改进SSD的输电铁塔鸟窝检测

输电铁塔作为架空输电线的重要组成部分,其安全状况将影响整个电力系统的运行,鸟窝的搭建作为影响输电线路正常运行的重要因素之一,需要对此进行监控.而现有的监控手段不仅效率低,还需要耗费大量的人力物力.本文针对这一现象提出一种基于SSD算法的实时检测方法,并在SSD的网络结构基础上将前置网络VGGNet替换为ResNet-101,提高其特征提取能力,并将Softmax loss用Focal loss代替以改善SSD算法中样本不平衡的问题,并利用数据增广提高数据多样性,以提高模型的鲁棒性.实验结果表明本文所提出的方法检测精度对比原SSD算法在准确度和召回率上分别有3.17%和6.35%的提升.     

基于阶段聚焦损失和并行增广策略的 遥感图像场景分类

随着深度学习的不断普及,卷积神经网络已成为遥感图像场景分类的主要手段,然而当前的研究主要集中于多网络主干的信息融合以及注意力机制等领域,在提高分类精度的同时也带来更高的计算复杂度。针对上述问题,分别从改进卷积神经网络损失函数和设计新的样本训练策略两个角度出发,在不增加计算复杂度的前提下,提升卷积神经网络的分类性能。首先,在对传统交叉熵和Focal loss损失函数进行分析的基础上,提出一种阶段聚焦损失函数,该损失函数可以在训练阶段对卷积网络进行有侧重的性能挖掘。其次,设计了一种并行样本训练策略,将采用Gridmask算法增广后的样本图像和原始样本图像,分为两路输入卷积网络进行并行训练,进一步提升卷积网络的分类性能。实验结果表明,所提出的算法分别在AID和NWPU-RESISC45两个大规模数据库上取得了96.72%和93.95%的检测精度,可以显著提升遥感图像场景分类的性能。     

低旁瓣反射面天线设计

反射面天线设计在追求高增益的同时,往往也要求天线具有较低的旁瓣水平,以防止对其他系统产生干扰。通过分析不同天线形式反射面天线旁瓣的影响因素,研究了低旁瓣反射面天线设计时参数选取原则和方法,总结了低旁瓣反射面天线的设计方法。给出了一个实际工程应用的超低旁瓣双偏置反射面天线的设计实例和性能仿真计算,天线仿真的口径效率优于59%,旁瓣优于-40dB,验证了分析与设计方法的正确性,该方法可用于指导低旁瓣反射面天线的工程设计。     

p-on-n长波、甚长波碲镉汞红外焦平面器件技术研究

As注入掺杂的p-on-n结构具有暗电流小、R₀A值高、少子寿命长等优点,是长波、甚长波碲镉汞红外焦平面器件发展的重要趋势。介绍了由昆明物理研究所研究制备的77 K温度下截止波长为9.5 μm、10.1 μm和71 K下14.97 μm 的p-on-n长波、甚长波碲镉汞红外焦平面器件,对器件的响应率、NETD、暗电流及R₀A等性能参数进行测试分析。测试结果表明,器件的有效像元率在99.78%~99.9%之间,器件的NETD均小于21 mK。实现了p-on-n长波、甚长波碲镉汞红外焦平面器件的有效制备。     

Calibrating effective focal length for central catadioptric cameras using one space line

In camera calibration, focal length is the most important parameter to be estimated, while other parameters can be obtained by prior information about scene or system configuration. In this paper, we present a polynomial constraint on the effective focal length with the condition that all the other parameters are known. The polynomial degree is 4 for paracatadioptric cameras and 16 for other catadioptric cameras. However, if the skew is 0 or the ratio between the skew and effective focal length is known, the constraint becomes a linear one or a polynomial one with degree 4 on the effective focal length square for paracatadioptric cameras and other catadioptric cameras, respectively. Based on this constraint, we propose a simple method for estimation of the effective focal length of central catadioptric cameras. Unlike many approaches using lines in literature, the proposed method needs no conic fitting of line images, which is error-prone and highly affects the calibration accuracy. It is easy to implement, and only a single view of one space line is enough with no other space information needed. Experiments on simulated and real data show this method is robust and effective.     

