A Simple and Effective Method for RGB-T Salient Object Detection
Zhengyi Liu, Bin Zhu, Yacheng Tan, Haitao Chu · 2022
Salient object detection aims to segment the prominent targets in the image which will be applied to intelligent traffic and transportation. Segmenting targets in complex scenes using only RGB images is very challenging, therefore, some researchers acquire thermal infrared images as auxiliary information to segment objects. The task that utilizes both color information and thermal infrared information for salient object detection is called RGB-T SOD. However, recent studies of RGB-T SOD have not fully exploited the information in the different modalities or ignored the problem of positional information during denoise process. In this paper, we first extract the features of both modalities using two ConvNeXt backbones with shared parameters that can model long-range information and local information. Then the redundant information in the auxiliary modality is suppressed by using a feature purification module that both eliminate the noise and retain the positional information of the features before feature fusion. In addition, we introduce a feature guide module to enhance the representation of spatial detail information in the features. Through the perfect cooperation of the three parts, our proposed method shows outstanding performance and outperforms the state-of-the-art methods on three publicly available RGB-T SOD datasets.