16 February 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Excellent performance has been demonstrated by convolutional neural network (CNN) in salient object detection for optical remote sensing images (ORSI-SOD). However, the limitations of CNN's feature extraction using sliding window approach hinder the capture of global representations. Therefore, an end-to-end detection model, known as adaptive dual-stream sparse transformer network (ADSTNet), has been proposed for ORSI-SOD and is assisted by the vision transformer. It effectively addresses the compensation issue of global and local information in ORSI-SOD. In particular, an adaptive interaction encoder has been devised, amalgamating the multi-scale sparse transformer (MST) and the pyramid atrous attention (PAA) to constitute the adaptive dual-stream sparse encoder (ADSE). This encoder collaborates with the CNN to enhance longrange dependency modeling and preserve global information more effectively base on local features. Additionally, a directional feature reconfiguration (DFR) is constructed to extract texture details from multiple directional dimensions. Finally, we propose the adaptive feature cascade decoder (AFCD) that synthesizes content information from the foreground, edges, and background to enhance the representational capacity of the image. Furthermore, a structural loss function, known as the weight compensation mechanism, is introduced to balance the performance of boundary and salmap segmentation losses. The proposed model has been demonstrated to outperform 26 stateof-the-art (SOTA) ORSI-SOD methods across eight evaluation metrics on two standard datasets, as evidenced by extensive experiments. Furthermore, to verify its robustness, the generalization performance of the model on the latest challenging ORSI4199 dataset is reported. The code and results for this work can be found at https://github.com/JieZzzoo/ADSTNet
2 July 2024 in IEEE Transactions on Geoscience and Remote Sensing
In the realm of optical remote sensing imagery, tackling the intricate task of detecting salient objects poses challenges that expose limitations in prevailing CNN- and Transformer-based methodologies, particularly in dealing with intricate object topologies. To address this, we present a novel solution, the Recurrent Adaptive Graph Reasoning Network (RAGRNet), tailored to capturing patterns of salient objects within complex and irregular remote sensing images. For the first time, our network orchestrates a simultaneous modeling of both object regions and boundaries, ushering in a comprehensive comprehension of their intrinsic correlations. This pioneering approach involves the deployment of a Multi-Granularity Boundary-Focused Module (MGBF), which effectively crafts enriched boundary features. Going further, the Recurrent Adaptive Graph Reasoning Module (RAGR) scrutinizes the semantic interplay between region and boundary representations, thereby fortifying the learning process of intricately structured object features. Concluding the process, a specialized decoder, termed the Heterogeneous Semantic Cooperative Interconnection (HSCI), systematically fuses region and boundary information in a progressively unified manner. This intricate interplay harnesses their synergistic complementarity to bolster the precision and sophistication of modeling salient objects. In rigorous evaluations across diverse datasets, our proposed network outperforms 38 state-of-the-art (SOTA) methodologies, a testament to its remarkable efficacy. Code and results will be available at https://github.com/JieZzzoo/RAGRNet.
14 July 2025 in IEEE Transactions on Geoscience and Remote Sensing
Based on current theoretical insights and ideas in salient object detection in optical remote sensing images (RSI-SOD), the utilization of boundaries typically involves enhancing object features through boundary-guided assistance. However, this initial approach overlooks the intrinsic connections between regions and boundaries, which are essential for capturing rich contextual features and refining boundary details from a global perspective, thereby providing robust support for the mutual enhancement of information. To further capture the complex relationships between regions and boundaries, we propose a novel RSI-SOD network, the Multi-Strategy Region and Boundary Interaction Network (MRBINet), which employs multiple strategies to analyze potential key information of the objects. Specifically, we design the Boundary-Region Split Module (BRS) to decompose image information and obtain features of regions and boundaries. Based on the obtained information, the Boundary-Guided Reasoning Refinement Module (BGRR) and the Multi-Source Dynamic Cross-Attention Module (MDC) are employed to explore intrinsic connections between regions and boundaries through graph reasoning and interactive attention, respectively, during the SOD process, effectively constraining diffuse regional information through robust boundary guidance. Finally, we utilize the Collaborative Enhancement Cascade Decoder Module (CECD) to recombine and reinforce the refined regions and boundaries, aiming to restore the intrinsic features of the target object. Extensive experiments on three benchmark datasets demonstrate that our method outperforms classical approaches, showing competitive performance. Furthermore, the intrinsic features enhancement of region and boundary contributes to the performance improvement of the RSI-SOD method. The code and results for this work can be found at https://github.com/JieZzzoo/MRBINet.
3 July 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Salient object detection in optical remote sensing images (ORSI-SOD) is a challenging and relatively novel research field. While many convolutional neural networks (CNNs) have been employed for salient object detection in natural scenes, they often overlook the unique properties of optical remote sensing images. These images typically lack clear edges and sharp details, posing challenges in capturing essential features. To address these issues, we propose a new method for ORSI-SOD called the Multi-level Interactive Reverse Guided network (MIRGNet), which coordinates high- and low-level features in a synchronized manner. The model consists of two key components: Global-Pixel Coordination Module (GPC) and the Multi-Level Feature Interactive Module (MFI), which facilitate coordination between the encoder and decoder. Specifically, GPC extracts low-level features and fully mobilizes them to refine detection, while MFI captures high-level features and enhances the representation of salient regions through multi-level interaction. And, we present a Multi-Scale Receptive Field Module (MRF), which strengthens feature representation through the hierarchical combination of various unbalance dilated convolutions, while reducing adverse grid effects and enhancing inter-pixel continuity. Additionally, to promote feature interaction, a Reverse Guided Module (RG) is introduced to compensate for the insufficient information in shallow features by leveraging deep features more effectively. Extensive experiments on two public datasets employing eight evaluation metrics demonstrate that MIRGNet outperforms 30 state-of-the-art methods. The code and results of our method could be found at https://github.com/JieZzzoo/MIRGNet.