Abstract:

Recently, saliency detection methods based on Fully Convolutional Network（FCN） have made great progress compared with the methods using handcrafted feature, however, there are still some problems to be solved in images with complex scenes. This paper proposes a novel feature-guided salient object detection model through investigating the important role of deep features in convolutional neural network. Based on the decode-encode architecture of feature pyramid network, firstly, it designs a Global Guidance Module（GGM） upon the bottom-up pathway, to extract the location information of salient objects. And then, it further builds a contextual Residual Module（RM） to extract more detail information at different output sides. Moreover, it employs Guiding Flows（GF） to aggregate the features generated by GGM and RM. The proposed approach uses deep semantics to guide the feature learning in shallower layers, which makes the network focus on the salient object and ignore the irrelevant background. Experimental results show that the proposed model improves the performance compared to the state-of-the-art methods on five benchmark datasets.

Key words: saliency detection, fully convolutional network, feature guidance, multi-scale and multi-level feature, residual structure

摘要：

近年来，基于全卷积网络的显著性物体检测方法较手工选取特征的方法已经取得了较大的进展，但针对复杂场景图像的检测仍存在一些问题需要解决。提出了一种新的基于全局特征引导的显著性物体检测模型，研究深层语义特征在多尺度多层次特征表达中的重要作用。以特征金字塔网络的编解码结构为基础，在自底而上的路径中，设计了全局特征生成模块（GGM），准确提取显著性物体的位置信息；构建了加强上下文联系的残差模块（RM），提取各侧边输出的多尺度特征；采用特征引导流（GF）融合全局特征生成模块和残差模块，利用深层语义特征去引导浅层特征提取，高亮显著目标的同时抑制背景噪声。实验结果表明，在5个基准数据集上与11种主流方法相比，该模型具有优越性。

关键词: 显著性检测, 全卷积网络, 特征引导, 多尺度和多层次特征, 残差结构