基于异构特征聚合的局部视图扭曲型纸币识别  

Local View Distorted Banknote Recognition Based on Heterogeneous Feature Aggregation

作　　者：郭玉慧 梁循[1] GUO Yu-Hui;LIANG Xun(School of Information,Renmin University of China,Beijing 100872)

机构地区：[1]中国人民大学信息学院,北京100872

出　　处：《计算机学报》2022年第1期98-114,共17页Chinese Journal of Computers

基　　金：国家自然科学基金(No.62072463);国家社会科学基金(No.18ZDA309);北京市自然科学基金(No.4172032)资助.

摘　　要：如何识别同一物体的不同结构的表现形式,对于机器而言,是一个比较困难的识别工作.本文以易变形的纸币为例,提出了一种基于异构特征聚合的局部视图扭曲型纸币识别方法.首先利用灰度梯度共生矩阵、Haishoku算法和圆形LBP分别获得纹理风格、色谱风格和纹理,这些特征从不同的角度描述了局部纸币图像,然后通过VGG-16、ResNet-18和DenseNet-121网络学习这些不变形特征得到输出特征,将输出特征聚合后输入识别层Softmax,达到三模型融合效果,进而识别局部视图扭曲型纸币.实验结果表明,多特征聚合和不同类型模型融合可以最大可能地捕获图像的语义,在准确率、精度、召回率和F1上优于基于单特征和双特征的识别,且优于单类模型和两类模型融合的识别性能,此外,在准确率和时间复杂度等评价标准下,与已有主流方法相比都取得了相对较好的效果.The same object may has different forms of manifestation for different structures in an unrestricted environment,how to recognize such objects is a relatively difficult recognition task for the machine.The banknote is a kind of object that can be easily distorted,as a result,in this paper,taking the distorted banknotes as an example,a local view distorted banknote recognition method based on heterogeneous feature aggregation was proposed.The texture style,color spectrum style and texture of local view distorted banknotes from multiple views were obtained,and these features describe local view distorted banknote images from different perspectives,so as to capture the semantics of local view distorted banknote images as much as possible.As a result,firstly,the gray gradient co-occurrence matrix was used to obtain texture style by carrying out secondary statistical calculation,the Haishoku algorithm was used to obtain color spectrum style,and the circular LBP was used to obtain texture.Then,considering that the multi-view invariant features describe the distorted banknote image in the local view,the VGG-16,ResNet-18 and DenseNet-121 networks were used for each type of feature respectively in the proposed method,and these three types deep models were fused,that is,these three types models did not recognize separately,but learn the invariant feature to get the output feature,which was normalized and aggregated with the other two output features.The VGG-16 network learned invariant feature of texture style to obtain output feature,the ResNet-18 network learned invariant feature of color spectrum style to obtain output feature,and the DenseNet-121 network learned invariant feature of texture to obtain output feature.These output features were aggregated to obtain aggregated features.After aggregated,aggregated features were input into the recognition layer Softmax to achieve the fusion of three types models,and recognize local view distorted banknote images.We had carried out a lot of experiments to verify the effective

关 键 词：纸币识别 局部视图扭曲 不变形特征 特征聚合 模型融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]