基于FPGA的电视制导目标检测算法研究

发布时间：2018-07-20 19:41

【摘要】：如今,电视制导武器得到了越来越广泛的应用,在电视制导技术的不断向前发展的过程中,数字图像处理技术也不断地更新和进步,尤其是图像目标检测算法的研究和应用显得格外重要,与此同时深度学习领域也取得了突破性的进展,卷积神经网络的架构可以得到更鲁棒的特征。本课题提出了一种基于卷积神经网络的目标检测算法,解决了传统人工选取特征准确性不足的缺点;同时引用回归思想取代了滑动窗口或者建议区域的遍历式的工作方式,加快了目标检测的速度;同时在训练中加入难例挖掘,提高了训练之后网络的准确性。并进行三种数据集的训练和测试,对数据进行详尽分析。为了更好的满足实时性,将算法映射到FPGA上进行加速,实现了数据转换模块和卷积神经网络的设计,因为大量的卷积计算是算法复杂度的集中体现。根据FPGA的并行处理的优势和卷积神经网络并行结构的特点,对并行性进行了充分的分析和利用。由于FPGA的资源受限,对存储系统尤其是二维卷积计算的缓存安排进行精心的布置,有效的提高了片上资源的使用效率。
[Abstract]:Nowadays, TV guided weapons are more and more widely used. In the process of the continuous development of TV guidance technology, the digital image processing technology is constantly updated and improved. Especially, the research and application of image target detection algorithm is very important. At the same time, a breakthrough has been made in the field of deep learning, and the architecture of convolution neural network can get more robust features. In this paper, a target detection algorithm based on convolution neural network is proposed, which solves the shortcoming of traditional artificial selection, and uses regression to replace the traversal mode of sliding window or suggested region. The speed of target detection is accelerated and the accuracy of network after training is improved by adding hard case mining in training. Three kinds of data sets are trained and tested, and the data are analyzed in detail. In order to better meet the real-time performance, the algorithm is mapped to FPGA to accelerate the design of data conversion module and convolution neural network, because a large number of convolution computation is the concentrated embodiment of algorithm complexity. According to the advantages of FPGA parallel processing and the characteristics of convolution neural network parallel structure, the parallelism is fully analyzed and utilized. Due to the resource limitation of FPGA, the cache arrangement of storage system, especially two-dimensional convolution computing, is carefully arranged, which effectively improves the efficiency of the use of on-chip resources.
【学位授予单位】：长春理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN791;TP183;TJ765.3

【参考文献】