Camouflage Target Detection Method with Mutual Compensation of Local-Global Features
In the field of camouflage object detection (COD), the latest proposed methods mainly use the local features of the camouflaged target to complete the COD task. The output prediction map has problems such as rough target contours and incomplete objects. In response to the above problems, this paper...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Journal of Computer Engineering and Applications Beijing Co., Ltd., Science Press
2025-02-01
|
| Series: | Jisuanji kexue yu tansuo |
| Subjects: | |
| Online Access: | http://fcst.ceaj.org/fileup/1673-9418/PDF/2311097.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the field of camouflage object detection (COD), the latest proposed methods mainly use the local features of the camouflaged target to complete the COD task. The output prediction map has problems such as rough target contours and incomplete objects. In response to the above problems, this paper proposes a camouflaged target detection method based on local-global feature mutual compensation, which uses local features and global features to compensate for each other to detect camouflaged targets. Firstly, a non-local feature enhancement module (N-LFEM) is designed to use a non-local mechanism to interact with adjacent local areas and enhance local feature expression. Then, a local-global feature interaction module (L-GFIM) is constructed to average local features to obtain global features, and perform mutual compensation of local features and global features. Finally, a local-global feature cross-covariance module (L-GFCCM) is designed to obtain spatial indicators through semantic alignment and cross-covariance to locate the area where the camouflaged target is located, and select the feature map with the highest similarity to output. Experiments on 3 public datasets show that this algorithm is better than the other 8 latest models. On the COD10K dataset, the mean absolute error (MAE) reaches 0.028. |
|---|---|
| ISSN: | 1673-9418 |