基于上下文感知跨层特征融合的光场图像显著性检测

Saliency Detection Based on Context-aware Cross-layer Feature Fusion for Light Field Images

光场图像的显著性检测是视觉跟踪、目标检测、图像压缩等应用中的关键技术。然而,现有深度学习方法在处理特征时,忽略特征差异和全局上下文信息,导致显著图模糊,甚至在前景与背景颜色、纹理相似或者背景杂乱的场景中,存在检测对象不完整以及背景难抑制的问题,因此该文提出一种基于上下文感知跨层特征融合的光场图像显著性检测网络。首先,构建跨层特征融合模块自适应地从输入特征中选择互补分量,减少特征差异,避免特征不准确整合,以更有效地融合相邻层特征和信息性系数;同时利用跨层特征融合模块构建了并行级联反馈解码器(PCFD),采用多级反馈机制重复迭代细化特征,避免特征丢失及高层上下文特征被稀释;最后构建全局上下文模块(GCM)产生多尺度特征以利用丰富的全局上下文信息,以此获取不同显著区域之间的关联并减轻高级特征的稀释。在最新光场数据集上的实验结果表明,该文方法在定量和定性上均优于所比较的方法,并且能够精确地从前/背景相似的场景中检测出完整的显著对象、获得清晰的显著图。

Saliency detection of light field images is a key technique in applications such as visual tracking,target detection,and image compression.However,the existing deep learning methods ignore feature differences and global contextual information when processing features,resulting in blurred saliency maps and even incomplete detection objects and difficult background suppression in scenes with similar foreground and background colors,textures,or background clutter.A context-aware cross-layer feature fusion-based saliency detection network for light field images is proposed.First,a cross-layer feature fusion module is built to select adaptively complementary components from input features to reduce feature differences and avoid inaccurate integration of features in order to more effectively fuse adjacent layer features and informative coefficients;Meanwhile,a Parallel Cascaded Feedback Decoder(PCFD)is constructed using the cross-layer feature fusion module to iteratively refine features using a multi-level feedback mechanism to avoid feature loss and dilution of high-level contextual features;Finally,a Global Context Module(GCM)generates multi-scale features to exploit the rich global context information in order to obtain the correlation between different salient regions and mitigate the dilution of high-level features.Experimental results on the latest light field dataset show that the textual method outperforms the compared methods both quantitatively and qualitatively,and is able to detect accurately complete salient objects and obtain clear saliency maps from similar front/background scenes.