YOLSO: You Only Look Small Object

Small object detection is challenging and far from satisfactory. Most general object detectors suffer from two critical issues with small objects: (1) Feature extractor based on classification network cannot express the characteristics of small objects reasonably due to insufficient appearance information of targets and a large amount of background interference around them. (2) The detector requires a much higher location accuracy for small objects than for general objects. This paper proposes an effective and efficient small object detector YOLSO to address the above problems. For feature representation, we analyze the drawbacks in previous backbones and present a Half-Space Shortcut(HSSC) module to build a background-aware backbone. Furthermore, a coarse-to-fine Feature Pyramid Enhancement(FPE) module is introduced for layer-wise aggregation at a granular level to enhance the semantic discriminability. For loss function, we propose an exponential L1 loss to promote the convergence of regression, and a focal IOU loss to focus on prime samples with high classification confidence and high IOU. Both of them significantly improves the location accuracy of small objects. The proposed YOLSO sets state-of-the-art results on two typical small object datasets, MOCOD and VeDAI, at a speed of over 200 FPS. In the meantime, it also outperforms the baseline YOLOv3 by a wide margin on the common COCO dataset.     

A comprehensive survey on computer vision based concepts,methodologies, analysis and applications for automatic gun/knife detection

The ability to detect gun and gun held in hand or other body parts is a typical human skill. The same problem presents an imperative task for computer vision system. Automatic observer independent detection of hand held gun or gun held in the other body part, whether it is visible or concealed, provides enhance security in vulnerable places and initiates appropriate action there. Compare to the automatic object detection systems, automatic detection of gun has very few successful attempts. In the present scope of this paper, we present an extensive survey on automatic detection of gun and define a taxonomy for this particular detection system. We also describe the inherent difficulties related with this problem. In this survey of published papers, we examine different approaches used in state-of-the-art attempts and compare performances of these approaches. Finally, this paper concludes pointing to the possible research gaps in related fields.     

Organic mesh template-based laminated object manufacturing to fabricate ceramics with regular micron scaled pore structures

A new aqueous slurry-based laminated object manufacturing process for porous ceramics is proposed: firstly, an organic mesh sheet is pre-paved as a pore-forming template before slurry layer scraping; secondly, the 2D pattern is built with laser outline cutting of the dried mesh–ceramic composite layer; finally, the pore structure is formed after degreasing and sintering. Alumina parts with porosities of 51.5 %, round hole diameters of 80 ± 5 μm were fabricated using 70 wt. % solid content slurry and 100 mesh nylon net. Using an organic mesh as the framework and template not only reduces the risk of damage of the green body but also ensures the regularity, uniformity and connectivity of the micron scaled pore network. The layer-by-layer drying method avoids the delamination phenomenon and improves the paving density. The new method can realize the flexible design of the pore structure by using various organic mesh templates.     

多先验融合的图像显著性目标检测算法

为了更加准确地检测出图像中的显著性目标，提出了多先验融合的显著性目标检测算法。针对传统中心先验对偏离图像中心的显著性目标会出现检测失效的情况，提出在多颜色空间下求显著性目标的最小凸包交集来确定目标的大致位置，以凸包区域中心计算中心先验。同时通过融合策略将凸包区域中心先验、颜色对比先验和背景先验融合并集成到特征矩阵中。最后通过低秩矩阵恢复模型生成结果显著图。在公开数据集MSRA1000和ESSCD上的仿真实验结果表明，MPLRR能够得到清晰高亮的显著性目标视觉效果图，同时F，AUC，MAE等评价指标也比现有的许多方法有明显提升。     

