Search

CN-121982080-A - Multiple equipartition deviation correcting method and system based on machine learning

CN121982080ACN 121982080 ACN121982080 ACN 121982080ACN-121982080-A

Abstract

The invention relates to the technical field of image processing, in particular to a multiple equipartition deviation correcting method and system based on machine learning. The method comprises the steps of obtaining a multisource image set, dividing the multisource image set into units to be rectified, aggregating characteristic distribution of images in the same unit to be rectified, constructing characteristic distribution description corresponding to each unit to be rectified, comparing the characteristic distribution description of each unit to be rectified, determining deviation indexes of each unit to be rectified, performing equipartition rectification on each image in each unit to be rectified, inputting equipartition rectification results into a preset image classification model to generate classification results of each image, comparing the classification results of each image to identify characteristic deviation conditions of each image after rectification, taking the images with characteristic deviation conditions after rectification as comparison samples of the characteristic distribution description, and performing equipartition rectification in an iterative mode. The invention effectively weakens systematic distribution deviation caused by multi-batch, multi-equipment and multi-scene imaging, and improves automatic deviation correcting capability and mobility.

Inventors

  • ZHANG XU

Assignees

  • 誊展精密科技（深圳）有限公司

Dates

Publication Date
20260505
Application Date
20260203

Claims (10)

