基于深度学习的快速目标检测技术研究

发布时间：2018-01-24 05:02

本文关键词： 深度学习卷积神经网络目标检测　出处：《天津理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着目前计算机视觉技术的不断发展,目标检测作为计算机领域中最重要的一部分,同时也在智能交通、图像检索、信息采集多个领域有着重要的地位。然而很多传统的目标检测算法都是基于手工设计的特征和小训练样本的,往往受到光照、遮挡、环境变化因素等众多因素的影响,进而影响了目标检测算法的性能。而近几年发展越来越火热的深度学习作为一个有力的工具能帮助人们实现更好地在目标检测领域进行研究和探索。深度学习在目标检测领域以极大的优势超越了传统视觉算法,通过在大数据下自主学习得到的特征,不仅在数量上更是在性能上远远超出了人手工设计的算法特征。虽然深度学习有着传统视觉算法不能比拟的性能优势,但是深度学习的缺点也是显而易见。深度学习中越深层的模型包含了越多的参数,带来了计算量的大幅提升,导致无法实时运行。这对于将深度学习技术应用于人们的日常生活带来了诸多困难。庞大的计算量需求导致了深度学习不能实时的运行在嵌入式或是手持设备中。为了解决这个问题本文提出了基于模型中卷积核滤波器筛选策略的模型精简以及端到端的目标检测方法,具体工作内容如下:(1)针对目前深度学习模型中参数繁多、运算耗时久、占用体积大的问题,本文提出了一种基于平均响应值Average Response Energy(ARE)的卷积核滤波器筛选策略,主要是将深度模型中贡献度低的卷积核滤波器从原模型中移除,可以大量的减少原模型的权重参数,进而减少模型的占用体积与计算量。在本文的三个公共数据集上,实验结果也是表明了基于平均响应值的卷积核滤波器筛选策略对于降低模型体积和计算量是有效的,在不牺牲精度的情况下,能够大幅提升深度学习模型的实时性。(2)对于当前结合分割定位算法和深度学习分离式框架,本文通过在深度模型的最后一层卷积层后添加多尺度卷积特征提取层,训练了一个端到端的基于深度学习的目标检测框架。在PASCAL VOC 2007数据集上,实验结果也是表明了该端到端的深度学习目标检测算法可以有效降低模型的运算时间。
[Abstract]:With the development of computer vision technology, target detection, as the most important part of computer field, is also in the fields of intelligent transportation and image retrieval. Information collection plays an important role in many fields. However, many traditional target detection algorithms are based on hand-designed features and small training samples, often subject to light, occlusion. Environmental change factors and many other factors. In recent years, as a powerful tool, it can help people to better research and explore in the field of target detection. The field of target detection surpasses the traditional visual algorithm with great advantage. Through the autonomous learning under big data, not only in quantity but also in the performance of manual design far beyond the characteristics of the algorithm, although depth learning has the traditional visual algorithm can not compare with the performance advantages. However, the disadvantages of deep learning are obvious. The deeper the model is, the more parameters are included, which leads to a significant increase in computational complexity. This makes it difficult to apply deep learning technology to people's daily life. The huge demand for computing makes it impossible for deep learning to run in real-time on embedded or handheld devices. In order to solve this problem, this paper proposes a model reduction and end-to-end target detection method based on the filtering strategy of convolutional kernels in the model. The specific work is as follows: (1) aiming at the problem that there are many parameters in the current depth learning model, the computation time is long, and the volume is large. In this paper, a convolutional kernel filter screening strategy based on average response value Average Response energy is proposed. It is mainly to remove convolutional kernel filter from the original model, which can greatly reduce the weight parameters of the original model. And then reduce the volume and computation of the model. In this paper on the three common data sets. The experimental results also show that the convolution kernel filter screening strategy based on the average response value is effective to reduce the volume and computational complexity of the model without sacrificing the accuracy. Can greatly improve the depth of learning model of real-time. 2) for the current combination of segmentation and location algorithm and depth learning separation framework. In this paper, we add multi-scale convolution feature extraction layer after the last layer of the depth model. An end-to-end target detection framework based on in-depth learning is trained on the PASCAL VOC 2007 dataset. The experimental results also show that the end-to-end depth learning target detection algorithm can effectively reduce the operation time of the model.
【学位授予单位】：天津理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】