计算机与现代化

• 图像处理 • 上一篇    下一篇

一种基于SSD改进的目标检测算法

  

  1. (华北电力大学控制与计算机工程学院,北京102206)
  • 收稿日期:2019-08-13 出版日期:2020-03-03 发布日期:2020-03-03
  • 作者简介:苏蒙(1995-),男(壮),云南曲靖人,硕士研究生,研究方向:深度学习,计算机视觉,E-mail: wa97@outlook.com; 李为(1967-),女,教授,硕士,研究方向:智能电网软件技术,电力信息安全,E-mail: liwei@ncepu.edu.cn。
  • 基金资助:
    国家自然科学基金资助项目(61300132)

A Modified Object Detection Algorithm Based on SSD

  1. (School of Control and Computer Engineering, North China Electric Power University, Beijing 102206, China) 
  • Received:2019-08-13 Online:2020-03-03 Published:2020-03-03

摘要: SSD(Single Shot MultiBox Detector)是一种基于深度学习的目标检测算法,它作为当前最为主流的检测算法之一,在极大地提高检测速度的同时,还能保证一定的检测精度,但是仍难以满足实际应用的需求。本文在SSD模型的基础上,引入注意力机制,提出一种基于SSD改进的目标检测算法。注意力机制能够有效地提高卷积神经网络对图片特征的提取能力,从而进一步提高算法的检测精度。改进后的算法在Pascal VOC数据集上进行对比试验。实验结果表明,改进后的模型在Pascal VOC2007测试集上的检测精度达到78.5% mAP(mean Average Precision),比改进前提高4.2个百分点,在Pascal VOC2012测试集上的检测精度达到77.1% mAP,比改进前提高4.7个百分点。

关键词: 目标检测, SSD, 深度学习, 卷积神经网络, 注意力机制

Abstract: Single Shot MultiBox Detector(SSD), one of the most prevalent object detection algorithms based on deep learning, greatly shortens the time of object detecting, whose accuracy of detection is acceptable. However its accuracy cannot meet the need of practical application. Attention mechanism is helpful in improving the performance of convolutional neural network and the accuracy of detection. This paper modifies SSD with attention mechanism and proposes a modified object detection algorithm based on SSD. The modified SSD has been tested with Pascal VOC dataset. The modified SSD achieves 78.5% mAP in Pascal VOC2007 test set and 77.1% mAP in Pascal VOC2012 test set, which are higher than original SSD 4.2 percentage points and 4.7 percentage points separately.

Key words: object detection, SSD, deep learning, convolutional neural network, attention mechanism

中图分类号: