摘要
针对胶囊网络(capsule network,CapsNet)特征提取结构单一和数据处理中参数量过大的问题,提出多尺度混合注意力胶囊网络模型。首先,在网络初始端添加不同尺度的卷积核来多角度提取特征,并引入混合注意力机制,通过聚焦更具分辨性的特征区域来降低复杂背景干扰。其次,采用局部剪枝算法优化动态路由,减少参数量,缩短模型训练时间。最后,在海洋鱼类数据集F4K(Fish4Knowledge)上验证,结果表明,与传统残差网络(residual network50,ResNet-50)、双线性网络(bilinear convolutional neural network,B-CNN)、分层精简双线性注意力网络(spatial transformation network and hierarchical compact bilinear pooling,STN-H-CBP)以及CapsNet模型相比,该算法识别精度为98.65%,比ResNet-50模型提升了5.92%;训练时间为2.2 h,相比于CapsNet缩短了近40 min,验证了该算法的可行性。
A multi-scale hybrid attention capsule network(CapsNet)model is proposed to solve the problem of insufficient feature extraction due to single feature extraction structure in CapsNet and excessive amount of parameters in data processing.First,convolution kernels of different scales are added at the initial end of the network to extract features at multiple angles,and channel attention(CA)mechanism and spatial attention(SA)mechanism are introduced to reduce complex background interference by focusing on features of more resolved regions.Second,a local pruning algorithm is adopted to optimize the dynamic routing algorithm,which reduces calculation parameters and training time.Finally,validation on open marine fish data set F4 K(Fish 4 Knowledge)shows that the model recognition accuracy in this paper is 98.65%compared with traditional residual network50(ResNet-50),bilinear convolutional neural network(B-CNN),spatial transformation network and hierarchical compact bilinear pooling(STN-H-CBP)and CapsNet models,5.92%higher than ResNet-50 model;The training time is 2.2 h,which is nearly 40 min shorter than that of CapsNet,which verifies the feasibility of the proposed algorithm.
作者
许学斌
刘燊莲
路龙宾
刘晨光
XU Xuebin;LIU Shenlian;LU Longbin;LIU Chenguang(School of Computer Science and Technology,Xi′an University of Posts&Telecommunications,Xi′an,Shaanxi 710121,China;Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing,Xi′an University of Posts&Telecommunications,Xi′an,Shaanxi 710121,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2022年第11期1158-1164,共7页
Journal of Optoelectronics·Laser
基金
国家自然科学基金面上项目(61673316)
陕西省重点研发计划项目(2017GY-071,2018GY-135)
陕西省教育厅项目(16JK1697)
陕西省技术创新引导计划项目(2017XT-005)
咸阳市科技计划项目(2017K01-25-3)
西安邮电大学研究生创新基金(CXJJLY202004)资助项目
作者简介
许学斌(1974-),男,副研究员,博士/博士后,主要研究方向人工智能、生物特征识别,E-mail:ccp9999@126.com
