基于分层基因优选多特征融合的图像材质属性标注  

Stratified Gene Selection Multi-Feature Fusion for Image Material Attribute Annotation

作　　者：张红斌[1] 邱蝶蝶 邬任重 蒋子良 武晋鹏 姬东鸿[2] ZHANG Hong-Bin;QIU Die-Die;WU Ren-Zhong;JIANG Zi-Liang;WU Jin-Peng;JI Dong-Hong(Software School,East China Jiaotong University,Nanchang 330013;School of Cyber Science and Engineering,Wuhan University,Wuhan 430072)

机构地区：[1]华东交通大学软件学院,南昌330013 [2]武汉大学国家网络安全学院,武汉430072

出　　处：《自动化学报》2020年第10期2191-2213,共23页Acta Automatica Sinica

基　　金：国家自然科学基金(61762038,61861016);教育部人文社会科学研究一般项目(20YJAZH1142);江西省自然科学基金(20202BABL202044);江西省科技厅重点研发计划(20171BBG70093,20192BBE50071);江西省教育厅科学技术研究项目(GJJ190323)资助。

摘　　要：图像材质属性标注在电商平台、机器人视觉、工业检测等领域都具有广阔的应用前景.准确利用特征间的互补性及分类模型的决策能力是提升标注性能的关键.提出分层基因优选多特征融合(Stratified gene selection multi-feature fusion,SGSMFF)算法:提取图像传统及深度学习特征;采用分类模型计算特征预估概率;改进有效区域基因优选(Effective range based gene selection,ERGS)算法,并在其中融入分层先验信息(Stratified priori information,SPI),逐层、动态地为预估概率计算ERGS权重;池化预估概率并做ERGS加权,实现多特征融合.在MattrSet和Fabric两个数据集上完成实验,结果表明:SGSMFF算法中可加入任意分类模型,并实现多特征融合;平均值池化方法、分层先验信息所提供的难分样本信息、"S+G+L"及"S+V"特征组合等均有助于改善材质属性标注性能.在上述两个数据集上,SGSMFF算法的精准度较最强基线分别提升18.70%、15.60%.Material attribute annotation can be broadly applied in many different scenarios in large-scale product image retrieval, robotics and industrial inspection. Accurately utilizing the complementarity between different image features and the decision abilities of classification models is the key factor to improve the final annotation performance.To address the problem, a novel algorithm called stratified gene selection multi-feature fusion(SGSMFF) for material attribute annotation is proposed. Both the traditional and deep learning image features are extracted firstly. Then any classification model is utilized to compute the estimated probability of each image feature. The traditional effective range based gene selection(ERGS) algorithm is modified in turn and the stratified priori information(SPI) obtained from two perspectives is integrated into the modified ERGS algorithm to dynamically compute the ERGS weight of each estimated probability. Two pooling strategies i. e. Maximum and Average are proposed to complete the final multifeature fusion procedure. The proposed SGSMFF algorithm is validated on two different datasets: MattrSet and Fabric.Experimental results demonstrate that any classification model can be integrated into the innovative SGSMFF algorithm.Several fundamental factors such as the proposed Average pooling strategy, the hard negative information provided by the stratified priori information, and the feature combinations including "S + G + L" and "S + V" all help improve thefinal annotation performance. Our approach significantly outperforms state-of-the-art baseline about 18.70 % and 15.60 %on the above datasets respectively.

关 键 词：材质属性标注 分层基因优选 多特征融合 预估概率 分层先验信息 难分样本信息 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]