基于选择性注意机制的目标区域提取研究  

Research on Target Region Extraction Based onSelective Attention Mechanism

作　　者：王奕婷 高美珍 洪家平 陈英豪 乔丰收 WANG Yiting;GAO Meizhen;HONG Jiaping;CHEN Yinghao;QIAO Fengshou(School of Artificial Intelligence,Beijing Normal University,Beijing 100875;College of Physics and Electron Science,Hubei Normal University,Huangshi 435002;College of Computer and Information Engineering,Hubei Normal University,Huangshi 435002;No.91919 Troops of PLA,Huanggang 438000;Wuhan Marin Communication Research Institute,Wuhan 430079)

机构地区：[1]北京师范大学人工智能学院,北京100875 [2]湖北师范大学物理与电子科学学院,黄石435002 [3]湖北师范大学计算机科学与技术学院,黄石435002 [4]中国人民解放军91919部队,黄冈438000 [5]武汉船舶通信研究所,武汉430079

出　　处：《舰船电子工程》2021年第5期39-43,共5页Ship Electronic Engineering

基　　金：湖北省自然科学基金项目(编号:2018CFB709)资助。

摘　　要：针对目前引入视觉注意机制的目标检测算法实时性不够且难以检测多个目标的问题,论文提出一种基于选择性注意机制的区域提取方法。通过分析输入信号的频域幅值分布规律,构建相应的映射函数,区分不同幅度成分的信息;再进行邻域平均消除部分抖动并作差,去除相对冗余的成分;利用相位信息和变换后的结果重构到空间域,并进行绝对平方运算增强各部分的差异,最后进行滤波和直方图均衡化来增强可视化效果。结果表明,该方法相较Itti模型和GBVS模型计算量显著降低,平均每张图片的处理时间减少一个数量级,且能够更精准地提取目标的轮廓信息。Aiming at the problem that the current target detection algorithm with visual attention mechanism is not real-time enough and cannot adapt to multi-target detection,a region extraction method based on selective attention mechanism is proposed.By analyzing the frequency domain information distribution of the input signal,the corresponding mapping function is constructed to distinguish the information of different amplitude components.Then the neighborhood average is used to eliminate partial jitter and the relative redundant components are removed by subtraction.The phase information and the transformed result are used to recon⁃struct the spatial domain,and the absolute square operation is carried out to enhance the difference of each part.Finally,the filter⁃ing and histogram equalization are carried out to enhance the visualization effect.The results show that compared with the Itti model and GBVS model,the calculation amount of this method is significantly reduced,the processing time of each image is reduced by an order of magnitude,and the contour information of the target can be extracted more accurately.

关 键 词：选择性注意机制 目标区域 频域分析 视觉显著 映射变换 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]