Salient object detection simulates the attention mechanism of human behavior to grasp the most attractive objects in the images. Recently edge information has been introduced to enhance the sharp contour in RGB image saliency detection. Inspired by it, we probe into the edge-guided RGB-D image saliency detection. There are two key problems need to be solved. One is how to extract edge information from cross-modal color and depth information, the other is how to fuse the edge feature into double-stream saliency detection network. To solve these two issues, a cross-modal edge-guided salient object detection for RGB-D image is proposed. Based on double-stream U-Net framework, edge information is extracted from the deep and shallow block of both modalities. The feature in deep layer contains sematic information implying where are the object boundaries, so the features of both modalities are directly fused. The feature in shallow layer provides more detailed spatial information, so a gated fusion layer is utilized to fuse the features of both modalities to filter out the depth image noise. Extracted edge feature is fed into decoder combining with color and depth feature to achieve edge-guided cross-modal decoding process. Experimental results show our model outperforms SOTA models based on the edge guidance and gated fusion strategies in cross-modal double-stream network.

显着物体检测可模拟人类行为的注意力机制，以抓住图像中最吸引人的物体。最近，边缘信息已被引入以增强RGB图像显着性检测中的清晰轮廓。受此启发，我们探索了边缘引导的RGB-D图像显着性检测。有两个关键问题需要解决。一种是如何从交叉模式的颜色和深度信息中提取边缘信息，另一种是如何将边缘特征融合到双流显着性检测网络中。为了解决这两个问题，提出了一种用于RGB-D图像的跨模态边缘引导显着目标检测方法。基于双流U-Net框架，从两种模式的深浅块中提取边缘信息。较深层中的要素包含语义信息，表明对象边界在哪里，因此这两种模态的特征是直接融合在一起的。浅层中的特征提供了更详细的空间信息，因此门控融合层用于融合两种模态的特征以滤除深度图像噪声。将提取的边缘特征与颜色和深度特征相结合，馈入解码器，以实现边缘引导的交叉模态解码过程。实验结果表明，该模型在交叉模式双流网络中基于边缘引导和门控融合策略的性能优于SOTA模型。将提取的边缘特征与颜色和深度特征相结合，馈入解码器，以实现边缘引导的交叉模态解码过程。实验结果表明，在交叉模式双流网络中，基于边缘引导和门控融合策略的模型优于基于SOTA的模型。将提取的边缘特征与颜色和深度特征相结合，馈入解码器，以实现边缘引导的交叉模态解码过程。实验结果表明，该模型在交叉模式双流网络中基于边缘引导和门控融合策略的性能优于SOTA模型。