目标检测中的特征融合方法  

Feature Fusion Method for Object Detection

作　　者：郑晗 储珺[1] ZHENG Han;CHU Jun(School of Software Engineering,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学软件学院,南昌330063

出　　处：《南昌航空大学学报（自然科学版）》2022年第4期59-67,共9页Journal of Nanchang Hangkong University(Natural Sciences)

基　　金：国家自然科学基金(62162045)。

摘　　要：随着硬件资源的开发,深度学习技术广泛应用于视觉任务,基于深度学习的目标检测方法已经成为主流。特征融合是检测框架中用来提升模型性能的普遍方式,目前还没有较为系统的总结工作。因此,以深度学习为背景,本文首先从局部和整体两个层面介绍了特征融合,在结构上按照层、流以及空间的融合形式分析了目标检测中典型的以及改进的特征融合方法。然后分析了骨干网络和颈部网络中常用的特征融合技术,如加深网络、扩大感受野、加权融合等。最后提出了特征融合未来的研究方向,分析了多模态融合、自适应融合以及注意力机制等发展前景,为后续工作提供一些有益的指导。With the development of hardware resources, deep learning technology is widely used in computer vision tasks. Object detection methods based on deep learning have become the mainstream. Feature fusion is a common way to improve model performance in object detection, and there is no systematic summary work yet. Therefore, with the background of deep learning, this paper firstly introduces feature fusion from the local and overall levels, and analyzes the typical and improved feature fusion methods in object detection according to the fusion form of layer, flow and space in structure. Then the commonly used feature fusion techniques in the backbone network and neck network are analyzed, such as deepening the network, expanding the receptive field, and weighted fusion. Finally, the future research direction of feature fusion is proposed, and the development prospects of multimodal fusion, adaptive fusion and attention mechanism are analyzed, which provides useful guidance for follow-up work.

关 键 词：目标检测 特征融合 深度学习 特征金字塔 多尺度特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]