基于特征对齐和特征融合的半监督目标检测算法  

Semi-supervised Object Detection Algorithm Based on Feature Alignment and Feature Fusion

作　　者：汤文兵[1] 李菲 TANG Wenbing;LI Fei(School of Computer Science and Engineering,Anhui University of Science and Technology,Anhui Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《重庆工商大学学报(自然科学版)》2025年第1期35-41,共7页Journal of Chongqing Technology and Business University:Natural Science Edition

摘　　要：目的针对半监督目标检测导致数据特征表示不充分,数据样本类不均衡等问题,提出一种基于特征对齐和特征融合的半监督目标检测方法。方法在常见的半监督目标检测框架中,伪标签是完全根据分类分数生成的,然而,高置信度预测并不总是保证准确的bbox定位。为了解决定位不准确问题和特征表示不充分问题,受Consistent Teacher中的FAM-3D算法启发,考虑分类和定位的最优特征可能在不同尺度上,引入T-head特征对齐头算法,在Unbiased Teacher V2中成功地将分类和定位分支进行对齐,并且引入ASFF,通过空间过滤冲突信息的方法来抑制不一致性,从而提高了特征的尺度不变性,实现特征在空间上的融合;通过学习不同特征图之间的联系来解决特征金字塔内部的不一致性问题。结果根据实验结果,改进的算法在COCO数据集、VOC数据集上都有一定的比例提升。结论改进的算法可以有效减轻数据表示不充分和数据样本类不均衡问题,同时也提高了算法的精度。Objective In response to issues such as insufficient data feature representation and imbalanced sample classes in semi-supervised object detection,a semi-supervised object detection method based on feature alignment and feature fusion was proposed.Methods In common semi-supervised object detection frameworks,pseudo-labels are generated solely based on classification scores.However,high-confidence predictions do not always fully guarantee accurate bbox positioning.In order to solve problems of inaccurate positioning and insufficient feature representation,inspired by the FAM-3D algorithm in the Consistent Teacher,considering that the optimal features for classification and positioning may be at different scales,the T-head feature alignment head algorithm was introduced and the classification and positioning branches were successfully aligned in Unbiased Teacher V2.Additionally,ASFF was introduced to suppress inconsistency by spatially filtering conflict information,thereby improving the scale invariance of features and achieving spatial fusion of features.The internal inconsistencies within the feature pyramid were addressed by learning the connections between different feature maps.Results According to experimental results,the improved algorithm demonstrated certain performance improvements on the COCO dataset and VOC dataset.Conclusion The proposed algorithm effectively alleviates issues of insufficient data representation and imbalanced sample classes while also enhancing algorithm accuracy.

关 键 词：目标检测 半监督学习 特征对齐 特征金字塔 ASFF 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]