目标分割、模型训练方法和装置，及存储介质
By combining background filter frames and supervision information from deep networks, the problem of selecting supervision information in video target segmentation is solved, achieving higher segmentation accuracy and target recognition, and improving the model's representation ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JD DIGITS HAIYI INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2023-04-24
- Publication Date
- 2026-07-17
AI Technical Summary
In existing technologies, the accuracy of video target segmentation is affected by factors such as changes in target appearance and scale, occlusion, and disappearance. Furthermore, models based on online few-shot learning face difficulties in selecting and generating supervised information, which limits segmentation performance.
By introducing background filter frames and combining pixel-level matching and few-shot learning modules, richer foreground target information is obtained, reducing inter-domain differences. Deep network encoding of background filter frames provides more representative supervision information, and an auxiliary decoder is used to guide feature extraction, thereby improving the accuracy of the target segmentation model.
It improves the accuracy of video target segmentation, enhances target recognition, reduces background interference, and improves the representation ability and segmentation performance of the target model.
Smart Images
Figure CN116612473B_ABST