基于连续度的多示例稀疏分布目标检测改进算法.docVIP

  基于连续度的多示例稀疏分布目标检测改 进算法# 高婷，王爽，焦李成** 5 10 （西安电子科技大学智能感知与图像理解教育部重点实验室） 摘要：针对目标尺寸不同导致复杂背景下目标检测中的高虚警率问题，本文对基于多示例的 稀疏分布目标检测算法进行了改进，提出了一种基于连续度的目标检测算法。该算法与一般 的多尺度目标检测算法不同之处在于：本文算法通过对测试图像的合理缩放，利用图像中目 标在多个尺度上稳定一致的性质，通过参考目标的连续度给出最终的目标检测结果。该算法 基于目标的连续度使得目标可被准确检测，同时极大抑制了背景中的虚警。实验表明，本文 算法具有检测准确率高、虚警率低的优点，可用于高分辨复杂背景图像中的目标检测。 关键词：目标检测；多示例学习；目标连续度 中图分类号：TP751 15 An Object Detection Algorithm Based on Multi-Instance Sparse Distribution and the Consistence of Object GAO Ting, WANG Shuang, JIAO Licheng (Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of 20 25 30 35 40 China, Xidian University.) Abstract: To solve the high false alarming problem in complex background which involved by the different size of the objects , based on the multi-instance sparse distribution object detection algorithm , this article proposes an new algorithm based on the consistence of the object. The difference between this algorithm and other ordinary multi-scale object detection algorithms is that this algorithm gives the final result of the object detection according to the consistence of object in reasonable scaling of the test image which based on the nature that the object in the image is stable and consistent on several continual scales. This algorithm is based on the consistence of the object which makes the object to be accurately detected, meanwhile greatly inhibited the false alarming in the background. Experiments shows that the algorithm proposes in this paper which has advantages of high accuracy and low false alarming can be used in the image with large-scale、high resolution and complex background for object detection. Key words: Object detection; Multi-Instance Learning; Consistence of Object 0 引言 随着计算机互联网的广泛普及以及现代科学技术的高速发展，图像作为一种直观、信息 丰富、质量高的信息媒介，成为人类信息处理的一种最有效的形态。由此，数字图像处理作 为一种对图像信息进行分析处理的手段，俨然成为国内外学者研究的热点。其中，目标检测 和识别作为图像处理的一个重要分支，在智能交通系统、智能监控系统、军事目标检测及医 学导航手术中、雷达通信等方面具有广泛的应用价值。 目标识别，是一种基于目标的几何和统计特征的图像分割应用，它将目标的分割和识别 合二为一，其准确性和实时性是整个系统的一项重要