基于卷积网络的物体检测应用研究
发布时间：2018-12-21 10:59
【摘要】：提出一种基于卷积神经网络改进的行人检测方法。改进主要涉及两个方面,包括如何决定CNN样本迭代学习次数和如何进行重合窗口的合并。第一,关于CNN样本迭代次序问题,在顺序迭代训练多个CNN分类模型的基础上,提出一种基于校验集正确率及其在迭代系列分类器中展现出的稳定性并进行更优模型选择的策略,以使最终选择的分类器推广能力更优。第二,提出了一种不同于非极大值抑制的多个精确定位回归框合并机制。精确定位回归框的获取以CNN检测过程输出的粗定位框作为输入,然后对每个粗定位框应用CNN精确定位过程并获得对应的精确定位回归框,最后对多个精确定位回归框进行合并,合并过程考虑了每个精确定位回归框的正确概率。更精确来说,最终的合并窗口基于多个相关的精确定位回归框的概率加权求和方式获得。针对提出的两个改进,在国际上广泛使用的行人检测公共测试数据集ETH上进行了一系列实验。实验结果表明,提出的两个改进方法均能有效地提高系统的检测性能,在相同的测试条件下,融合两个改进的方法相比Fast R-CNN算法检测性能提升了5.06%,达到了 40.01%的检测结果。提出一种基于卷积神经网络的车型分类方法,首先建立了十万数量级的不同车型不同场景的车型分类数据库,通过设计网络结构并训练卷积神经网络模型,在其自定义测试集上进行分类器性能验证。实验过程中对样本进行了对齐和扩边界操作,对比了不同训练迭代次数产生的CNN分类器,最终得到的CNN分类器在车型分类数据库测试集上的平均正确率达到了 95.49%。实验结果表明,本文提出的卷积神经网络模型在车型分类任务上取得了较好的分类精度。
[Abstract]:An improved pedestrian detection method based on convolution neural network is proposed. The improvement mainly involves two aspects, including how to determine the number of CNN sample iterative learning and how to merge the overlap window. First, with regard to the iterative order of CNN samples, on the basis of sequential iterative training of multiple CNN classification models, a strategy based on the correct rate of check set and its stability in iterative series classifier is proposed to select a better model. In order to make the final choice of classifier promotion ability better. Secondly, a combination mechanism of multiple exact location regression frames is proposed, which is different from non-maximum suppression. The accurate location regression frame is obtained by using the coarse positioning box output from the CNN detection process as input, and then the CNN precise positioning process is applied to each coarse positioning box and the corresponding accurate positioning regression box is obtained. Finally, the multiple accurate positioning regression boxes are merged. The merging process takes into account the correct probability of each exact location regression box. More precisely, the final merge window is based on the probability weighted summation of multiple correlated precise location regression frames. Aiming at the two improvements proposed, a series of experiments have been carried out on ETH, a common test data set for pedestrian detection, which is widely used in the world. The experimental results show that the proposed two improved methods can effectively improve the detection performance of the system. Under the same test conditions, the two improved methods can improve the detection performance of the Fast R-CNN algorithm by 5.06 steps. The test results are 40.01%. This paper presents a vehicle classification method based on convolution neural network. Firstly, a model classification database of different models with different scenes of 100,000 orders of magnitude is established, and the network structure is designed and the convolutional neural network model is trained. The classifier performance is verified on its custom test set. During the experiment, the samples are aligned and expanded, and the CNN classifier produced by different training iterations is compared. The average correct rate of the CNN classifier on the test set of vehicle classification database is 95.49. The experimental results show that the proposed convolution neural network model has achieved better classification accuracy in vehicle classification task.
【学位授予单位】：南京信息工程大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
本文编号：2388817
[Abstract]:An improved pedestrian detection method based on convolution neural network is proposed. The improvement mainly involves two aspects, including how to determine the number of CNN sample iterative learning and how to merge the overlap window. First, with regard to the iterative order of CNN samples, on the basis of sequential iterative training of multiple CNN classification models, a strategy based on the correct rate of check set and its stability in iterative series classifier is proposed to select a better model. In order to make the final choice of classifier promotion ability better. Secondly, a combination mechanism of multiple exact location regression frames is proposed, which is different from non-maximum suppression. The accurate location regression frame is obtained by using the coarse positioning box output from the CNN detection process as input, and then the CNN precise positioning process is applied to each coarse positioning box and the corresponding accurate positioning regression box is obtained. Finally, the multiple accurate positioning regression boxes are merged. The merging process takes into account the correct probability of each exact location regression box. More precisely, the final merge window is based on the probability weighted summation of multiple correlated precise location regression frames. Aiming at the two improvements proposed, a series of experiments have been carried out on ETH, a common test data set for pedestrian detection, which is widely used in the world. The experimental results show that the proposed two improved methods can effectively improve the detection performance of the system. Under the same test conditions, the two improved methods can improve the detection performance of the Fast R-CNN algorithm by 5.06 steps. The test results are 40.01%. This paper presents a vehicle classification method based on convolution neural network. Firstly, a model classification database of different models with different scenes of 100,000 orders of magnitude is established, and the network structure is designed and the convolutional neural network model is trained. The classifier performance is verified on its custom test set. During the experiment, the samples are aligned and expanded, and the CNN classifier produced by different training iterations is compared. The average correct rate of the CNN classifier on the test set of vehicle classification database is 95.49. The experimental results show that the proposed convolution neural network model has achieved better classification accuracy in vehicle classification task.
【学位授予单位】：南京信息工程大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
【参考文献】
相关期刊论文 前2条
1 冯兴辉;张旭;陈礼贵;宋杰;;直线特征测量数据的平滑去噪[J];机械制造;2015年04期
2 刘远社;傅里叶变换与测不准原理[J];西南民族大学学报(自然科学版);2003年05期
，本文编号：2388817
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2388817.html
