激光与光电子学进展, 2021, 58 (2): 0215003, 网络出版: 2021-01-11
一种基于多尺度特征融合的目标检测算法 下载: 1419次
Multiscale Feature Fusion-Based Object Detection Algorithm
机器视觉 卷积神经网络 目标检测 特征金字塔 特征融合 machine vision convolution neural network object detection feature pyramid feature fusion
摘要
基于深度学习的目标检测器RetinaNet和Libra RetinaNet均是使用特征金字塔网络融合多尺度特征,但上述两个检测器存在特征融合不充分的问题。鉴于此,提出一种多尺度特征融合算法。该算法是在Libra RetinaNet的基础上进一步扩展,通过建立两条自底向上的路径构建两个独立的特征融合模块,并将两个模块产生的结果与原始预测特征融合,以此提高检测器的精度。将多尺度特征融合模块与Libra RetinaNet结合构建目标检测器并在不同的数据集上进行实验。实验结果表明,与Libra RetinaNet检测器相比,加入模块后的检测器在PASCAL VOC数据集和MSCOCO数据集上的平均精度分别提高2.2个百分点和1.3个百分点。
Abstract
The RetinaNet and Libra RetinaNet object detectors based on deep learning employ feature pyramid networks to fuse multiscale features. However, insufficient feature fusion is problematic in these detectors. In this paper, a multiscale feature fusion algorithm is proposed. The proposed algorithm is extended based on Libra RetinaNet. Two independent feature fusion modules are constructed by establishing two bottom-up paths, and the results generated by the two modules are fused with the original predicted features to improve the accuracy of the detector. The multiscale feature fusion module and Libra RetinaNet are combined to build a target detector and conduct experiments on different datasets. Experimental results demonstrate that the average accuracy of the added module detector on PASCAL VOC and MSCOCO datasets is improved by 2.2 and 1.3 percentage, respectively, compared to the Libra RetinaNet detector.
张涛, 张乐. 一种基于多尺度特征融合的目标检测算法[J]. 激光与光电子学进展, 2021, 58(2): 0215003. Tao Zhang, Le Zhang. Multiscale Feature Fusion-Based Object Detection Algorithm[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215003.