基于改进候选区域网络的红外飞机检测

为较好地解决防空武器成像系统对空中红外飞机的检测问题。首先简要地概括了卷积神经网络的兴起和应用,其次在引入基于深度学习的目标检测模型Faster R-CNN的基础上,详细地介绍了经典K-means聚类算法的工作原理、实现流程、存在的弊端以及该算法的主要改进手段,并利用K-means聚类算法对Faster R-CNN锚点框的生成方式进行了改进。最后在CAFFE框架平台下进行了多次仿真实验,测试集来源于自建的专用于空中红外飞机检测任务的数据集,实验结果表明本文采用的改进手段可以在保证较高平均准确率AP的同时提高检测速度,并且给出了最适用于本文自建数据集利用聚类算法的k值。

Infrared aircraft detection based on improved region proposal network

The purpose of this paper is to better solve the problem of airborne infrared aircraft detection by air defense weapon imaging system.Firstly,the rise and application of convolutional neural networks are briefly summarized.Secondly,based on the introduction of the deep learning target detection model Faster R-CNN,the working principle,implementation process,existing drawbacks and the main improvement methods of the classical K-means clustering algorithm are introduced in detail.The K-means clustering algorithm is used to improve the generation of anchor frame of Faster R-CNN.Finally,a number of simulation experiments were conducted under the CAFFE framework platform.The test set was derived from a self-built data set dedicated to airborne infrared aircraft detection tasks.The experimental results show that the proposed improved method can improve the detection speed while ensuring a high average precision value,and the k value that is most suitable for the self-built data set in this paper to use clustering algorithm is given.