基于改进SSD的车辆小目标检测方法  

Detecting method of small vehicle targets based on improved SSD

作　　者：李小宁 雷涛[1] 钟剑丹[1] 唐自力[3] 蒋平[1] LI Xiaoning;LEI Tao;ZHONG Jiandan;TANG Zili;JIANG Ping(Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610209,China;University of Chinese Academy of Sciences,Beijing 100049,China;No.63870 Troops of PLA,Xi’an 714200,China)

机构地区：[1]中国科学院光电技术研究所,四川成都610209 [2]中国科学院大学,北京100049 [3]中国人民解放军63870部队,陕西西安714200

出　　处：《应用光学》2020年第1期150-155,共6页Journal of Applied Optics

基　　金：中国科学院青年创新促进会(2016336)

摘　　要：地面车辆目标检测问题中由于目标尺寸较小,目标外观信息较少,且易受背景干扰等的原因,较难精确检测到目标。围绕地面小尺寸目标精准检测的问题,从目标特征提取的角度提出了一种特征融合的子网络。该子网络引入了重要的局部细节信息,有效地提升了小目标检测效果。针对尺度、角度等的变换问题,设计了基于融合层的扩展层预测子网络,在扩展层的多个尺度空间内匹配目标,生成目标预测框对目标定位。在车辆小目标VEDAI(vehicle detection in aerial imagery)数据集上的实验表明,算法保留传统SSD(single-shot multibox detector)检测速度优势的同时,在精度方面有了明显提升,大幅提升了算法的实用性。For the task of detecting objects on the ground such as vehicles,it’s difficult to obtain good detection results for the reason of small size and little appearance information of objects and the interference of complex background.Aiming at the problem of accurate localization of small size targets on the ground,a sub-network with feature fusion was proposed from the perspective of target feature extraction.The fusion network introduced important context information,and effectively improved the precision of small target detection.To solve the problem of transformation of sizes and angles,a extensional feature pyramid network was designed as prediction module based on fused feature.Prediction boxes were generated on different scales of extensional feature layers to match the specific objects.Experiments were conducted on the small vehicle target data set VEDAI(vehicle detection in aerial imagery).Results indicate that while the algorithm retains the advantages of detection speed of traditional SSD(single-shot multibox detector),it can significantly improve the accuracy and greatly improve the practicability of the algorithm.

关 键 词：计算机视觉 目标检测 深度学习 车辆小目标 特征融合 

分 类 号：TN206[电子电信—物理电子学] TP391.4[自动化与计算机技术—计算机应用技术]