Bi-projection for 360°image object detection bridged by RoI Searcher

Object detection on 360°images is a vital component of 3D environment perception. The existing methods either treat panoramic images (usually represented in equirectangular projection—ERP) as normal FoV images and endure the distortions or project them into the less-distortion format and narrow the FoV, leading to unsatisfactory performance in practical applications. To solve this problem, we propose a dual-projection 360°object detection network named Bip R-CNN, consisting of three modules: a bi-projection feature extractor, a cross-projection region-of-interest (RoI) searcher, and a classification and regression predictor. Specifically, we extract the equirectangular and corresponding dual-cubemap features simultaneously from the input images. Besides, Projection-Inter Feature Fusion and Projection-Intra Feature Fusion are designed to allow the mutual interaction between the bi-projective features and promote the integration of features at different scales, respectively. In the proposed cross-projection RoI Searcher, we search for the bounding box (BBox) locations on cubemap from the corresponding ERP spherical proposals, bridging the RoIs of two different projection formats at feature level. Finally, the cube proposals are used to detect objects in the last predictor module. Considering the scarceness of the existing panoramic dataset (only indoor scenes), we propose an efficient approach to convert conventional datasets into annotated panoramic datasets without manual intervention, increasing the diversity of panoramic datasets. Extensive experiments are conducted on the synthetic and real-world datasets with spherical criteria, demonstrating our superiority to other state-of-the-art solutions.     

全局关系注意力引导场景约束的高分辨率遥感影像目标检测

高分辨率遥感影像中地物目标往往与所处场景类别息息相关,如能充分利用场景对地物目标的约束信息,有望进一步提升目标检测性能。考虑到场景信息和地物目标之间的关联关系,提出全局关系注意力(RGA)引导场景约束的高分辨率遥感影像目标检测方法。首先在多尺度特征融合检测器的基础网络之后,加入全局关系注意力学习全局场景特征;然后以学到的全局场景特征作为约束,结合方向响应卷积模块和多尺度特征模块进行目标预测;最后利用两个损失函数联合优化网络实现目标检测。在NWPU VHR-10数据集上进行了4组实验,在场景信息约束的条件下取得了更好的目标检测性能。     

基于Grad-CAM与KL损失的SSD目标检测算法

鉴于Single Shot Multibox Detector（SSD）算法对中小目标检测时会出现漏检甚至错检的情况,提出一种改进的SSD目标检测算法,以提高中小目标检测的准确性.运用Gradient-weighted Class Activation Mapping（Grad-CAM）技术对检测过程中的细节作可视化处理,并以类激活图的形式呈现各检测层细节,分析各检测层的类激活图发现SSD算法中待检测目标的错检以及中小目标的漏检现象与回归损失函数相关.据此,采用Kullback-Leibler（KL）边框回归损失策略,利用Non Maximum Suppression（NMS）算法输出最终预测框.实验结果表明,改进算法相较于已有检测算法具有更高的准确率以及稳定性.     

基于特征复用的膨胀-残差网络的SAR图像超分辨重建

对于合成孔径雷达(SAR)图像,传统的超分辨重建方法对视觉特征的人为构造十分依赖,基于普通卷积神经网络(CNN)的超分辨重建方法对微小目标的重建能力较弱,对边缘轮廓的保真度较差。针对以上问题,该文提出一种基于特征复用的膨胀-残差卷积超分辨网络模型,同时引入感知损失,实现了精确的SAR图像4倍语义级超分辨。该方法为增加网络感受野,采用膨胀-残差卷积(DR-CNN)结构用于限制模型中特征图分辨率的严重损失,提高网络对微小细节的敏感度;为实现不同层级的特征最大化利用,将不同层级的特征图进行级联,形成一种特征复用结构(FRDR-CNN),以此大幅度提升特征提取模块的效率,进一步提升超分辨精度;针对SAR图像特殊的相干斑噪声干扰,引入感知损失,使得该方法在恢复图像边缘和精细的纹理信息方面具有优越表现。文中实验表明,与传统算法以及目前较为流行的几种全卷积神经网络超分辨重建算法相比,该文采用的FRDR-CNN模型在视觉上对小物体的超分辨重建能力更强,对边界等轮廓信息的重建更准确,客观指标中的峰值信噪比(PSNR)和结构相似性指数(SSIM)分别为33.5023 dB和0.5127,边缘保持系数(EPD-ROA)在水平和垂直方向上分别为0.4243和0.4373。      

Ferret: An RFID-enabled pervasive multimedia application

This paper presents the design and implementation of Ferret, a system for locating nomadic augmented with RFID tags and visually displaying them to a user in real-time. We present a novel algorithm to infer location of tagged objects using the location of a camera and reader that observes them. We also present techniques to refine location estimates using multiple observations and a method to display and update object locations on a video camera screen. An experimental evaluation of the Ferret prototype shows that (i) Ferret can refine object locations to only 1% of the reader’s coverage region in less than 2 min with small error rate (2.22%); (ii) Ferret can detect nomadic objects with 100% accuracy when the moving distances exceed 20 cm; and (iii) Ferret is robust against different movement patterns of user’s mobility.     

