Lu Wei ; Chen Si-Bao ; Ding Chris H. Q. ; Tang Jin ; Luo Bin

Abstract

Remote sensing (RS) visual tasks have gained significant academic andpractical importance. However, they encounter numerous challenges that hindereffective feature extraction, including the detection and recognition ofmultiple objects exhibiting substantial variations in scale within a singleimage. While prior dual-branch or multi-branch architectural strategies havebeen effective in managing these object variances, they have concurrentlyresulted in considerable increases in computational demands and parametercounts. Consequently, these architectures are rendered less viable fordeployment on resource-constrained devices. Contemporary lightweight backbonenetworks, designed primarily for natural images, frequently encounterdifficulties in effectively extracting features from multi-scale objects, whichcompromises their efficacy in RS visual tasks. This article introduces LWGANet,a specialized lightweight backbone network tailored for RS visual tasks,incorporating a novel lightweight group attention (LWGA) module designed toaddress these specific challenges. LWGA module, tailored for RS imagery,adeptly harnesses redundant features to extract a wide range of spatialinformation, from local to global scales, without introducing additionalcomplexity or computational overhead. This facilitates precise featureextraction across multiple scales within an efficient framework.LWGANet wasrigorously evaluated across twelve datasets, which span four crucial RS visualtasks: scene classification, oriented object detection, semantic segmentation,and change detection. The results confirm LWGANet's widespread applicabilityand its ability to maintain an optimal balance between high performance and lowcomplexity, achieving SOTA results across diverse datasets. LWGANet emerged asa novel solution for resource-limited scenarios requiring robust RS imageprocessing capabilities.