多尺度特征提取和多级别特征融合的显著性目标检测方法

目前主流的显著性目标检测方法通常采用短连接加权的方式融合多级别特征信息,这种方式无法精准有效的控制信息流的传递。而且,现有的检测方法通常采用单一的特征检测,导致显著性目标区域与背景的边界不连续、易模糊。因此,本文提出一种多尺度特征提取和多级别特征融合的显著性目标检测方法。首先,利用不同扩张率的空洞卷积获取多尺度的上下文信息,弥补单一特征检测带来的不足。其次,提出一个多级别特征融合模块,该模块有效的利用浅层特征、深层特征和全局上下文特征信息之间的分布特性进行融合,不仅可以抑制噪声的传递,而且可以更有效地恢复显著性目标的空间细节结构信息。在5个公开的数据集上进行的实验结果表明: 相比较其它13种主流的检测方法,本文方法检测的显著图边缘轮廓连续性更好、空间结构细节信息更清晰,在综合指标(F-measure)、平均绝对误差(MAE)、结构化度量(S-measure)、精准率-召回率(PR)曲线和F-score曲线等指标上均有明显的提升。

Salient Object Detection Based on Multi-scale Feature Extraction and Multi-level Feature Fusion

Current mainstream detection methods fuse multi-level feature information through short connection to add feature maps, which cannot accurately and effectively control the information. In addition, existing salient detection methods usually use single feature detection, which results in discontinuous and fuzzy boundary between the saliency object region and the background. A new salient object detection method based on multi-scale feature extraction and multi-level feature fusion is proposed in this paper. Firstly, we obtain multi-scale context information by using the dilated convolution of different expansion rates to make up for the deficiencies caused by single feature detection. Secondly, a multi-level feature fusion module is designed, which fuses low-level feature, high-level feature and global context feature information for different distribution characteristics of them. It can not only restrain the transmission of noise, but also restore the spatial detail structure information of the saliency object effectively. Experiments on five public datasets show that compared with other thirteen mainstream detection methods, the saliency map predicted by our method has better continuity of edge contours and clearer details of spatial structure details. Moreover, our method is superior to other methods in terms of F-measure, mean absolute error (MAE), structure measure (S-measure), precision-recall (PR) curves and F-score curves.