基于张量分解融合RGB-D图像的物体识别

为了充分利用RGB-D图像的深度图像信息，提出了基于张量分解的物体识别方法。首先将RGB-D图像构造成一个四阶张量，然后将该四阶张量分解为一个核心张量和四个因子矩阵，再利用相应的因子矩阵将原张量进行投影，获得融合后的RGB-D数据，最后输入到卷积神经网络中进行识别。RGB-D数据集中三组相似物体的识别结果表明，利用张量分解融合RGB-D图像的物体识别准确率高于未采用张量分解的物体识别准确率，并且单一错分实例的准确率最高可提升99%。     

Stream-Based Data Sampling Mechanism for Process Object

Process object is the instance of process. Vertexes and edges are in the graph of process object. There are different types of the object itself and the associations between object. For the large-scale data, there are many changes reflected. Recently, how to find appropriate real-time data for process object becomes a hot research topic. Data sampling is a kind of finding c hanges o f p rocess o bjects. There i s r equirements f or s ampling to be adaptive to underlying distribution of data stream. In this paper, we have proposed a adaptive data sampling mechanism to find a ppropriate d ata t o m odeling. F irst o f all, we use concept drift to make the partition of the life cycle of process object. Then, entity community detection is proposed to find changes. Finally, we propose stream-based real-time optimization of data sampling. Contributions of this paper are concept drift, community detection, and stream-based real-time computing. Experiments show the effectiveness and feasibility of our proposed adaptive data sampling mechanism for process object.     

一种基于改进NMS的牛脸检测方法

针对牲畜牛身份认证的多牛脸检测场景，本文给出一种基于改进Faster R-CNN的牛脸检测方法。使用Inception v2替换ZF网络作为Faster R-CNN的基础网络，模型精度得到显著提升；针对多牛检测场景对NMS（Non-Maximum Suppression）进行相应优化，使模型的召回率得到显著提升。通过和其他目标检测模型对比实验，本文的改进模型在精确率和召回率上均优于其他模型。     

一种基于小字典不对等语料的跨语言词嵌入方法

双语词嵌入通常采用从源语言空间到目标语言空间映射,通过源语言映射嵌入到目标语言空间的最小距离线性变换实现跨语言词嵌入。然而大型的平行语料难以获得,词嵌入的准确率难以提高。针对语料数量不对等、双语语料稀缺情况下的跨语言词嵌入问题,该文提出一种基于小字典不对等语料的跨语言词嵌入方法,首先对单语词向量进行归一化,对小字典词对正交最优线性变换求得梯度下降初始值,然后通过对大型源语言(英语)语料进行聚类,借助小字典找到与每一聚类簇相对应的源语言词,取聚类得到的每一簇词向量均值和源语言与目标语言对应的词向量均值,建立新的双语词向量对应关系,将新建立的双语词向量扩展到小字典中,使得小字典得以泛化和扩展。最后,利用泛化扩展后的字典对跨语言词嵌入映射模型进行梯度下降求得最优值。在英语—意大利语、德语和芬兰语上进行了实验验证,实验结果证明该文方法可以在跨语言词嵌入中减少梯度下降迭代次数,减少训练时间,同时在跨语言词嵌入上表现出较好的正确率。     

深度学习识别光网络单元故障的设计与应用

为解决依赖装维上门鉴别光网络单元故障带来的不便,可以从机器视觉入手实现自动化故障识别。近年,ImageNet挑战赛的成功推动了物体识别技术的跨越式发展,特别是基于卷积的深度学习技术在视觉识别方面已经达到人类水平,为光网络单元故障的自动识别提供了技术基础。文章对识别光网络单元的工作状态进行了研究,将设备工作状态分为7个场景,提出了利用手机APP采集图片识别故障的解决方案并投入了实际生产;重点阐述了深度学习模块的设计与实现,提出一种通过算法整合的方式综合运用物体检测和图像分类算法,分3阶段逐步求精,解决了图片过滤,光网络单元型号和状态识别等问题,实现了基于计算机视觉自动识别光网络单元故障。从数据上看产品的端到端准确率超过84%,识别速度达到10 FPS,月均提供服务超过1万人次,在减少用户等待的同时节约了人力资源。