论文总字数：24548字

摘 要

随着科技的快速发展，深度学习技术凭借其自动提取特征和优秀的学习能力等特性成为目标检测的主流技术，但在有限的硬件条件下，检测速度和检测精度还远落后于人类的识别水平，限制了计算机视觉技术在机器人领域的研究进展。因此，在有限硬件的情况下，如何在根本上保证提高目标检测速度的同时又能提高目标检测的精度，是目标检测技术需要大量时间去研究和解决的一个技术问题。

通过对目标检测算法YOLO进行研究，针对有限硬件情况下，对模型结构和训练过程进行学习和应用，确保检测速度和模型精度，达成目标实时检测，主要工作如下：

（1）学习神经网络，详细的阐述了CNN网络的基本构成，并研究基于CNN网络的各个实时目标检测算法，对这些算法进行对比，分析各自的优缺点，总结归纳了他们的创新点。

（2）分析YOLO算法的设计原理，以Tiny-YOLOv3模型为基础，搭建相关环境配置，改进算法，创建自己的数据集，在保证速度的情况下，能准确完成杯子和笔的检测。

本文基于Tiny-YOLOv3算法进行深入学习，提出了用Tiny-YOLOv3改进算法进行实时目标检测，这种算法在速度和精度上性能良好，通过制作杯子和笔的数据集，并改进相关配置文件，在数据集上训练网络，经过训练得到一个稳定的权重结果，最后将其匹配网络，确保检测能力，验证了其准确快速的完成目标检测的能力。

关键词：深度学习 YOLO算法 目标检测 卷积神经网络

Abstract

With the rapid development of science and technology, it is no longer impossible for the machine to complete the target detection. Deep learning technology has become the mainstream technology of real-time target detection by virtue of its automatic feature extraction and excellent learning ability. However, the detection speed and precision are far behind the level of human recognition under the limited hardware condition, which limits the development of computer vision technology in the field of robot. Therefore, in the case of limited hardware, how to improve the speed of target detection and the accuracy of target detection is a technical problem that needs a lot of time to study and solve.

In this paper, we study the object detection algorithm YOLO, and study and apply the model structure and training process to ensure the detection speed and model accuracy, so as to achieve real-time target detection. The main work is as follows:

(1) Learning neural network, elaborated the basic structure of CNN in detail, and studied each algorithm based on CNN, compared these algorithms, analyzed their advantages and disadvantages, summarized their innovation points.

(2) The design principle of YOLO algorithm is analyzed. Based on the Tiny-YOLOv3 model, the relevant environment configuration is built, the algorithm is improved, and the data set is created. Under the condition of ensuring the speed, the detection of cups and pens can be completed accurately.

In this paper, we use the improved Tiny-YOLOv3 algorithm for real-time object detection based on the study of Tiny-yolov3 Algorithm, which has good performance in speed and precision, the network is trained on the data set, and a stable weight result is obtained after the training. Finally, it is matched with the network to ensure the detection ability and verify its ability to complete the target detection accurately and quickly.

Keywords：Deep learning; YOLO algorithm; Object detection; Convolutional neural networks

目 录

摘 要 I

Abstract II

目 录 III

第一章 绪论 1

1.1 课题背景及研究意义 1

1.2 国内外研究现状 1

1.3 本文主要内容及结构安排 3

第二章 实时目标检测算法及相关理论 5

2.1 实时目标检测算法 5

2.1.1 R-CNN 5

2.1.2 SPP-net 5

2.1.3 Faster R-CNN 6

2.2 基于深度学习的实时目标检测算法——YOLO 6

2.2.1 YOLOv1 7

2.2.2 YOLOv2 7

2.2.3 YOLOv3 8

2.3 本章小结 10

第三章 Tiny-YOLOv3网络框架理论介绍及改进 11

3.1 CNN基本结构 11

3.2 YOLO算法原理实现 15

3.2.1 边界框预测 15

3.2.2 非极大值抑制 16

3.2.3 锚框（Anchor Box） 17

3.2.4 YOLO的损失函数 18

3.3 Tiny-YOLOv3的检测框架 18

3.4 Tiny-YOLOv3算法的改进 20

3.5 本章小结 22

第四章 Tiny-YOLOv3的目标检测实验与结果分析 23

4.1 实验设备和环境 23

4.2 目标检测数据集 23

4.3 Tiny-YOLOv3网络训练 25

4.4 算法改进前后实验结果分析 27

4.4.1目标检测衡量指标 27

4.4.2改进前后检测精度对比 27

4.4.3训练时间和检测速度 29

4.4.4不足及优化思路 30

4.5 本章小结 30

第五章 总结与展望 31

5.1 总结 31

5.2 展望 31

参考文献 33

致谢 35

请支付后下载全文，论文总字数：24548字