基于注意力感知的RGB-D显著性检测  

Attention-aware for RGB-D salient object dectection

作　　者：李琦 戴蒙 张晴 刘云翔 LI Qi;DAI Meng;ZHANG Qing;LIU Yun-xiang(School of Computer Science and Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China)

机构地区：[1]上海应用技术大学计算机科学与信息工程学院,上海201418

出　　处：《计算机工程与设计》2023年第2期511-517,共7页Computer Engineering and Design

基　　金：上海市自然科学基金项目(19ZR1455300、21ZR1462600)。

摘　　要：为在进行RGB-D显著性检测时能高效利用RGB信息与深度信息,对跨模态的融合进行研究。区别于现有方法忽略图像中干扰因素的存在或不考虑图像初始信息的缺陷,设计一个基于注意力机制的跨模态融合模块(CFM)。通过空间注意力与通道注意力有效过滤深度特征以及调制后特征的不必要信息,集成RGB特征与深度特征,让二者实现更好的互补。为保证全局信息与初始彩色信息的不丢失,通过残差连接将原始特征作为补充信息。在5个数据集上使用5个评价指标进行的实验结果表明,与当前主流模型相比,该模型具有优越性。To utilize RGB-data and depth data efficiently in RGB-D significance detection, cross-mode fusion was studied. Diffe-rent from the existing methods, which ignored the existence of interference factors in the image or failed to consider the initial information of the image, a cross-modal fusion module(CFM) based on the attention mechanism was designed. The unnecessary information of depth feature and modulated feature were effectively filtered through spatial and channel attention, and RGB feature and depth feature were integrated to make them better complement each other. To prevent the global information and the initial color information from losing, the original feature was used as supplementary information through residual connection. Results of experiments on five data sets with five evaluation indexes show that the proposed model is superior to the current mainstream model.

关 键 词：显著性检测 跨模态 空间注意力 通道注意力 全局信息 彩色信息 残差连接 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]