  1. 1. The multiple equipartition deviation correcting method based on machine learning is characterized by comprising the following steps: s1, acquiring a multi-source image set, and dividing the multi-source image set into a plurality of units to be rectified; step S2, aggregating the characteristic distribution of the images in the same unit to be rectified, and constructing the characteristic distribution description corresponding to each unit to be rectified according to the aggregation result; S3, comparing the characteristic distribution description of each unit to be rectified, and determining the offset index of each unit to be rectified; Step S4, inputting the average deviation correction result into a preset image classification model to generate classification results of the images; And S5, taking the image with the characteristic deviation after deviation correction as a comparison sample of the characteristic distribution description, and iteratively executing equipartition deviation correction until the deviation index in each unit to be corrected is converged into the target equipartition interval.
  2. 2. The machine learning-based multiple equipartition correction method according to claim 1, wherein in step S1, dividing the multi-source image set into a plurality of units to be corrected includes: extracting an acquisition equipment identifier, acquisition time, a shooting scene identifier and image resolution of each image in a multi-source image set; Dividing images which are identical in identifier of the multi-source image centralized acquisition equipment and have the acquisition time within the same time window into first candidate deviation rectifying units; Dividing images which are shot Jing Bie in the residual multisource image set and have the same identification and the image resolution within the same preset resolution interval into second candidate deviation rectifying units; And performing constraint reassignment on the first candidate deviation rectifying unit and the second candidate deviation rectifying unit to divide the units to be rectified.
  3. 3. The machine learning based multiple equipartition correction method according to claim 2, wherein performing constraint reassignment includes: marking candidate deviation rectifying units, of which the sample numbers are lower than the lower limit of the preset sample numbers, in the first candidate deviation rectifying unit and the second candidate deviation rectifying unit as units to be redistributed; Calculating the imaging condition similarity of each image in the unit to be redistributed and other candidate deviation rectifying units, and if the imaging condition similarity between any unit to be redistributed and any candidate deviation rectifying unit is higher than a preset similarity threshold value, merging the image into the corresponding candidate deviation rectifying unit to obtain an initial unit to be rectified; and according to the acquisition time span and the resolution ratio distribution of each initial unit to be rectified, sub-candidate rectification units are divided again to obtain the units to be rectified.
  4. 4. The machine learning based multiple equipartition correction method according to claim 3, wherein calculating the similarity of the imaging conditions of each image in the unit to be reassigned and other candidate correction units comprises: Taking the difference value between the brightness of each image in the unit to be redistributed and the brightness average value of the images in other candidate deviation correcting units as the brightness average value deviation value; respectively calculating the overlapping degree of each image in the unit to be redistributed and the acquisition time window overlapping degree of other candidate deviation correcting units and the overlapping degree of the resolution ratio interval; And taking the overlapping degree of the time window and the overlapping degree of the resolution interval as similarity basic items, taking the brightness average deviation amount as a similarity reduction item, calculating a comprehensive similarity score, and taking the comprehensive similarity score as the similarity of imaging conditions.
  5. 5. The machine learning based multiple equipartition correction method according to claim 3, wherein performing sub-candidate correction unit subdivision includes: If the span of the acquisition time of any initial unit to be rectified exceeds the upper limit of the time window, splitting the initial unit to be rectified into a plurality of first sub candidate rectification units according to the acquisition time; If the distribution of the resolution of any initial unit to be rectified spans more than two resolution intervals, splitting the initial unit to be rectified into a plurality of second sub candidate rectification units according to the resolution; and carrying out intersection operation on the first sub candidate deviation rectifying unit and the second sub candidate deviation rectifying unit which belong to the same initial deviation rectifying unit, and taking each intersection operation result as the deviation rectifying unit.
  6. 6. The machine learning-based multiple equipartition deviation correcting method according to claim 1, wherein determining the deviation index of each unit to be corrected in step S3 includes: comparing the characteristic distribution description of each unit to be rectified with the reference characteristic distribution description of a preset comparison sample so as to calculate the offset index of each unit to be rectified, wherein the acquisition method of the comparison sample comprises the following steps: Selecting an image set which is in a preset target time window and has contrast and brightness meeting the preset imaging quality requirement from a multi-source image set, calculating statistics of image features based on the image set, describing the statistics of the image features as reference feature distribution, and taking the corresponding image set as a comparison sample.
  7. 7. The machine learning-based multiple average deviation rectifying method according to claim 1, wherein performing average deviation rectifying in step S3 includes: performing layering division on the offset index according to the brightness offset increment size and the offset channel duty ratio of the offset index of each unit to be rectified; If the offset index of any unit to be rectified is divided into global offset levels, uniformly applying corresponding brightness offset increments and channel offset increments to all pixels of each image in the unit to be rectified according to the brightness offset increments and the channel offset increments of the offset index of the unit to be rectified, and performing overall brightness translation and channel proportional recalibration on the uniform distribution result to obtain a global offset rectification result; If the offset index of any unit to be rectified is divided into local offset levels, according to the local texture complexity of each image in the unit to be rectified, the increment of the offset index is redistributed to the corresponding local offset area in each image according to the local texture complexity weight, and brightness adjustment and contrast enhancement are carried out on the redistributed result to obtain a local offset rectification result.
  8. 8. The machine learning based multiple equipartition deviation correcting method according to claim 7, wherein the determining method of the local deviation area comprises: Marking a pixel neighborhood with local texture complexity exceeding a preset texture complexity threshold value in each image in the unit to be rectified as a texture candidate area; If the deviation between the local brightness average value and the contrast of any texture candidate region and the corresponding statistic corresponding to the integral image characteristic of the unit to be rectified exceeds a preset deviation threshold value, marking the texture candidate region as a local deviation region.
  9. 9. The machine learning-based multiple equipartition correction method according to claim 1, wherein the identifying of the corrected feature shift condition of each image in step S4 includes: if the classification result of any image before and after correction changes and the classification confidence of the predicted class of the image after correction is lower than a preset confidence threshold, marking the image as a class-unstable image; Based on the feature vector output by the middle layer of the image classification model, obtaining feature vector difference values before and after correction, taking the distance measurement between the feature vector difference values before and after correction and the feature distribution description of the corresponding unit to be corrected as a residual feature offset index, and marking the image as a feature abnormal image if the residual feature offset index of any image exceeds a preset feature offset threshold; And executing intersection operation on the category unstable images and the characteristic abnormal images, and taking an intersection operation result as an image with the characteristic deviation after deviation correction.
  10. 10. A multiple equipartition deviation correcting system based on machine learning, characterized in that it is used for executing the multiple equipartition deviation correcting method based on machine learning according to claim 1, the multiple equipartition deviation correcting system based on machine learning comprises: The data acquisition module is used for acquiring a multi-source image set and dividing the multi-source image set into a plurality of units to be rectified; the feature aggregation module is used for aggregating the feature distribution of the images in the same unit to be rectified and constructing feature distribution descriptions corresponding to the units to be rectified according to the aggregation result; The device comprises a equipartition correction module, a correction module and a correction module, wherein the equipartition correction module is used for comparing the characteristic distribution description of each unit to be corrected and determining the offset index of each unit to be corrected; The image classification module is used for inputting the average deviation correction result into a preset image classification model to generate classification results of the images; And the iterative optimization module is used for iteratively executing equipartition deviation correction by taking the image with the characteristic deviation after deviation correction as a comparison sample of the characteristic distribution description until the deviation index in each unit to be corrected is converged into the target equipartition interval.

