【专题】基于深度学习及FPGA的装备目标检测研究

本文应用深度学习技术实现海天背景下基于可见光、红外方式成像的舰船及角反、烟幕干扰的目标检测,这也是反舰导弹作战使用的关键技术之一。采集的可见光与红外成像目标检测数据集涵盖实施典型干扰下的态势场景,贴近实战;结合四种不同的目标检测机制,选取YOLOV3、Faster R-CNN、SSD及CenterNet四种典型模型分别进行训练与验证,通过对比分析进一步提高弱小目标、复杂干扰态势的的检测,可以实现端到端的高精度装备目标检测模型。在确保精度的前提下基于现场可编程门阵列（FPGA）进行软硬件协同设计,通过对比分析选定基于Vitis AI的实施方案,经过模型的量化、编译与优化,可在保证检测效率的前提下快速实现模型的小型化部署,便于进行装备移植。研究结果表明,该研究内容可有效提高现役反舰导弹目标检测的准确率。     

Hyperspectral imaging and remote trace detection of cis-1,3,4,6-tetranitrooctahydroimidazo-[4,5 d] imidazole (BCHMX) compared with traditional explosives using laser induced fluorescence

cis-1,3,4,6-Tetranitrooctahydroimidazo-[4,5 d] imidazole (BCHMX) is an advanced energetic compound that expected to spread worldwide in the near future. Since, no approved remote detection methods were reported in current literature for this material, we performed hyper-spectral imaging and laser induced fluorescence (LIF) to a BCHMX sample under low laser fluence for determining the optimum laser wavelength used in any future BCHMX-LIF based remote detection systems. For this purpose, an experimental setup consisted of a sun spectrum lamp and hyper-spectral camera was built to illuminate and image white powder samples of BCHMX in comparison with the traditional explosives, HMX (1,3,5,7-tetranitro-1,3,5,7-tetraazacyclooctane), RDX (1,3,5-trinitro-1,3,5-triazacyclohexane), PETN (2,2-Bis[(nitroxy)methyl]propane-1,3-diyldinitrate). The imaging reveals strong BCHMX sample absorption contrast among other samples at wavelength ranging from 400 to 410 nm. When light source was replaced by a 405 nm laser diode illuminator, a strong BCHMX sample LIF at the spectral range from 425 to 700 nm was observed under low laser fluence condition of 0.1 mJ/cm². Finally, we demonstrated successfully the ability of the 405 nm LIF and the hyperspectral imaging technique to detect finger print traces of BCHMX on white cellulose fabric from a distance of 15 m and a detection limit of 1 μg/cm².     

结合YdUaVa颜色模型和改进MobileNetV3的视频烟雾检测方法

为降低视频烟雾检测中的虚警率和提升检测效率,提出Y^dU^aV^a颜色模型,该模型可以表征烟雾的空间域分布特性和时间域变化特性。利用该颜色模型快速筛选出疑似烟雾图像块,降低虚警率和提升运算效率。提出改进的MobileNetV3网络结构,用于提取图像深度特征并对疑似烟雾图像块进行分类识别,检测视频中是否存在烟雾。视频烟雾检测仿真结果表明:该方法准确率和检测帧率高,虚警率低。     

改进非凸估计与非对称时空正则化的红外小目标检测方法

针对复杂背景下的红外小目标检测,在非对称时空正则化约束的非凸张量低秩估计算法基础上,提出了一种新的核范数估计方法代替原算法中的估计方法。提出基于结构张量与多结构元顶帽(Top-Hat)滤波的自适应权重张量对目标张量进行约束,增强目标张量稀疏性的同时抑制其中残存的强边缘结构。实验结果表明,所提改进算法能够更好地消除图像中强边缘结构对检测结果的影响,在保证检测率的情况下,较原算法具有更低的虚警率。     

一种复合模式的人脸识别系统设计

通过综合考察红外 可见光图像人脸识别的优缺点,利用红外温谱图进行人脸检测与识别的优势,在研究可见光和红外人脸温谱图识别技术以及两者融合技术的基础上,提出和设计了一种新的红外 可见光复合模式的人脸识别系统。初步实验结果表明,这种识别方法不论从理论上还是从实验上都是可行的,系统具有良好的识别能力。