基于深度学习的红外遥感信息自动提取

为了提高红外遥感图像地物 信息自动提取的精确性,同时避免人工提取遥感 信息的低效性,提出了一种基于UNet深度学习模型 的遥感信息提取算法。该算法用于从红外遥感图像中分割 出5类地物信息(包括道路、建筑、树木、农田和水 体)。首先,对分辨率高但数量较少的训练数 据进行小像幅的随机裁剪,并对其进行相应的数据增 强处理。然后搭建UNet深度学习模型,并用它 自动提取遥感图像的特征信息。采用交叉熵损失函数 以及Adam反向传播优化算法对该模型进行训练,并对测 试样本中的5幅遥感图像进行精确的地物信息提取。最后,运 用Jaccard指数对测试结果进行精度评定。实验结果表明,该 方法对高分辨率红外遥感图像信息和可见光 遥感图像信息进行了充分融合,对于不同种类地物 的定位和分类都取得了较高精度。     

基于特征融合与通道注意力的目标检测研究

目标检测作为计算机视觉领域的重要研究分支，受到了广泛关注。目前，特征融合已成为提高目标检测准确率的重要方法，基于特征金字塔网络（FPN）的特征融合方法结合了多维度感受野来弥补信息丢失的思想，改进了FPN，并取得了良好的效果。在众多以FPN为基础结构的特征金字塔模型中，BiFPN既包含了级联信息传递，也包含了跨层特征融合，DyFPN包含了多感受野Inception模块，也包含了动态门机制。受BiFPN和DyFPN的启发，文中提出了一种新的基于FPN且包含注意力机制的特征金字塔网络结构——CAI-BiFPN。CAI-BiFPN沿用了Inception-FPN的思想，在BiFPN的基础上加入了Inception模块，并引入了通道注意力和SE模块。该结构应用了BiFPN的分支注意力与SE模块的通道注意力，将Inception模块放置于BiFPN第4层和第6层，SE模块则放置在BiFPN的第5层。通过一系列简单的改进，相较于BiFPN，CAI-BiFPN的AP_s和AP_l提升了0.7个百分点，AP从31.0%提升到31.3%，提升了0.3个百分点。     

基于一种卷积神经网络的目标快速跟踪技术

特定动态目标的快速检测及跟踪,是计算机视觉领域重要的课题。改变特征图在YOLOv3卷积神经网络中的选取位置,通过收集相关网络数据(类似模式分析、统计建模和计算学习视觉对象类别数据集合,即PASCAL Visual Object Classes数据集)构建自定义数据集合进行训练,使用面积的交并比完成辅助类别的联合,构建了能够实时检测特定目标在相关可视对象类检测数据集合上mAP@75达到47.41的检测器。联合卡尔曼滤波和匈牙利算法,通过将面积信息加入到匈牙利算法的代价矩阵中,改善了使用原方法产生大量ID切换(ID switch)的问题。该方法满足快速识别与跟踪的要求,在使用一张NVIDIA GeForce GTX 1060 6GB GPU条件下,平均速度能达到0.109 7 s/帧。     

CCA-FPN: Channel and content adaptive object detection

Feature pyramid network (FPN) is a typical detector commonly for solving the issue of object detection at different scales. However, the lateral connections in FPN lead to the loss of feature information due to the reduction of feature channels. Moreover, the top-down feature fusion will weaken the feature representation in the process of feature delivery because of features with different semantic information. In this paper, we propose a feature pyramid network with channel and content adaptive feature enhancement module (CCA-FPN), which uses a channel adaptive guided mechanism module (CAGM) and multi-scale content adaptive feature enhancement module (MCAFEM) to alleviate these problems. We conduct comprehensive experiments on the MS COCO dataset. By replacing FPN with CCA-FPN in ATSS, our models achieve 1.3 percentage points higher Average Precision (AP) when using ResNet50 as backbone. Furthermore, our CCA-FPN achieves 0.3 percentage points higher than the AugFPN which is the state-of-the-art FPN-based detector.     

