiming at the abnormal defect of definition, invoking and generating an anti-network deblurring algorithm to learn an image blur kernel, reversely generating a clear image, synchronously adopting a bilateral filtering algorithm to reserve edge details, and generating a definition repair image; Aiming at the exposure abnormal defect, if the exposure is over, a multi-scale Retinex algorithm is called to adjust a brightness channel and inhibit a highlight region, and if the exposure is under, a self-adaptive histogram equalization algorithm is called to enhance local contrast and generate an exposure repair image; aiming at the color distortion defect, a white balance correction algorithm is called to calculate an image gray average value, the proportion of RGB channels is adjusted, the true color of the vehicle is restored, and a color restoration image is generated; Aiming at the defect of slight angle deviation, based on the feature points identified by the vehicle key part detection model, invoking a perspective transformation algorithm to adjust the image visual angle, correcting the angle deviation and generating an angle restoration image; sequentially applying various repair algorithms to the original repairable image according to the repair priority sequence to generate a comprehensive repair image; and carrying out edge smoothing treatment on the comprehensive repair image, eliminating transition marks among different repair algorithms, and generating a final repaired image.

Description

Automatic detection method for image quality of second-hand vehicle based on image recognition Technical Field The invention relates to the technical field of secondary handcart photographing quality detection, in particular to a secondary handcart image quality automatic detection method based on image recognition. Background In recent years, with the rapid transformation of a second-hand vehicle transaction to an 'on-line + off-line' fusion mode, a vehicle image becomes a core carrier for presenting vehicle conditions, supporting different-place transactions and defining responsibility, and higher requirements are put on standardization, authenticity and detail integrity of photographing quality of the second-hand vehicle. The traditional second-hand vehicle photographing relies on manual photographing and subjective auditing, lacks a unified quality standard and an automatic detection mechanism, is prone to the problems of missed photographing of key parts, blurred images, light distortion and the like, not only affects the judgment of a customer on vehicle conditions and the pricing efficiency of a merchant, but also can cause transaction disputes due to insufficient quality of image evidence. At present, a plurality of processing methods related to the second-hand vehicle images are proposed, the methods are mainly used for carrying out preliminary checking by manually making shooting guidance and combining simple image definition and size screening rules, and then judging whether the images meet the transaction requirements or not by means of manual checking, however, the traditional methods are insufficient in systematic detection and standardization of shooting quality, on one hand, the shooting quality is influenced by the fact that the shooting of key parts of the second-hand vehicle is lack of uniformity and the requirements of the second-hand vehicle transaction scene are not combined, on the other hand, the condition that the images can completely present the core information such as vehicle appearance, engine cabin and interior trim is difficult to ensure only through basic image parameter screening, on the other hand, the traditional manual checking mode is low in efficiency and strong in subjectivity, the image detection requirements of mass online transactions are difficult to deal with, and the problem that whether the images have falsification, shielding and the like influence on the storage value cannot be accurately identified, so that the reliability of the second-hand vehicle images is reduced. Disclosure of Invention The invention provides a method for generating an adaptive detection image through collecting vehicle images and shooting scene information, noise reduction and standardized preprocessing, and lays a quality detection foundation. Four types of core quality characteristics such as definition and exposure are extracted, a pre-training model is input to be combined with an evaluation rule, quality grade and defect identification are automatically generated, low-efficiency subjective manual review is replaced, and the requirement of mass transaction detection is met. And (5) invoking corresponding algorithm targeted optimization aiming at the repairable image, and re-checking the quality characteristics to ensure that the repairable image meets the preset standard. The automatic detection method for the quality of the second-hand vehicle image based on image identification, which is characterized in that a closed loop of acquisition, pretreatment, feature extraction, intelligent grading, repair and verification is integrally constructed, a unified quality detection standard is established, the image can be guaranteed to completely present the core information of the vehicle, the problem that falsification, shielding and the like affect the value of the stored certificate is avoided, and the detection efficiency and the reliability of the image are greatly improved, comprises the following steps: Collecting a vehicle image to be detected and shooting scene information, and generating an original image data set; preprocessing an original image data set, and generating an adaptive detection image through image noise reduction and format standardization processing; Based on the adaptive detection image, four core quality characteristics of definition, exposure, angle compliance and shooting integrity are extracted, and a multi-dimensional quality characteristic set is generated; inputting a multi-dimensional quality feature set into a pre-trained quality classification model, and generating an image quality grade and a defect type identifier by combining a preset quality evaluation rule, wherein the image quality grade comprises high quality, repairable and unrepairable; Aiming at the image marked as repairable, a corresponding image repairing algorithm is called based on the defect type to carry out targeted optimization, and a repaired image is generated; And re-extracting qua