Description

Multiple equipartition deviation correcting method and system based on machine learning Technical Field The invention relates to the technical field of image processing, in particular to a multiple equipartition deviation correcting method and system based on machine learning. Background In the field of image recognition and image processing, recognition accuracy and feature extraction effects depend to a large extent on the quality of the image data, the consistency of distribution, and the balance between different acquisition sources. However, in practical applications, the images often come from multiple types of devices, multiple sampling conditions, and multiple imaging environments, such as cameras with different resolutions, monitoring devices under different illumination conditions, mobile terminals with different angles and shooting distances, and the like. These factors cause significant differences in brightness, contrast, color channel distribution, texture details, noise level, etc. of the images, forming multiple offsets such as batch offset, equipment offset, scene offset, etc., so that similar images show uneven distribution among different data sources. The traditional deviation correcting method generally adopts a normalization strategy of global mean value stretching, histogram equalization, white balance adjustment or fixed rules to correct the overall brightness, color proportion or local texture of the image. Although the offset problem can be relieved to a certain extent in the prior art, when the multi-batch image style change and the nonlinear offset relation are faced, only the global distribution is corrected, and multi-dimensional deviations such as texture, structural details, local area brightness and the like are difficult to consider. Disclosure of Invention Accordingly, the present invention is directed to a multiple average deviation rectifying method and system based on machine learning, so as to solve at least one of the above-mentioned problems. In order to achieve the above purpose, a multiple equipartition deviation correcting method based on machine learning comprises the following steps: s1, acquiring a multi-source image set, and dividing the multi-source image set into a plurality of units to be rectified; step S2, aggregating the characteristic distribution of the images in the same unit to be rectified, and constructing the characteristic distribution description corresponding to each unit to be rectified according to the aggregation result; S3, comparing the characteristic distribution description of each unit to be rectified, and determining the offset index of each unit to be rectified; Step S4, inputting the average deviation correction result into a preset image classification model to generate classification results of the images; And S5, taking the image with the characteristic deviation after deviation correction as a comparison sample of the characteristic distribution description, and iteratively executing equipartition deviation correction until the deviation index in each unit to be corrected is converged into the target equipartition interval. Aiming at the multiple offset problems of images in the aspects of brightness, contrast, color channel distribution, texture details and the like under the acquisition conditions of multiple devices, multiple scenes and multiple batches, the application introduces a feature distribution modeling and layering equipartition correction mechanism divided according to units to be corrected, is not limited to the traditional coarse granularity processing of executing unified histogram equalization or fixed normalization parameters on the whole library image, groups the images according to acquisition devices, time windows, scenes and resolutions, and describes brightness offset increment, channel offset increment and texture offset degree of each group relative to the reference feature distribution through offset indexes, and then selects global or local correction strategies according to increment size of the offset indexes and offset component proportion layering, thereby realizing the collaborative adjustment of brightness translation, channel proportion recalibration and local texture enhancement on different levels. The method has the advantages that images are input into an image classification model which is trained in advance and comprises an intermediate characteristic layer after correction, classification results before and after correction are compared with intermediate characteristic vector changes, and an unstable-class image and a characteristic abnormal image are taken as abnormal samples to be brought into an iterative correction closed loop, so that correction parameters can gradually approach a target uniform partition area in multiple iterations, excessive trowelling of key textures and structural details is avoided while batch deviation, equipment deviation and scene deviation are restrained, th