TUOD遮挡图像库的设计与实现

Design and Implementation of Tsinghua University Occlusion Image Database

遮挡问题是复杂场景图像中一个普遍存在的现象,探索遮挡对图像认知的影响规律、建立具有抗遮挡能力的认知模型直接关系到计算机视觉技术的实际应用,是一个迫切需要解决的科学问题。通过研究复杂场景图像中的遮挡问题,探索遮挡对图像认知的影响规律,建立一个评估检测识别算法的抗遮挡能力、研究图像认知模型及抗遮挡规律的TUOD(Tsinghua University Occlusion Database)遮挡图像库。首先,根据遮挡对图像识别的影响,提出遮挡部件、遮挡面积、遮挡关系、遮挡复杂度4个维度的图像遮挡属性,建立了图像遮挡程度量化标准;其次,基于遮挡维度提出一个新的层次化图像库组织结构,以此为基础进行数据库构建。从PASCALVOC和ImageNet中进行图像筛选和处理,构建了一个包括飞机、车辆、人、动物4大类,共2 100张图片的TUOD遮挡图像库。利用TUOD图像库,结合机器学习理论,通过实验比较分析不同遮挡维度对Faster R-CNN算法的影响。实验表明,TUOD遮挡图像库能够为算法的抗遮挡能力提供量化评估标准。TUOD遮挡图像库的建立为提高抗遮挡算法的性能奠定了基础,具有实用性。

Occlusion is a common phenomenon in images characteristic of complex scenes.Discovering the pattern of how occlusion affects image cognition and establishing a cognition model insusceptible to occlusion is closely related to the utilization of computer vision technologies,and it is also an important and pressing scientific problem to be solved.By analyzing occlusion in complex scenes and how occlusion affects image cognition,this paper established the Tsinghua University Occlusion Image Database for evaluating the anti-occlusion performance of algorithms and studying image cognition model.Firstly,based on occlusion’s impact on image cognition,this paper proposed a 4-dimension occlusion attribute including occluded part,occluded area,occlusion relationship and occlusion complexity,as well as a quantification standard for the extent of occlusion.Then we proposed a novel hierarchical dataset structure,based on which the database could be constructed.This paper established TUOD database which consists of 2 100 images extracted from PASCAL VOC and Image Net databases.Those images covered 4 major object types:aeroplane,car,person and animal.An experiment was conducted to analyze the influence of each dimension of occlusion attribute on the performance of Faster R-CNN using images in TUOD.As is shown in the aforementioned experiment,TUOD database can provide quantitative criteria for the anti-occlusion performance of algorithms and thus it is highly practical and lays the foundation for improving the anti-occlusion performance of object recognition algorithms in complex scenes.