基于注意力机制的显著性目标检测方法  

Salient object detection method based on the attention mechanism

作　　者：王凯诚 鲁华祥[1,3,4] 龚国良 陈刚[1] WANG Kaicheng;LU Huaxiang;GONG Guoliang;CHEN Gang(Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China;School of Future Technology,University of Chinese Academy of Sciences,Beijing 100089,China;Center for Excellence in Brain Science and Intelligence Technology,Chinese Academy of Sciences,Shanghai 200031,China;Semiconductor Neural Network Intelligent Perception and Computing Technology Beijing Key Lab,Beijing 100083,China)

机构地区：[1]中国科学院半导体研究所,北京100083 [2]中国科学院大学未来技术学院,北京100089 [3]中国科学院脑科学与智能技术卓越创新中心,上海200031 [4]半导体神经网络智能感知与计算技术北京市重点实验室,北京100083

出　　处：《智能系统学报》2020年第5期956-963,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61701473);中国科学院STS计划项目(KFJ-STS-ZDTP-070);北京市科技计划项目(Z181100001518006);中国科学院国防科技创新基金项目(CXJJ-17-M152);中国科学院战略性先导科技专项(A类)(XDA18040400).

摘　　要：针对目前主流的基于全卷积神经网络的显著性目标检测方法,受限于卷积层感受野大小,低层特征缺少全局性的信息,而高层特征由于多次池化操作分辨率较低,无法准确地预测目标边缘等细节的问题,本文提出了基于注意力的显著性目标检测方法。在ResNet-50网络中加入注意力精炼模块,利用训练样本的显著真值图对空间注意力进行有监督的学习,使得不同像素位置的相关性更准确。通过深度融合多尺度的特征,用低层特征优化高层特征,精修网络的预测结果使其更加准确。在DUT-OMRON和ECSSD数据集上的测试结果显示,本文方法能显著提升检测效果,F-measure和平均绝对误差都优于其他同类方法。Salient object detection simulates human visual mechanism.At present,the mainstream methods are based on fully convolutional neural networks.Limited by the receptive fields of convolution layers,low-level features lack a global description of images,whereas high-level features are too coarse to accurately segment details of objects,such as edges,because of multi-stage downsampling operations.To solve this problem,we propose a salient object detection method based on the attention mechanism.We introduce novel attention refinement modules.The ground-truth attention calculated from the training datasets is employed to supervise spatial attention.Through this method,the network learns more accurate position relevance between different pixels.In addition,to refine the output salient maps,we gradually combine the multi-scale features and optimize low-layer features with high-layer features.Sufficient experiments on DUT-OMRON and ECSSD datasets have demonstrated that the proposed method outperforms the others in terms of the value of the F measure and mean absolute error.

关 键 词：显著性目标检测 深度学习 全卷积神经网络 视觉注意力 多尺度特征 图像处理 人工智能 计算机视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]