融合GIoU和Focal loss的YOLOv3目标检测算法

YOLOv3目标检测算法检测速度快且精度较高,但存在对小目标检测能力不足、边界框定位不准确等问题。提出了一种基于YOLOv3改进的目标检测算法,该算法在YOLOv3的基础上,对网络中的残差块增加旁路连接,进一步进行特征重用,以提取更多的特征信息。同时,采用GIoUloss作为边界框的损失,使网络朝着预测框与真实框重叠度较高的方向去优化。在损失函数中加入Focal loss,减小正负样本不平衡带来的误差。在PASCAL VOC和COCO数据集上的实验结果表明,该算法能够在不影响YOLOv3算法实时性的前提下,提高目标检测的mAP。该算法在PASCAL VOC 2007测试集上达到83.7mAP（IoU=0.5）,在COCO测试集上比YOLOv3算法提升2.27mAP（IoU[0.5,0.95]）。     

一种改进的Focal Loss在语义分割上的应用

传统基于深度学习的语义分割方法使用的损失函数为交叉熵,而交叉熵并不能解决训练数据中的样本非均衡性问题。语义分割任务属于像素级分类,样本的非均衡性问题在其中体现得十分突出。文章提出了一种改进的Focal Loss作为损失函数来自动解决训练样本的非均衡性。该损失函数等同于在标准交叉熵上加上一个权重,该权重能够自动增加困难样本的交叉熵损失值,同时保持简单样本的交叉熵损失值。将Focal Loss作为DeepLabv3+的损失函数,并将DeepLabv3+的Backbone替换为ResNet-18,再使用Cityscapes数据集作为训练样本,分别使用交叉熵和Focal Loss作为损失函数来对模型进行训练。实验结果表明,改进的Focal Loss损失函数相比于交叉熵获得的语义分割精度更高,且能够有效缓解训练样本的非均衡性问题。     

红外焦平面图像稳像处理的FPGA实现

红外末制导中,弹载传感器在飞行过程中的抖动给目标检测算法以及跟踪目标的系统控制误差信号带来了极为不利的影响,必须进行实时的稳像处理,才能使导引头信息处理机正常工作。本文对弹上导引头信息处理机中的实时稳像预处理做了详尽的分析,提出了一种实时高效的稳像算法,并选用了Xilinx公司的易于实现高速硬件算法的现场可编程门阵列（FPGA）4000系列加以实现。实验结果表明,稳像效果极佳,完全符合实时制导要求。     

An adaptive observer framework for accurate feature depth estimation using an uncalibrated monocular camera

This paper presents a novel solution to the problem of depth estimation using a monocular camera undergoing known motion. Such problems arise in machine vision where the position of an object moving in three-dimensional space has to be identified by tracking motion of its projected feature on the two-dimensional image plane. The camera is assumed to be uncalibrated, and an adaptive observer yielding asymptotic estimates of focal length and feature depth is developed that precludes prior knowledge of scene geometry and is simpler than alternative designs. Experimental results using real camera imagery are obtained with the current scheme as well as the extended Kalman filter, and performance of the proposed observer is shown to be better than the extended Kalman filter-based framework.     

Measurement of temperature induced X-ray tube transmission target displacements for dimensional computed tomography

We characterised the displacement of a transmission X-ray tube target parallel to the beam axis due to thermal effects using interferometry. Displacements of up to 10 μm with a sensitivity of about 0.4 μm/W deposited power were measured relative to the X-ray tube's mechanical mounting position. They significantly influence the scale factor of dimensional computed tomography measurements by several 10⁻⁴ at high geometric magnifications. Finite element simulations revealed that the displacement is due to thermal deformation of the entire target assembly and helped to derive a linearised model that predicted the displacements within a few tenths of a micrometre. Radiographic measurements of a calibrated standard confirmed that the mechanical displacement of the target corresponded with the X-ray focal spot position. The findings help to increase the accuracy of dimensional X-ray computed tomography measurements and to ameliorate the design of transmission X-ray tubes.