An efficient semi-automatic labeling method for remote sensing image target detection dataset construction
By introducing a pre-trained YOLOv8 model and HBM and SFS algorithms, the problems of high manpower requirements and low efficiency in the annotation process of remote sensing image target detection datasets are solved, realizing efficient and low-cost dataset production, which is suitable for high-quality annotation of large datasets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2024-04-15
- Publication Date
- 2026-07-21
AI Technical Summary
The existing remote sensing image target detection dataset annotation process suffers from problems such as high manpower requirements, long annotation time, difficulty in guaranteeing quality, and difficulty in efficiently processing the annotation of a large number of small targets.
A pre-trained YOLOv8 model is introduced for fully automatic annotation. Harmonic Background Modelling (HBM) and Seed-filling Foreground Segmentation (SFS) algorithms are used to supplement missed targets and delete erroneous annotations. The model is optimized through multiple rounds of iterative training to reduce manual intervention.
It significantly improves the efficiency of dataset annotation, reduces production costs and time, outputs high-quality large datasets, is suitable for current large model training, and frees up the workload of manual annotation.
Smart Images

Figure CN118411723B_ABST