基于多特征注意力循环网络的显著性检测  

Salient Object Detection Using Multi-Scale Features with Attention Recurrent Mechanism

作　　者：卢珊妹 郭强[1,2] 王任 张彩明 Lu Shanmei;Guo Qiang;Wang Ren;Zhang Caiming(School of Computer Science and Technology,Shandong University of Finance and Economics,Jinan 250014;Shandong Key Laboratory of Digital Media Technology,Jinan 250014;Software College,Shandong University,Jinan 250010)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东省数字媒体技术重点实验室,济南250014 [3]山东大学软件学院,济南250010

出　　处：《计算机辅助设计与图形学学报》2020年第12期1926-1937,共12页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61873145,61802229);山东省自然科学省属高校优秀青年联合基金(ZR2017JL029);山东省自然科学基金(ZR2018BF007);山东省高等学校青创科技计划(2019KJN045);山东省优势学科人才团队。

摘　　要：特征表达是图像显著性检测的关键,现有方法所提取的特征缺乏一定的可辨识性.为此,提出多尺度上下文特征提取机制和注意力循环机制来解决这一问题.多尺度上下文特征提取机制通过空洞卷积增大高层特征的感受野来获取丰富的上下文语义特征,并采用向量聚合策略对特征进行融合.为增强融合特征的可辨识性,利用注意力机制自适应地对卷积特征增加权重以区分每个像素的重要性,使注意力集中于显著性区域,并抑制背景中的干扰信息.在此基础上,采用循环网络能够有效地在空间位置上对卷积特征进行逐步细化,进一步调整了显著性区域及其边缘,从而生成精确的显著图.该方法在5个常用的数据集上与8种相关方法进行了比较.实验结果表明,该方法不仅能够生成更加准确与完整的显著图,而且其MAE和最大F-measure量化性能也有所提升.Feature representation is a key component to salient object detection.However,the features extracted by existing methods lack capability of discrimination for salient object detection.Hence,multi-scale context features extraction mechanism and attention recurrent mechanism are proposed to solve this problem.Specifically,the multi-scale context features extraction mechanism uses atrous convolution,which can expand receptive fields of high-level convolution features to obtain rich context sematic features,and adopts a vector aggregation strategy to fuse these features.In order to enhance the discriminative power of fused features,the proposed method adopts a attention mechanism to distinguish the importance of each pixel by adaptively increasing the weight of convolution features.The attention mechanism can focus on salient regions while suppressing the interference information in the background.Furthermore,a recurrent network is presented to gradually refine the convolution features in spatial position,and adjust salient regions to generate accurate saliency maps.The proposed method is compared with eight state-of-the-art methods on five benchmark datasets.Experimental results show that the method can generate more accurate and complete saliency maps,and achieves good performance in both mean absolute error and maximum F-measure.

关 键 词：空洞卷积 多尺度特征 注意力机制 循环网络 显著性检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]