CN-121982715-A - Scene annotation judging method based on flow

CN121982715ACN 121982715 ACN121982715 ACN 121982715ACN-121982715-A

Abstract

The invention discloses a scene annotation judging method based on a process, which relates to the technical field of scene annotation, and comprises the steps of feature index construction, scene annotation analysis, scene annotation judgment and early warning prompt, wherein the index is constructed through the VPR features of a constructed sub-map, so as to execute the geographic priori corresponding to a target scene, search and obtain each candidate reference image matched with a scene image corresponding to the target scene, and execute the analysis judgment between the scene image and each candidate reference image, based on the judging result, the geographic reordering of candidate reference images corresponding to the target scene is respectively executed, the final positioning result of the target scene and the flow optimization of the corresponding annotation of the analysis target scene are confirmed, so that the accurate and efficient annotation of the corresponding flow of the scene is realized, the problem of mismatching of different scenes caused by the lack of semantic constraint in the current annotation is intelligently solved, and the scale and resource consumption of online retrieval are effectively reduced.

Inventors

HU XIAOBIN
LI MENGZHE
WANG HAIRONG
LV XIAOBAO

Assignees

曙光天玑数据科技(江苏)有限公司
中科曙光南京研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. The scene annotation judging method based on the flow is characterized by comprising the following steps: Step one, constructing a feature index, namely constructing a VPR feature construction index of each sub map by collecting each historical geotag image with GPS coordinates, and simultaneously acquiring scene image data of a target scene based on equipment multiple sources; step two, scene annotation analysis, namely executing geographic priori corresponding to the target scene based on the acquired scene image corresponding to the target scene, and searching to obtain candidate reference images matched with the scene image corresponding to the target scene; Judging scene annotation, namely judging the matching degree of each candidate reference image matched with the scene image corresponding to the target scene, executing geographic reordering of the candidate reference images corresponding to the target scene when the matching degree of the candidate reference images matched with the scene image corresponding to the target scene is qualified, analyzing a final positioning result corresponding to the target scene, acquiring feedback data in the labeling process corresponding to the target scene when the matching degree of the candidate reference images matched with the scene image corresponding to the target scene is unqualified, and analyzing flow optimization of labeling corresponding to the target scene; and fourthly, early warning prompt is carried out when the matching degree of the candidate reference images matched with the scene images corresponding to the target scene is unqualified.
2. The scene annotation judging method based on the process of claim 1, wherein the constructing the VPR feature construction index of each sub map comprises the following specific construction processes: Based on collecting each historical geotag image with GPS coordinates, utilizing DBSCAN clustering algorithm, dividing the library into a plurality of sub-maps with compact space according to the levels of the GPS coordinates and scene class based on the images in the image library, extracting visual feature vectors from all reference images in each sub-map by using a VPR model, and finally using FAISS retrieval tools to construct VPR feature indexes of each sub-map.
3. The scene annotation judging method based on the process of claim 1, wherein the device-based multi-source acquisition of scene image data of the target scene comprises the following specific acquisition processes: scene image data corresponding to a target scene is monitored and obtained in the multi-source direction and angle of a vehicle-mounted camera, unmanned aerial vehicle aerial photography, satellite remote sensing and monitoring equipment, image preprocessing is carried out on the scene image data, a data format is unified, the resolution and the pixel depth are included, and a scene image in a unified standard format is obtained.
4. The scene annotation judging method based on the process according to claim 1, wherein the geographic priori corresponding to the execution target scene is specifically executed as follows: Inputting a scene image corresponding to the target scene into VLMS, further obtaining VLMS rough GPS coordinates of the scene image corresponding to the target scene by predicting a high-performance large language model, extracting macroscopic geographic features in the scene image corresponding to the target scene, further executing scene large category labeling on the scene image corresponding to the target scene according to the extracted features, and simultaneously identifying and outputting all display characters in the scene image corresponding to the target scene by utilizing an OCR technology.
5. The scene annotation judging method based on the process of claim 1, wherein the searching obtains each candidate reference image matched with the scene image corresponding to the target scene by the following specific searching process: Based on the VPR model which is the same as the reference set, further performing visual feature extraction on the scene image corresponding to the target scene to obtain visual feature data of the scene image corresponding to the target scene, comparing the visual feature data of the scene image corresponding to the target scene with the reference visual feature data set of each sub-map VPR feature construction index, and if the visual feature data of the scene image corresponding to the target scene is contained in the reference visual feature data set of a certain sub-map VPR feature construction index, taking the reference image of the sub-map VPR feature construction index pair target as a candidate reference image matched with the scene image corresponding to the target scene, and performing the same so on to obtain each candidate reference image matched with the scene image corresponding to the target scene.
6. The scene annotation judging method based on the process of claim 1, wherein the matching degree judgment is carried out on each candidate reference image matched with the scene image corresponding to the target scene, and the specific judging process is as follows: Extracting a similarity value between a target scene corresponding scene image and each candidate reference image, comparing the similarity value between the target scene corresponding scene image and each candidate reference image with a preset reference similarity threshold interval, if the similarity value between the target scene corresponding scene image and a certain candidate reference image is not contained in the preset reference similarity threshold interval, judging that the matching degree of the scene image corresponding to the target scene and the candidate reference image is not qualified, otherwise, judging that the matching degree of the scene image corresponding to the target scene and the candidate reference image is qualified.
7. The method for determining scene annotation based on flow according to claim 1, wherein the performing the geographic reordering of the candidate reference images corresponding to the target scene is performed as follows: And extracting GPS coordinates and VLMS predicted coordinates corresponding to each candidate reference image from each candidate reference image matched with the scene image corresponding to the target scene under the condition that the matching degree is qualified, simultaneously obtaining the average radius of the earth, and respectively calculating the geographic distance between the GPS coordinates and VLMS predicted coordinates of each candidate reference image by utilizing a Ha Fuxin formula.
8. The scene annotation judging method based on the process of claim 1, wherein the final positioning result corresponding to the analysis target scene is as follows: And based on the geographic distances corresponding to the candidate reference images obtained through calculation, sequencing the geographic distances corresponding to the candidate reference images according to the sequence from near to far, and selecting the scene name and the corresponding GPS coordinates of the first candidate reference image ranked after sequencing as the final positioning result of the scene image corresponding to the target scene.
9. The flow-based scene annotation decision method of claim 1, wherein the feedback data comprises geographic a priori bias data and annotation standard fuzzy data.
10. The scene annotation judging method based on the process according to claim 1, wherein the process optimization of the corresponding annotation of the analysis target scene comprises the following specific analysis processes: And when the feedback data in the labeling process of the corresponding scene image of the target scene is labeling standard fuzzy data, judging that the reason of the matching abnormality of the corresponding scene image of the target scene is labeling standard fuzzy, optimizing the classification rule of scene classification or the quantization rule of feature labeling, and analyzing to obtain the flow optimization of the corresponding labeling of the target scene.