面向显著性目标检测的SSD改进模型

传统显著性目标检测方法常假设只有单个显著性目标,其效果依赖显著性阈值的选取,并不符合实际应用需求。近来利用目标检测方法得到显著性目标检测框成为一种新的解决思路。SSD模型可同时精确检测多个不同尺度的目标对象,但小尺寸目标检测精度不佳。为此,该文引入去卷积模块与注意力残差模块,构建了面向多显著性目标检测的DAR-SSD模型。实验结果表明,DAR-SSD检测精度显著高于SOD模型;相比原始SSD模型,在小尺度和多显著性目标情形下性能提升明显;相比MDF和DCL等深度学习框架下的方法,也体现了复杂背景情形下的良好检测性能。     

一种基于词袋模型的图像分类方法

采用词袋模型(BoW)对图像进行分类,并针对传统词袋模型存在的不足进行了改进,提出了一种特征软量化的方式。软赋值量化通过将局部显著特征量化(SIFT)为与其距离最近的若干个视觉单词,并对其进行加权,由此保存特征空间中的距离信息,从而解决硬赋值量化造成的特征空间信息损失问题。通过在Caltech 101数据库进行实验,验证了本文方法的有效性,实验结果表明,该方法能够大幅度提高图像分类的性能。     

TDFSSD: Top-Down Feature Fusion Single Shot MultiBox Detector

Object detection across different scales is challenging as the variances of object scales. Thus, a novel detection network, Top-Down Feature Fusion Single Shot MultiBox Detector (TDFSSD), is proposed. The proposed network is based on Single Shot MultiBox Detector (SSD) using VGG-16 as backbone with a novel, simple yet efficient feature fusion module, namely, the Top-Down Feature Fusion Module. The proposed module fuses features from higher-level features, containing semantic information, to lower-level features, containing boundary information, iteratively. Extensive experiments have been conducted on PASCAL VOC2007, PASCAL VOC2012, and MS COCO datasets to demonstrate the efficiency of the proposed method. The proposed TDFSSD network is trained end to end and outperforms the state-of-the-art methods across the three datasets. The TDFSSD network achieves 81.7% and 80.1% mAPs on VOC2007 and 2012 respectively, which outperforms the reported best results of both one-stage and two-stage frameworks. In the meantime, it achieves 33.4% mAP on MS COCO test-dev, especially 17.2% average precision (AP) on small objects. Thus all the results show the efficiency of the proposed method on object detection. Code and model are available at: https://github.com/dongfengxijian/TDFSSD.     

利用改进特征金字塔模型的SAR图像多目标船舶检测

深度学习模型中的特征金字塔网络（Feature Pyramid Network,FPN）常被用作合成孔径雷达（Synthetic Aperture Radar,SAR）图像中多目标船舶的检测。针对复杂场景下多目标船舶检测问题,提出了一种基于改进锚点框的FPN模型。首先将特征金字塔模型嵌入传统的RPN（Region Proposal Network）并映射成新的特征空间用于目标检测,然后利用基于形状相似度距离（Shape Similar Distance,SSD）度量的Kmeans聚类算法优化FPN的初始锚点框,并使用SAR船舶数据集测试。实验结果表明,所提算法目标检测精确率达到98.62%,在复杂场景下与YOLO、Faster RCNN、FPN based on VGG/ResNet等模型进行对比,模型准确率提高,整体性能更好。     

一种轻量化的水下目标检测算法SG-Det

基于深度学习的目标检测算法在水下进行检测主要存在两个困难：水下设备的存储和计算能力有限;水下图像模糊且小生物聚集。这种局限性要求水下目标检测算法要做到轻量且高效,因此现有的目标检测算法不能完全满足水下目标检测的需求。为此本文在SSD(single shot MultiBox detector)的网络框架上进行改进,设计了一种轻量化的一阶段检测模型SG-Det。一方面,借鉴GhostNet的思想,对ShuffleNetV2网络进行了重构,提出一种新的轻量化特征提取网络SGnet。此网络进一步减少了模型参数量,使模型大小更适合部署在水下设备。另一方面,网络主要是利用6个不同尺度的特征图检测不同大小的生物,为此设计了一种引入双分支注意力机制的跨尺度特征融合模块(cross-scale feature fusion module, AFF)。模块首先引入注意力机制对输入的特征在全局通道和局部通道两方面进行加权,突出有用信息,从而减少背景等无关信息的干扰。然后选取非线性化程度更高的第4层分别增强前3层的语义信息,以较小的代价使前3层在识别小物体方面有更好的表现。模型在中国水下机器人大赛提供的水下...     

网络流量特征选择方法中的分治投票策略研究

特征选择作为机器学习过程中的预处理步骤,是影响分类性能的关键因素.网络流量具有数据量大,特征维度高的特点,如何快速提取特征子集,并提高分类效率对于基于机器学习的流量分类方法具有重要意义.本文提出基于分治与投票策略的特征提取方法,将数据集分裂为多个子集,分别执行特征提取算法,利用投票方法获得最后的特征子集.实验表明可有效提高特征提取的时间效率,同时使分类器取得良好的分类准确率.     

