高级检索
    黄继鹏, 史颖欢, 高阳. 面向小目标的多尺度Faster-RCNN检测算法[J]. 计算机研究与发展, 2019, 56(2): 319-327. DOI: 10.7544/issn1000-1239.2019.20170749
    引用本文: 黄继鹏, 史颖欢, 高阳. 面向小目标的多尺度Faster-RCNN检测算法[J]. 计算机研究与发展, 2019, 56(2): 319-327. DOI: 10.7544/issn1000-1239.2019.20170749
    Huang Jipeng, Shi Yinghuan, Gao Yang. Multi-Scale Faster-RCNN Algorithm for Small Object Detection[J]. Journal of Computer Research and Development, 2019, 56(2): 319-327. DOI: 10.7544/issn1000-1239.2019.20170749
    Citation: Huang Jipeng, Shi Yinghuan, Gao Yang. Multi-Scale Faster-RCNN Algorithm for Small Object Detection[J]. Journal of Computer Research and Development, 2019, 56(2): 319-327. DOI: 10.7544/issn1000-1239.2019.20170749

    面向小目标的多尺度Faster-RCNN检测算法

    Multi-Scale Faster-RCNN Algorithm for Small Object Detection

    • 摘要: 小目标是指图像中覆盖区域较小的一类目标.与常规目标相比,小目标信息量少,训练数据难以标记,这导致通用的目标检测方法对小目标的检测效果不好,而专门为小目标设计的检测方法往往复杂度过高或不具有通用性.在分析现有目标检测方法的基础上,提出了一种面向小目标的多尺度快速区域卷积神经网络(faster-regions with convolutional neural network, Faster-RCNN)检测算法.根据卷积神经网络的特性,修改了Faster-RCNN的网络结构,使网络可以同时使用低层和高层的特征进行多尺度目标检测,提升了以低层特征为主要检测依据的小目标检测任务的精度.同时,针对训练数据难以标记的问题,使用从搜索引擎上获取的数据来训练模型.因为这些训练数据与任务测试数据分布不同,又利用下采样和上采样的方法对目标高分辨率的训练图像进行转化,使训练图像和测试图像的特征分布更类似.实验结果表明:所提出的方法在小目标检测任务上的平均精度均值(mean average precision, mAP)可以比原始的Faster-RCNN提高约5%.

       

      Abstract: Normally, small object is the object which only covers a small part of a whole image. Compared with regular object, small object has less information and the training data of small object is difficult to be marked. This leads to the poor performance when directly employing the previous object detection methods for small object detection. Moreover, the detection methods designed for small object are often too complex or not generic. In this paper, we propose a small object detection algorithm named multi-scale Faster-RCNN. According to the characteristics of convolutional neural network, the structure of Faster-RCNN is modified, such that the network can integrate both the low-level and high-level features for multi-scale object detection. Through such a manner, the accuracy of small object detection is improved. Simultaneously, with the goal of solving the problem that training data is difficult to be marked, we use training data crawled from search engine to train the model. Because the distribution of crawled data is different from the real test data’s, training images in which objects have high resolution are transformed by means of down sampling and up sampling. It makes the feature distribution of training images and test images more similar. The experiment results show that the mean average precision (mAP) of proposed approach can be up to 5% higher than the original Faster-RCNN’s in the task of small object detection.

       

    /

    返回文章
    返回