隔级融合特征金字塔与CornerNet相结合的小目标检测

PDF下载HTML

Title:

Small target detection based on a combination of feature pyramid and CornerNet

作者:

赵文清^1,3, 孔子旭¹, 赵振兵²

1. 华北电力大学 控制与计算机工程学院，河北 保定 071003;
2. 华北电力大学 电气与电子工程学院，河北 保定 071003;
3. 复杂能源系统智能计算教育部工程研究中心，河北 保定 071003

Author(s):

ZHAO Wenqing^1,3, KONG Zixu¹, ZHAO Zhenbing²

1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. School of Electrical and Electronic Engineering, North China Electric Power University, Baoding 071003, China;
3. Engineering Research Ce

关键词:

CornerNet; 小目标检测; 卷积; 特征图; 隔级融合; 上下融合; 旁路连接; 特征金字塔

Keywords:

CornerNet; small target detection; convolution; feature map; interval fusion; upper and lower fusion; bypass connection; feature pyramid

分类号:

TP391

DOI:

10.11992/tis.202004033

摘要:

为弥补CornerNet中小目标语义信息弱的缺陷，提出隔级融合特征金字塔的方法，提高小目标平均准确率。 对骨干网络后半部分融合后的4个特征图进行提取，将尺寸较小的特征图进行2次卷积，得到2个新的特征图；运用上下融合、隔级融合和旁路连接的思想，生成融合后的特征图并将其组成特征金字塔。将改进后的算法与当前主流CornerNet、Faster RCNN、RetinaNet算法在MS COCO数据集上进行比较，结果表明，改进后算法在对小目标进行检测时，小目标平均准确率有较大提高。隔级融合特征金字塔在CornerNet上能有效融合高低层特征图，使融合后的特征图有较强的语义信息，提高CornerNet网络的小目标平均准确率。

Abstract:

To improve the problem of the weak semantic information of the small target in CornerNet, a method of the hierarchical fusion feature pyramid is proposed to increase the average accuracy of the small target. The method first extracts the four feature maps after the fusion of the second half of the backbone network, then convolves the feature maps with a smaller size twice to obtain two new feature maps, and finally uses the ideas of the upper and lower fusion, interlevel fusion, and bypass connection to generate a fused feature map and form it into a feature pyramid. The result shows that the average accuracy for small targets obtained by our algorithm has been greatly improved compared with those by current mainstream algorithms, such as CornerNet, Faster RCNN, and RetinaNet on the MS COCO dataset, which demonstrates great superiority. The inter-level fusion feature pyramid can effectively fuse high-level and low-level feature maps on CornerNet, so that the fused feature maps have strong semantic information, and improve the average accuracy of the small targets of the CornerNet network.