基于跳跃连接金字塔模型的小目标检测

PDF下载HTML

Title:

Skip feature pyramid network with a global receptive field for small object detection

作者:

单义^1,2, 杨金福^1,2, 武随烁^1,2, 许兵兵^1,2

1. 北京工业大学 信息学部, 北京 100124;
2. 计算智能与智能系统北京重点实验室, 北京 100124

Author(s):

SHAN Yi^1,2, YANG Jinfu^1,2, WU Suishuo^1,2, XU Bingbing^1,2

1. Beijing University of Technology, Faculty of Information Technology, Beijing 100124, China;
2. Beijing Key Laboratory of Computational Intelligence and Intelligence System, Beijing 100124, China

关键词:

跳跃连接金字塔; 全局感受野; 目标检测; 深度学习; 特征提取; 卷积神经网络; 空洞卷积; 图像处理

Keywords:

skip feature pyramid network; global receptive field; object detection; deep learning; feature extraction; convolutional neural network; dilated convolution; image processing

分类号:

TP183

DOI:

10.11992/tis.201905041

摘要:

随着深度学习的发展，目标检测已经获得了较高的精度和效率。但是小目标的检测仍然是一个挑战。小目标检测准确率较低的重要原因是没有充分利用高层特征的语义信息和低层特征的细节信息之间的关系。针对上述问题，本文提出一种基于跳跃连接金字塔模型的小目标检测方法。与其他的目标检测方法不同，本文提出利用跳跃连接金字塔结构来融合多层高层语义特征信息和低层特征图的细节信息。而且为了更好地提取不同尺度物体对应的特征信息，在网络模型中采用不同大小的卷积核和不同步长的空洞卷积来提取全局特征信息。在PASCAL VOC和MS COCO数据集上进行了实验，验证了算法的有效性。

Abstract:

With the development of deep learning, objects can be detected with high accuracy and efficiency. However, the detection of small objects remains challenging. The main reason for this is that the relationship between high-level semantic information and low-level feature maps is not fully utilized. To solve this problem, we propose a novel detection framework, called the skip feature pyramid network with a global receptive field, to improve the ability to detect small objects. Unlike previous detection architectures, the skip feature pyramid architecture fuses high-level semantic information with low-level feature maps to obtain detailed information. To extract global information from a network, we apply a global receptive field (GRF) with convolution kernels of different sizes and different dilated convolution steps. The experimental results on PASCAL VOC and MS COCO datasets show that the proposed approach realizes significant improvements over other comparable detection models.