【摘要】 目标检测是计算机视觉领域中的一个重要研究方向,针对目标检测算法中存在的模型庞大、多尺度目标检测等问题,基于HRNet (high resolution net)和自适应空间特征融合(adaptively spatial feature fusion, ASFF)提出一种多尺度特征融合目标检测算法.首先,利用通道拆分(channel split)操作和深度可分离卷积(depthwise separable convolution, Dwconv)改进HRNet的基础模块,结合CSPNet改进HRNet的分支结构,减少模型的参数量,在得到轻量化L-HRNet三个分支后使用空间特征金字塔EESP (extremely efficient spatial pyramid)模块获得不同感受野大小特征,并将其融合后加强特征;其次,使用ASFF模块自适应融合EESP模块输出多尺度特征,该模块为3个分支的特征分配不同的特征融合权重,自适应融合重要的空间特征;最后,引入SIoU (shape-aware IoU)作为边界框定位损失函数,综合考量边界框回归之间的角度关系、中心点距离关系以及边界框的形状关系,使得预测框与真实框之间的损失度量更加准确,整体参数量为5.7 M,在公开数据集PASCAL VOC上达到了85.1%的mAP,在MS COCO上的实验结果表明, mAP0.5-0.95达到了38.7%,在模型参数量较少的同时保持了较高的检测性能.更多还原

【Abstract】 Object detection is an important research direction in the field of computer vision. To address the challenges associated with complex models and multi-scale object detection in object detection algorithms, a multi-scale feature fusion object detection algorithm based on HRNet and ASFF is proposed. Firstly, the basic module of HRNet is improved by channel split operation and Dwconv, and the branch structure of HRNet is improved in combination with CSPNet to reduce the number of model parameters. After improving the three branches of the lightweight L-HRNet, the EESP module is adopted to obtain features of different receptive field sizes, and the features are further enhanced by fusion.Secondly, the ASFF module is adopted to adaptively fuse the multi-scale features output by the EESP module. This module assigns different spatial weights to features of the three branches and adaptively fuses important spatial features.Finally, shape-aware IoU(SIoU) is introduced as the bounding box localization loss function, which comprehensively considers angle relationship, center point distance relationship, and shape relationship between bounding box regressions,making the loss measurement between predicted boxes and ground truth boxes more accurate. The overall number of parameters is 5.7M, achieving 85.1 % mAP on the PASCAL VOC public dataset. Experimental results on MS COCO2017 show that the m AP0.5-0.95 reaches 38.7 %, maintaining high detection performance with fewer model parameters.更多还原