Dense Attention Fluid Network for Salient Object Detection in Optical Remote Sensing Images (2011.13144v1)

Abstract: Despite the remarkable advances in visual saliency analysis for natural scene images (NSIs), salient object detection (SOD) for optical remote sensing images (RSIs) still remains an open and challenging problem. In this paper, we propose an end-to-end Dense Attention Fluid Network (DAFNet) for SOD in optical RSIs. A Global Context-aware Attention (GCA) module is proposed to adaptively capture long-range semantic context relationships, and is further embedded in a Dense Attention Fluid (DAF) structure that enables shallow attention cues flow into deep layers to guide the generation of high-level feature attention maps. Specifically, the GCA module is composed of two key components, where the global feature aggregation module achieves mutual reinforcement of salient feature embeddings from any two spatial locations, and the cascaded pyramid attention module tackles the scale variation issue by building up a cascaded pyramid framework to progressively refine the attention map in a coarse-to-fine manner. In addition, we construct a new and challenging optical RSI dataset for SOD that contains 2,000 images with pixel-wise saliency annotations, which is currently the largest publicly available benchmark. Extensive experiments demonstrate that our proposed DAFNet significantly outperforms the existing state-of-the-art SOD competitors. https://github.com/rmcong/DAFNet_TIP20

• The paper introduces DAFNet, a novel end-to-end network that leverages a dense attention fluid structure with global context-aware mechanisms for optical remote sensing images.
• The paper demonstrates significant improvements over 15 state-of-the-art models using metrics such as F-measure, MAE, and S-measure on a newly constructed 2,000-image dataset.
• The paper highlights the importance of integrating multi-level attention cues and global context to enhance feature discrimination and robustly segment salient objects.

Dense Attention Fluid Network for Salient Object Detection in Optical Remote Sensing Images

Salient object detection (SOD) in optical remote sensing images (RSIs) represents a challenging task distinct from SOD in natural scene images (NSIs). This complexity arises due to the inherent characteristics of RSIs, including diverse background patterns, variable object scales and orientations, and potential noise. This paper introduces the Dense Attention Fluid Network (DAFNet), an architecture designed specifically to address these challenges by leveraging a novel attention mechanism adapted for SOD in optical remote sensing contexts.

DAFNet adopts an end-to-end encoder-decoder architecture integrating a Dense Attention Fluid (DAF) structure and a Global Context-aware Attention (GCA) mechanism. The GCA module enhances feature representation through global context capture, employing a global feature aggregation approach and a cascaded pyramid attention (CPA) framework. This adaptation allows for the management of scale variation in objects, which is a prominent issue in optical RSI-based SOD.

The DAF structure facilitates the propagation of shallow-level attention cues into deeper network layers, improving the accuracy and consistency of the high-level feature attention maps. This structure emulates a consecutive flow of attention information across hierarchical network levels, bolstering the DAFNet's robustness in segmenting salient objects from intricate backgrounds.

The authors constructed a new dataset comprising 2,000 optical RSIs, extended from the previously available ORSSD dataset, to provide a more exhaustive benchmark for evaluating SOD methodologies in optical RSIs. This newly constructed dataset emphasizes a variety of scenarios, including diverse object categories, scale variations, and additional imaging complexities such as shadows and illumination changes.

The experimental results demonstrated that DAFNet significantly outperformed 15 state-of-the-art SOD models across multiple performance metrics, including but not limited to F-measure, mean absolute error (MAE), and S-measure. This performance was consistent across the newly proposed dataset, showing marked improvements in handling both large-scale and small, intricate object features.

Key insights from this research underline the importance of integrating multi-level attention cues and exploiting global contextual dependencies in accurately detecting salient objects in RSIs. The fluid integration of these components in DAFNet not only enhances feature discrimination but also improves the interpretability of complex scenes.

Looking forward, this paper opens up new research avenues in SOD for optical RSIs by emphasizing attention-based architectures’ potential. Future investigations could explore more computationally efficient attention mechanisms, further refine global context modeling techniques, or adapt the DAFNet architecture for other domains within the image processing and computer vision fields. Additionally, research could be directed towards developing models that do not rely on extensive pixel-level annotations for training, improving the generalization across varied remote sensing datasets.

In summary, this paper's contributions significantly advance the field of SOD in optical RSIs, providing both a novel methodological framework with DAFNet and an extensive, challenging dataset to propel further innovations and research efforts.