首页|期刊导航|电讯技术|基于特征融合和对齐的领域自适应DINO

基于特征融合和对齐的领域自适应DINOOA

Domain Adaptive DINO with Feature Fusion and Alignment

中文摘要英文摘要

基于提升去噪锚框的检测 Transformer(Detection Transformer with Improved Denoising Anchor Boxes,DINO)是近年来新兴的检测器之一,其凭借卓越的检测性能获得了广泛关注.然而,DINO 作为一个监督模型难以从源域数据中学习到通用特征表示,其跨域泛化能力仍然面临巨大挑战,在处理具有领域差异的数据时性能会显著下降.为了提升 DINO 模型的泛化能力,提出了一种基于特征融合和对齐的领域自适应 DINO 算法.首先在骨干网络提出多尺度特征融合和全局特征对齐策略,提升模型对目标场景的整体适应性,然后在编码器提出局部特征对齐策略,缓解场景布局等因素对模型跨域性能的影响,最后在解码器提出层次特征融合和目标级对齐策略,提升模型的泛化和判别性能.所提方法在不同跨域任务中将目标领域检测的平均精度(mean Average Precision,mAP)提高了1.5%~17.4%.

Detection Transformer with improved denoising anchor boxes(DINO)is one of the emerging detectors in recent years and has gained considerable attention for its outstanding detection performance.However,DINO is difficult to learn generic representations from the source domain samples as a supervised model.Its cross-domain generalization capability still faces significant challenges and it may suffer a performance degradation when dealing with data that has domain shift.To improve the generalization ability of the DINO model,a domain adaptive DINO framework based on feature fusion and alignment is proposed.Firstly,it introduces multi-scale feature fusion and global feature alignment strategies in the backbone to boost the model's adaptability to target scenes.Then,a local feature alignment strategy is proposed in the encoder to mitigate the impact of factors like scene layout on the model's cross-domain performance.Finally,hierarchical feature fusion and object-level alignment strategies are introduced in the decoder to improve the model's generalization and discriminative capability.The proposed method improves the mean average precision(mAP)in the target domain by 1.5%to 17.4%across various cross-domain detection tasks.

朱思翰

西南电子技术研究所,成都 610036||复杂航空系统仿真全国重点实验室,成都 610036

信息技术与安全科学

目标检测领域自适应深度学习DINO检测器

object detectiondomain adaptationdeep learningDINO detector

《电讯技术》 2026 (9)

1436-1443,8

国家自然科学基金青年科学基金(62303433)

10.20079/j.issn.1001-893x.250417001

评论