增强区域全卷积网络下的炸点检测方法研究  

Research on Bomb-fall Detection Method Based on Advanced Region-based Fully Convolutional Networks

作　　者：刘峰[1,2] 赵广伟[1,2] 王向军 Liu Feng;Zhao Guangwei;Wang Xiangjun(National Key Laboratory of Precision Testing Techniques and Instrument,Tianjin University,Tianjin 300072;Key Laboratory of the Ministry of Education on Micro Optical Electronic Mechanical System Technology,Tianjin University,Tianjin 300072)

机构地区：[1]天津大学精密测试技术及仪器国家重点实验室,天津300072 [2]天津大学微光机电系统技术教育部重点实验室,天津300072

出　　处：《计算机辅助设计与图形学学报》2019年第3期412-420,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(51575388)

摘　　要：野外大视场环境下的炸点检测常采用图像帧间差分的方法,但由于弹体落地后炸点分布的情况复杂,对密集炸点的检测成为了难点问题.针对该问题,将炸点图像经过整理、分类,构建了炸点检测的专用数据集.在此基础上,对R-FCN模型的特征提取网络、区域推荐网络、位置敏感池化层和分类回归层进行了分析与改进,提出了增强区域全卷积网络用于单帧目标检测,并针对现在盲目多次尝试取最优训练结果的训练方法,提出了一种基于剪枝的网络模型训练方法.在野外大视场炸点图像专用数据集上进行了对照实验,最终平均检测率为83.73%,检测率明显提高.在Pascal VOC数据集上与其他常用算法进行了对比实验,结果表明了该算法的有效性.The method of image difference was often used in the detection of bomb-fall in the wild large field of view.However,due to the complexity of the situation after the bomb landing,the detection of dense bomb-fall became a difficult problem.To deal with it,a unique dataset for detection of bomb-fall was constructed after sorting and classifying the images of bomb explosion.Then,we analyzed the feature extraction network,region proposal network,position-sensitive RoI pooling layer,and classification & regression layer of R-FCN and improved them.The modified network calls advanced region-based fully convolutional networks and is used for single-frame detection.A network model training method based on pruning is used instead of the training method that blindly tries several times to obtain optimal training results.The ablation experiment was carried out on the unique dataset for detection of bomb-fall.The final mAP(mean Average Precision)reached 83.73%,which achieved a good detection performance.Compared with other commonly used algorithms on the Pascal VOC dataset,the results show the effectiveness of the algorithm.

关 键 词：炸点检测 增强区域全卷积网络 JSP&P训练方法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]