基于深度学习的目标检测研究

发布时间：2018-05-25 05:19

本文选题：目标检测 + 深度学习　；参考：《北京交通大学》2017年硕士论文

【摘要】：目标检测作为图像处理和计算机视觉领域中的经典课题,在交通监控、图像检索、人机交互等方面有着广泛的应用。它旨在一个静态图像(或动态视频)中检测出人们感兴趣的目标对象。传统的目标检测算法中特征提取和分类决策分开进行,对特征选取的要求就更加严格,在面对复杂场景的时候很难得到理想效果。自Hinton教授提出深度学习理论,越来越多的研究学者尝试采用深度学习理念来解决目标检测问题,并且提出了不同的模型。不同的模型应用也不尽相同,通常采用卷积神经网络来处理目标检测问题。相比于传统目标检测算法,卷积神经网络中特征提取和模式分类并行进行,而且随着层数的增多可以更好的处理复杂场景,但是它对目标边缘的约束性太差。在这样的基础上,本文对传统算法和卷积神经网络做了深入的研究,实现了将传统算法和卷积神经网络相结合的目标检测算法。本文的主要工作和创新有:(1)针对传统目标检测算法一般是使用矩形框的方式得到目标的大致区域,而我们的需求是尽可能的获得目标的边缘轮廓问题,本文实现了一种改进的基于主动轮廓模型的目标检测算法,使轮廓尽可能的接近目标。(2)针对传统算法需要人工设计图像特征,不同场景模型不稳定问题而卷积神经网络分割不精确以及缺少相邻像素之间的约束的问题,本文将传统算法和卷积神经网络相结合,使用卷积神经网络进行图像“高层次”特征提取,超像素提取出图像“低层次”特征,可以适应不同的复杂场景,并且获得准确的目标边缘。在颐和园景点数据库中进行了充分的实验。通过结果可以看出使用我们的算法进行目标检测提取,可以很精确的提取出目标,而且目标的边缘约束性也非常强。(3)主要创新点:在使用超像素提取特征删除重复特征,减少特征冗余,降低特征的维度;同时将VGGNet(Visual Geometry Group Network)网络换成收敛速度更快的GoogleNet网络,提高了算法的速度。
[Abstract]:As a classical subject in the field of image processing and computer vision, target detection has been widely used in traffic monitoring, image retrieval, human-computer interaction and so on. It aims to detect the object of interest in a static image (or dynamic video). In the traditional target detection algorithm, the feature extraction and classification decision are carried out separately, the requirement of feature selection is more strict, and it is difficult to obtain ideal results in the face of complex scene. Since Professor Hinton put forward the theory of deep learning, more and more researchers have tried to solve the problem of target detection by using the concept of deep learning, and put forward different models. The application of different models is different. Convolution neural network is usually used to deal with the target detection problem. Compared with the traditional target detection algorithm, the convolution neural network features extraction and pattern classification parallel, and with the increase of the number of layers can better deal with the complex scene, but it is too bad to target edge constraint. On this basis, the traditional algorithm and the convolutional neural network are studied in this paper, and the target detection algorithm combining the traditional algorithm and the convolution neural network is realized. The main work and innovation of this paper are: 1) for traditional target detection algorithms, we usually use rectangular frame to get the approximate area of the target, and our requirement is to obtain the edge contour of the target as much as possible. In this paper, an improved target detection algorithm based on active contour model is implemented, which makes the contour as close as possible to the target. In this paper, the traditional algorithm and convolution neural network are combined to solve the problems of unstable scene models and inaccurate segmentation of convolutional neural networks and the lack of constraints between adjacent pixels. Using convolution neural network to extract "high level" feature of image, super-pixel can extract "low level" feature of image, which can adapt to different complex scene and obtain accurate target edge. A full experiment has been carried out in the Summer Palace scenic spot database. The results show that using our algorithm for target detection and extraction, we can extract the target accurately, and the edge constraint of the target is also very strong. At the same time, the VGGNet(Visual Geometry Group Network) network is replaced by a faster convergent GoogleNet network, which improves the speed of the algorithm.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】