显著性背景感知的多尺度红外行人检测方法

超大视场(U-FOV)红外成像系统探测范围大、不受光照限制,但存在尺度多样、小目标丰富的特点。为此该文提出一种具备背景感知能力的多尺度红外行人检测方法,在提高小目标检测性能的同时,减少冗余计算。首先,构建了4尺度的特征金字塔网络分别独立预测目标,补充高分辨率细节特征。其次,在特征金字塔结构的横向连接中融入注意力模块,产生显著性特征,抑制不相关区域的特征响应、突出图像局部目标特征。最后,在显著性系数的基础上构建了锚框掩膜生成子网络,约束锚框位置,排除平坦背景,提高处理效率。实验结果表明,显著性生成子网络仅增加5.94%的处理时间,具备轻量特性;超大视场(U-FOV)红外行人数据集上的识别准确率达到了93.20%,比YOLOv3高了26.49%;锚框约束策略能节约处理时间18.05%。重构模型具有轻量性和高准确性,适合于检测超大视场中的多尺度红外目标。     

多尺度局部区域响应累积的非滑窗快速目标检测算法

针对滑动窗口全局搜索检测目标搜索时间长的问题,提出一种多尺度局部区域响应累积的非滑窗快速目标检测算法。首先,提取检测目标多尺度可重叠局部区域作为训练样本,通过学习得到多尺度且具有判别能力的部件集,部件集中每个局部区域与检测目标有明确位置对应关系;然后,根据各投影检测器响应判断目标是否在某一区域出现,并利用多尺度目标局部区域的检测结果和位置约束进行投票,完成目标粗定位;其次,利用HOG特征提取和SVM相结合完成检测目标验证实现准确检测。该方法将多尺度部件模型、统计累积投票思想及分类器判决相结合,实现快速目标检测,大大减少滑动窗口逐像素搜索背景时所消耗时间,提高检测效率。      

Highly Soft Viscoelastic Robot Skin With a Contact Object-Location-Sensing Capability

This paper concerns the development of robot skin capable of accurately sensing the location of objects in area contact with the skin surface. There has been no report on tactile sensing which attained not only skin deformation detection but also contact object location sensing with high accuracy. In the category of optomechatronics technology, we apply optical fibers to transmit surface deformation information of soft skin for sensing the location of an object in contact with the soft skin accurately. In the paper, we illustrate the structure of the robot skin, and describe the principle of both detecting the position of the reflector chips and sensing the contact location of an object. The robot skin is characterized by the fact that the surface is low cost and easily replacable, and the sensing performance is robust against any electromagnetic disturbance. We then show experimental results for verifying the principles using a wedge-shaped object. For evaluating the sensing accuracy, comparisons are made: 1) between the location of a real convex of the object and that of the corresponding estimated polygon and 2) for the position of two vertices of the object when independent fitting and Lagrangian fitting methods are applied.     

Feedback-based object detection for multi-person pose estimation

In this paper, a novel method is proposed for increasing the performance through coupling of top-down models adjusting the object detector based on a new loss function. Generally, object detectors and keypoint estimators are sequentially used in real-time multi-person pose estimations; however, these two models are separately trained. Therefore, the results of the object detector are not optimized for the keypoint estimator. To solve this problem, we analyze the relationship between the two models and propose a feedback-based loss optimization in the object detector, based on the estimation results of the keypoint estimator. In addition, the resulting bounding box of the object detector is readjusted to improve the accuracy of the keypoint estimation model. The experimental results demonstrate that the proposed approach can perform real-time operations with a high frame rate similar to that of the baseline model. Moreover, it achieved an accuracy of 74.2 average precision (AP), which is higher than the state-of-the-arts model including the human detector used in the experiment.     

基于Focal-EIOU函数的被动式太赫兹图像违禁物品识别

针对被动式太赫兹安检系统因环境影响导致图像质量波动,从而影响识别算法,导致准确率大幅降低的问题,提出了基于Focal-EIOU损失函数的改进YOLOv4算法,并用被动式太赫兹人体安检图像对刀、枪违禁物品进行模型训练获得模型。建立不同环境、不同位置角度携带刀枪嫌疑物人员的太赫兹图像数据库,采用图像增广的方法构建丰富数据集;将YOLOv4的CIOU loss改进为Focal-EIOU loss,提高算法对太赫兹图像识别的鲁棒性,进而经过训练获得较优的模型。在本文的测试集中,使用改进后的算法训练的模型平均检测精确度(mAP)达到96.4%,检测速度在28 ms左右,交并比(IOU)平均值为0.95,在同等条件下高于常规算法,改善了检测识别的效果。实验结果表明,本文方法能够有效提高被动式太赫兹人体安检系统的嫌疑物识别准确率,有利于该项技术在人体安检领域的推广应用。