Description

Scene annotation judging method based on flow Technical Field The invention relates to the technical field of scene annotation, in particular to a scene annotation judging method based on a flow. Background Along with the rapid development of the internet, the number of network pictures is also rapidly increased, wherein the scene position labeling of pictures is greatly demanded in the analysis of the internet space information, but because of the characteristics of similar characteristics of ports of all airports, a single image classification-based algorithm cannot realize the picture labeling with high accuracy, so a scene labeling judging method based on a process is provided, and the efficient and accurate scene labeling is realized. The invention patent disclosed by the publication number CN107133325B is based on an Internet photo geographic space positioning method of a street view map, and comprises the steps of preprocessing a street view photo library, extracting and describing features, establishing a feature index, inquiring nearest neighbor features of each feature of a photo to be inquired according to the index, voting, trimming and smoothing voting results to obtain a most similar photo, defining a buffer area according to the distance between two known street view points by taking the most similar photo as a circle center, calculating similarity between the street view photo in the buffer area and the photo to be inquired, screening out a photo with high similarity as a similar photo set, extracting and matching features of the similar photo set and the photo to be inquired together, registering the photo by using an SFM algorithm, generating a relative position relation between sparse point cloud and a camera, calculating unknown external azimuth elements of the photo to be inquired according to the known street view point coordinates, and realizing positioning and gesture determination. Practice proves that the image positioning method provided by the invention can effectively and accurately position the electronic photo of any source of the Internet. The method mainly aims at realizing accurate positioning on the geographic space positioning of the Internet photo based on the feature index, does not carry out corresponding labeling on a scene image, has low accuracy, cannot realize high-accuracy picture labeling, lacks of intelligent distinguishing of large and small scene scenes, is single in algorithm, cannot realize high-efficiency labeling of the scenes, is based on the retrieval method, relies on a visual feature matching reference database, has the problems of poor expandability and perceived confusion on the global scale, converts positioning into a geographic cell classification task based on a classification method, but can cause poor generalization due to geographic space region unit division, and cannot guarantee accurate and high-efficiency labeling of the scenes. Disclosure of Invention Aiming at the technical defects, the invention aims to provide a scene annotation judging method based on a flow. The invention provides a scene annotation judging method based on a process, which comprises the following steps of firstly, constructing a feature index, namely, constructing VPR feature construction indexes of all sub-maps by collecting all historical geographic mark images with GPS coordinates, and simultaneously acquiring scene image data of a target scene based on equipment multiple sources. And secondly, scene annotation analysis, namely executing the geographic prior corresponding to the target scene based on the acquired scene image corresponding to the target scene, and searching to obtain each candidate reference image matched with the scene image corresponding to the target scene. And step three, scene annotation judgment, namely judging the matching degree of each candidate reference image matched with the scene image corresponding to the target scene, executing geographic reordering of the candidate reference images corresponding to the target scene when the matching degree of the candidate reference images matched with the scene image corresponding to the target scene is qualified, analyzing a final positioning result corresponding to the target scene, acquiring feedback data in the process of labeling corresponding to the target scene when the matching degree of the candidate reference images matched with the scene image corresponding to the target scene is unqualified, and analyzing flow optimization of labeling corresponding to the target scene. And fourthly, early warning prompt is carried out when the matching degree of the candidate reference images matched with the scene images corresponding to the target scene is unqualified. The scene annotation judging method based on the flow has the advantages that 1, the index is built through the VPR features of the sub-map, the geographic priori corresponding to the target scene is further executed, each candidate ref