Adjacent-Atrous Mechanism for Expanding Global Receptive Fields: An End-to-End Network for Multiattribute Scene Analysis in Remote Sensing Imagery

• Published in: IEEE transactions on geoscience and remote sensing Vol. 62; pp. 1 - 19
• Main Authors: Li, Zhengpeng, Hu, Jun, Wu, Kunyang, Miao, Jiawei, Wu, Jiansheng
• Format: Journal Article
• Language: English
• Published: New York IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adjacent-atrous global context modeling; attention mechanism; Computer applications; Context modeling; Convolutional neural networks; Feature extraction; global receptive fields; Image processing; Image resolution; local perception; Modelling; Modules; multiattribute scene understanding (MASU); Receptive field; Remote sensing; Scene analysis; Semantics; Spatial discrimination learning; Task analysis; Task complexity; Transformers
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3422007

The multiattribute scene understanding (MASU) tasks currently lie in capturing multiple attribute features and learning the complex correlations between different attributes. Traditional methods primarily focus on exploring multiscale local insights and employ direct approaches to merge global semantic data into the image models, thereby neglecting the full spectrum of global semantic features across different receptive fields. Furthermore, encapsulating a wide range of spatial details through deeper networks inevitably leads to a drastic increase in computational complexity. To address these challenges, we propose a novel end-to-end network named adjacent-atrous mechanism for expanding global receptive fields (AMEGRF-Net). Specifically, we introduce an efficient local-global feature learning paradigm that innovatively expands the model's receptive field to enhance scene understanding without incurring additional computational overhead. A local feature sensing (LFS) module is proposed to enhance the distinctiveness between different categories within the feature space while refining the spatial feature learning capability and interchannel synergy. We present an innovative adjacent-atrous mechanism, adjacent-atrous global context modeling module (AGCM), to combine a broader global receptive field with a complex relationship capturing mechanism, achieving deep modeling of the intricate relationships between attributes and labels. Through extensive comparative experiments on three challenging public datasets, the superior performance of AEGRF-Net in handling high-resolution remote sensing images for MASU has been clearly demonstrated.