IOE OpenIR  > 光电技术研究所博硕士论文
基于深度学习的对地目标检测技术研究
李小宁
Subtype硕士
2019-05-28
Degree Grantor中国科学院大学
Place of Conferral中国科学院光电技术研究所
Keyword小目标检测 卷积神经网络 语义增强 扩展层预测 正则化
Abstract

目标识别与检测是计算机视觉、多媒体应用等领域重要的研究内容之一。目
标识别的任务是对给定的输入图像,确定是否包含指定目标,在多目标检测中,
还需判断出不同目标所属的类别。目标检测任务则不仅需要确定是否包含目标,
还需要给出目标准确的位置信息。现今,图像采集和传输技术的飞速发展产生了
海量的图像数据,自动识别图像中的物体这一技术在诸多场景应用中至关重要。
在行为分析、语义理解等高层视觉处理与分析任务中,精准检测出目标是其重要
基础。此外,在视频实时监控城市公共安全,无人机及卫星航拍图像检测道路等
场景中,目标检测技术均得到了广泛应用。尽管该领域目前已取得许多突破,但
依然面临诸多挑战。如在图像背景复杂、光照不均、目标模糊、尺度过小、以及
目标被遮挡等的情况下,很难获取到理想的目标辨识效果。卷积神经网络在目标
检测中的应用使得模式识别的模型复杂度提升,也由此带来了计算成本的增加。研究实时性、精确性、稳定性较高的目标识别与检测算法成为了当前的热点问题。
 
本文从目标识别与检测的实际问题出发,围绕地面小尺寸目标精准定位的问
题,结合计算机视觉、深度学习、机器学习的相关算法进行了深入的研究。具体
研究内容包括:从目标特征表达和提取,目标预测框生成、提高定位精度等的角
度研究了不同光照、角度变化下的地面小目标检测问题,主要贡献如下: 

一、尽管目前基于深度学习的目标检测算法对于常规尺寸目标的取得了较好
的检测结果,但由于地面目标尺度较小,外观信息较少,图像背景复杂等的原因
导致小目标检测存在精度低、定位困难等问题。本文对几种经典的深度学习目标
检测框架针对小目标检测进行了实验分析与对比,从目标特征提取的角度提出了
一种基于特征融合的子网络来获取增强语义的小目标特征。该网络利用了多个层
次的深度特征图信息,构建了融合特征层,作为小目标预测网络的输入。对比许
多检测算法仅利用高层特征的信息表达,而缺失了对小目标而言较为关键的局部
细节信息,该方法有效地提升了小目标的特征表达。

二、针对尺度、角度等的变换问题,本文设计了一种基于融合层的扩展层预
测子网络,在扩展层的多个尺度空间内匹配目标,将每个层次的预测值与真实值
的偏差加权和作为损失函数训练模型,有效地提高了小尺寸目标的定位精度。 

三、深度学习检测模型虽有强大的表征能力,但随着模型复杂度的爆发式增
长,容易导致过拟合的问题。针对这一问题,本文从增加训练集量级、降低模型
复杂度的角度出发,做了一系列的改进以提高模型的泛化能力,包括数据集的增 广,BN(Batch Normalization)层归一化处理,L2 正则化操作等。这些工作进一步提升了检测模型的识别精度,同时加快了模型收敛速度。在两个公开数据集的实验表明,本文提出的小目标检测模型展现出较为明显的优势。 

Other Abstract

Object recognition and detection is an important research issue in the fields of computer vision, multimedia application and others. The task of object recognition is to judge whether a specified object is included for a given input image. For multi-class objects detection, it is also required to output the category of every object. 

In addition, the other fundamental task of object detection is to identify the accurate location of objects in image. Nowadays, a huge amount of image data was generated for the rapid development of image acquisition and transmission technology. The technology of automatically recognizing objects in images plays an important role in many applications in tasks of high-level visual processing and analysis such as behavior analysis and semantic understanding. In addition, object detection technology has been widely used in real-time monitoring of urban public security and road detection in aerial images from drones and satellites. Although many breakthroughs have been made, object detection still faces with various challenges. For instance, it is difficult to detect the object accurately under circumstances of complicated background, uneven illumination, blurred or small-size objects. The application of convolutional neural networks increases model complexity, which leads to a dramatic increase in computational cost. Object recognition and detection algorithm with high real-time, accuracy and stability has become a hot research topic. 

Around the issue of precise detection of small-scale objects on the ground, this paper conducted some researches combining with related algorithms of computer vision, machine learning. The specific research contents include: 1.research on the ground small-size objects detection on the condition of different light and various scales, from the angle of feature extraction and expression, prediction frame generation, and improvement of positioning accuracy. The main contributions of this paper are as follows: 

Firstly, the current object detection algorithms based on deep learning have performed well on objects of regular sizes. However, detection effect of small-size objects on the ground is not ideal, as the small objects with less appearance information are vulnerable to the interference of background, which leads to the difficulties of object detection. This paper compared several classical object detection frameworks based on convolutional neural networks on small objects detection. From the perspective of object feature extraction, this paper proposed a sub-network based on feature fusion to acquire enhanced semantic feature of small objects. The network has utilized information of feature maps from multiple levels of depth to obtain the fusion feature, which was used as input to the detection network. Many detection algorithms only use the information of high-level layer features, ignoring the local detail information which is critical to small objects. By comparing with these algorithms, our method effectively enhances the feature representation of small objects.

 Secondly, aiming at the objects in images of various scales and angles, this paper designs an extensional prediction network. The object is matched in multiple scale spaces of the extensional layers. The error between of the predicted value of each level and the ground truth value is used as the loss to train our model, which effectively improves the detection accuracy of the small-sized object. 

Finally, although models based on deep learning of object detection have powerful representation ability, however, the increase in complexity of the model is easy to cause the problem of over-fitting. This paper has made a series of improvements to improve the generalization ability of the model from the angle of increasing the size of training dataset and reducing the complexity of the model, which included dataset augmentation, BN layer normalization and L2-regularization. These methods further improved the detection accuracy of our model, and accelerated the speed of model convergence. Experiments were conducted with two public small-sized object datasets, which confirmed that the model proposed in this paper shows obvious advantages. 

MOST Discipline Catalogue工学
Language中文
Document Type学位论文
Identifierhttp://ir.ioe.ac.cn/handle/181551/9111
Collection光电技术研究所博硕士论文
Corresponding Author李小宁
Recommended Citation
GB/T 7714
李小宁. 基于深度学习的对地目标检测技术研究[D]. 中国科学院光电技术研究所. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
李小宁-基于深度学习的对地目标检测技术研(3338KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李小宁]'s Articles
Baidu academic
Similar articles in Baidu academic
[李小宁]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李小宁]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.