当前位置: X-MOL 学术Neurocomputing › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
MAFNet: Multi-style attention fusion network for salient object detection
Neurocomputing ( IF 6 ) Pub Date : 2021-01-01 , DOI: 10.1016/j.neucom.2020.09.033
Yanhua Liang , Guihe Qin , Minghui Sun , Jie Yan , Huiming Jiang

Abstract Salient object detection based on deep learning has become one of the research hotspots in computer vision. How to effectively extract useful information is a key issue for saliency detection. Most of the existing methods integrate features extracted from convolutional neural networks indiscriminately. However, the features of different layers have different characteristics, not all of them are useful for saliency detection and some even cause interferences. To solve above problem, we propose a Multi-style Attention Fusion Network (MAFNet). Specifically, MAFNet is mainly consists of dual-cues spatial attention module (DSA), dual attention intermediate representation module (DAIR), high-level channel attention module (HCA) and multi-level feature fusion module (MLFF). Among them, DSA aims to refine low-level features and filter background noise. DAIR uses two branches to adaptively integrate the spatial and semantic information of middle-level features. HCA obtains the semantic features of high-level blocks through two different channel-wise operations. Besides, MLFF effectively integrates the above multi-level features in a learnable manner. Finally, different from cross-entropy, cross-IOU loss guides the network to pay more attention to local details. Experimental results on six public datasets demonstrate that MAFNet’s outperformance is competitive in saliency detection and performs well on small object detection.

中文翻译:

MAFNet:用于显着目标检测的多风格注意力融合网络

摘要 基于深度学习的显着目标检测已成为计算机视觉领域的研究热点之一。如何有效地提取有用信息是显着性检测的关键问题。大多数现有方法不加区别地集成从卷积神经网络中提取的特征。然而,不同层的特征具有不同的特性,并非所有特征都对显着性检测有用,有些甚至会造成干扰。为了解决上述问题,我们提出了一种多风格注意力融合网络(MAFNet)。具体来说,MAFNet主要由双线索空间注意力模块(DSA)、双注意力中间表示模块(DAIR)、高级通道注意力模块(HCA)和多级特征融合模块(MLFF)组成。他们之中,DSA 旨在细化低级特征并过滤背景噪声。DAIR 使用两个分支来自适应地整合中层特征的空间和语义信息。HCA 通过两种不同的 channel-wise 操作获取高层块的语义特征。此外,MLFF 以可学习的方式有效地集成了上述多级特征。最后,与cross-entropy不同,cross-IOU loss引导网络更加关注局部细节。在六个公共数据集上的实验结果表明,MAFNet 在显着性检测方面的表现具有竞争力,在小物体检测方面表现良好。HCA 通过两种不同的 channel-wise 操作获取高层块的语义特征。此外,MLFF 以可学习的方式有效地集成了上述多级特征。最后,与交叉熵不同,交叉IOU损失引导网络更加关注局部细节。在六个公共数据集上的实验结果表明,MAFNet 在显着性检测方面的表现具有竞争力,在小物体检测方面表现良好。HCA 通过两种不同的 channel-wise 操作获取高层块的语义特征。此外,MLFF 以可学习的方式有效地集成了上述多级特征。最后,与cross-entropy不同,cross-IOU loss引导网络更加关注局部细节。在六个公共数据集上的实验结果表明,MAFNet 在显着性检测方面的表现具有竞争力,在小物体检测方面表现良好。
更新日期:2021-01-01
down
wechat
bug