摘要
小目标检测是计算机视觉领域具有挑战性的研究任务。针对小目标物体尺寸小、特征不明显、目标聚集等问题,提出了一种基于自适应特征融合和任务对齐的小目标检测算法C-SODNET。该算法在TOOD基础上进行优化与改进,引入ConvNeXt作为骨干网络,通过嵌入CBAM注意力机制和自适应特征融合模块的特征金字塔结构提升兴趣区域的特征提取能力,同时在检测头加入可变形卷积,显著改善了对于小目标物体的检测能力,最后引入CIoU回归损失函数来训练模型。实验结果表明,C-SODNET在VisDrone2019小目标检测数据集mAP_(50)为51.2%,相较于TOOD算法准确率提升了9.4%,小目标物体的精确率APs提高了7.3%,验证了算法的有效性。该算法可为高空或远距离场景小目标检测应用提供了有效解决方案。
Small target detection is a challenging research task in the field of computer vision.For the problems of small target object size,inconspicuous features and target aggregation,a small target detection algorithm C-SODNET based on adaptive feature fusion and task alignment is proposed.The algorithm is optimized and improved on the basis of TOOD by introducing ConvNeXt as the backbone network,improving the feature pyramid structure by embedding CBAM attention mechanism and adaptive feature fusion module feature extraction capability of the region of interest,while adding deformable convolution in the detection head significantly improves the detection capability for small target objects,and finally introducing CIoU regression loss function to train the model.The experimental results show that the mAP_(50) of C-SODNET in VisDrone2019 small target detection dataset is 51.2%,which improves the accuracy rate by 9.4%compared to the TOOD algorithm,and the accuracy rate APs of small target objects improves by 7.3%,which verifies the effectiveness of the algorithm.This algorithm can provide an effective solution for small target detection applications in high-altitude or long-range scenes.
作者
郑有凯
胡君红
田春欣
ZHENG Youkai;HU Junhong;TIAN Chunxin(College of Physical Science and Technology Central China Normal University,WuHan 430079,China)
出处
《激光杂志》
CAS
北大核心
2024年第2期36-42,共7页
Laser Journal
基金
国家自然科学基金(No.62101204)
湖北省自然科学基金(No.2020CFB474)。
关键词
小目标检测
注意力机制
任务对齐
通道注意力
自适应特征融合
small target detection
attention mechanism
task alignment
channel attention
adaptive feature fusion
