Live-action photo instance segmentation, classification and verification method

By constructing a lightweight cascaded network architecture and spatial mapping transformation, the problems of high computational cost and low recognition accuracy in real-world photo instance segmentation are solved, achieving efficient instance segmentation and automated verification of vector data.

CN122176449APending Publication Date: 2026-06-09CHINA AERO GEOPHYSICAL SURVEY & REMOTE SENSING CENT FOR LAND & RESOURCES
4 Cites 1 Cited by

Patent Information

Application Number
CN202610535240.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-22
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing technologies for instance segmentation of real-world photos suffer from problems such as high computational cost and low efficiency due to excessive model parameters, difficulty in extracting fine-grained features of ground objects in complex scenes, and lack of effective correlation and verification mechanisms between instance segmentation results and spatial attributes of vector data.

Method used

A lightweight cascaded network architecture is constructed, and detection boxes are used to guide segmentation and freeze the parameters of the large model encoder to enhance the ability to extract fine-grained features of highly similar ground features and occluded targets. Furthermore, a mapping relationship between vector data and real-world images is established through spatial mapping transformation to achieve automated closed-loop verification between heterogeneous data.

Benefits of technology

It significantly reduces the memory usage and time cost of model training and inference, improves recognition accuracy and operational efficiency, and solves the problems of false alarms and misreports in verification.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention provides a live-action photo instance segmentation and classification verification method, and relates to the technical field of computer vision. The method comprises the following steps: constructing a live-action verification data set according to spatial distribution of to-be-verified vector data; establishing an interpretation mark feature library containing texture and geometric features; constructing a cascade network architecture consisting of a target detection sub-network and a segmentation sub-network, guiding segmentation operation by using a detection frame, freezing segmentation encoder parameters in a training stage, and embedding interpretation mark features into a network trunk to obtain a lightweight cascade instance segmentation model; and projecting the vector data to an image coordinate system by using a space mapping transformation algorithm to obtain a theoretical projection area, and carrying out space overlapping degree and attribute consistency comparison on a model segmentation result and the theoretical projection area to generate classification verification correction information. According to the method, the problem that an effective association verification mechanism is lacked among heterogeneous data is solved, and automation and high efficiency of ground feature verification are realized.
Need to check novelty before this filing date? Find Prior Art