|
Natural camouflage is a crucial defense mechanism for species to ensure their survival. The most common strategies used in natural camouflage are background matching, environmental texture mimicry, and disruptive coloration. The Camouflage Object Segmentation (COS) task focuses on visually segmenting camouflaged objects that blend seamlessly with the surrounding background. This task poses a greater challenge than the Salient Object Segmentation (SOS) task. Currently, most disguised target segmentation models achieve their tasks by enhancing the texture and contour features of the target, fusing various hierarchical image features, or incorporating frequency domain information of the target. However, these existing models are built on a single framework that uses image context aggregation strategies to segment disguised objects. This approach overlooks the rich and diverse texture features in both spatial and frequency domains for targets. To address this issue, we introduce the cue of frequency components information as an auxiliary enhancement method for images. We also design a dual-stream encoder to process and fuse texture features from both RGB images and frequency perspective images, in order to refine the multi-level features of disguised target textures and contour segmentation. Specifically, we propose a frequency-aware aggregation module that fuses multi-scale features of target textures from a frequency perspective, which includes three different scales of offline discrete cosine transform modules and an image fusion module. A dual-stream encoder is designed to explore global and local texture expressions of targets in both spatial and frequency domains, that multi-layer in spatial domain encoder obtains frequency domain features as masks from the corresponding layers of frequency domain encoder. Additionally, a learnable transpose convolution decoder is used to enhance contour capture capability. Experimental results demonstrate that our model achieves state-of-the-art segmentation accuracy.
|
