Bidirectional mutual guidance transformer for salient object detection in optical remote sensing images

Salient object detection in optical remote sensing images presents great challenges due to the characteristics of remote sensing images such as cluttered background, varying object scales, and unstable imaging conditions, etc. In this paper, we present a Bidirectional Mutual Guidance Transformer (BMGT), which mitigates the locality issue of CNN-based models, and exploits the mutual guidance between global context-aware object representations and fine-grained boundary structures. It contains a hierarchically structured Transformer encoder that extracts multi-level multi-scale token representations, and a dual-stream cross-task MLP decoder that performs joint salient object detection and salient boundary detection in an end-to-end manner. In particular, the dual-stream decoder consists of two sub-branch networks with symmetric architectures, that are connected by a newly proposed Mutual Guidance MLP layer (MG-MLP). Through MG-MLP, salient object features and salient boundary features interact with each other, facilitating complementary learning at multiple network levels. Extensive evaluations demonstrate that our proposed method outperforms other existing methods in two public remote sensing image benchmarks. It proves that our BMGT is advantageous in exploiting long-range context dependencies as well as preserving fine-grained boundary structures.