Content-aware dynamic filter salient object detection network in multispectral polarimetric imagery

Abstract

Salient object detection (SOD) is widely applied in image segmentation, image fusion, and adaptive compression. However, the SOD of visible images in complex scenes remains a prominent problem due to the lack of low-level features. To solve this problem, a Content-aware Dynamic Filter salient object detection Network using visible and polarized mask images is proposed. It can use the prior information on polarization dimension to guide the SOD of visual features. First, to extract information from the MSPI, a salient polarization mask M composed of three channels is generated. Secondly, deep fused features of the M and RGB images are generated by Encoder and DenseNet fusion structures with receptive fields. Finally, the fused features guide the decoder to generate saliency maps through the content-aware dynamic filtering model. The indexes of the salient object detection results given in this paper are superior to the state-of-the-art algorithms, especially for objects in dim, low contrast, or high transparency and other complex scenarios.