High-precision weakly supervised target detection and segmentation method based on segmentation large model

By using a segmentation-based large model to generate high-quality pseudo-labels through spatial, instance, and semantic queries, the accuracy and speed issues of object detection and segmentation under weak supervision are solved, achieving high-precision target detection and segmentation results.

CN117635944BActive Publication Date: 2026-06-23HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HUAZHONG UNIV OF SCI & TECH
Filing Date
2023-12-04
Publication Date
2026-06-23

AI Technical Summary

Technical Problem

Existing weakly supervised object detection and segmentation methods struggle to achieve high-precision object localization and segmentation when utilizing limited annotation information, especially since image-level labels lack sufficient localization information, resulting in limited segmentation accuracy.

Method used

By designing a segmentation-based large model, suggestion boxes are generated using query hints based on spatial relevance, instance relevance, and semantic relevance. A dynamic bounding box pseudo-label generation method is constructed, and combined with a region of interest discarding strategy, a fully supervised object detection network is trained to generate high-quality pseudo-labels for training the segmentation-based large model.

Benefits of technology

It achieves high-precision target detection and segmentation, improves recall and segmentation accuracy, enhances robustness and training speed, and is suitable for object detection and segmentation tasks under weak supervision.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117635944B_ABST
    Figure CN117635944B_ABST
Patent Text Reader

Abstract

The application discloses a high-precision weakly supervised object detection and segmentation method based on a segmentation large model. The method first uses a classification clue to prompt the segmentation large model to generate a high-recall picture suggestion box, and then trains a weakly supervised object detection network based on the suggestion box. Meanwhile, the application also proposes a dynamic pseudo-label generation strategy to improve the quality of pseudo-labels generated by the weakly supervised network, and an interesting area / query discarding method to reduce the noise influence in the pseudo-labels. In addition, the method can be further extended to a weakly supervised object segmentation task, that is, a segmentation large model is prompted by a pseudo-label of weakly supervised detection to generate a high-quality instance pseudo-label. The instance pseudo-label can provide more detailed supervision information for the object segmentation network than previous weakly supervised object segmentation methods. The method is accurate and efficient, and achieves a precision and speed far exceeding other methods on multiple weakly supervised instance recognition benchmarks.
Need to check novelty before this filing date? Find Prior Art