MSNet

Abstract

Remote sensing object detection (RSOD) encounters challenges in effectively extracting features of small objects in remote sensing images (RSIs). To alleviate these problems, we proposed a Multi-Scale Network for Object Detection in Remote Sensing Images (MSNet) with multi-dimension feature information. Firstly, we design a Partial and Pointwise Convolution Extraction Module (P2CEM) 2 CEM) to capture feature of object in spatial and channel dimension simultaneously. Secondly, we design a Local and Global Information Fusion Module (LGIFM), designed local information stack and context modeling module to capture texture information and semantic information within the multi-scale feature maps respectively. Moreover, the LGIFM enhances the ability of representing features for small objects and objects within complex backgrounds by allocating weights between local and global information. Finally, we introduce Local and Global Information Fusion Pyramid (LGIFP). With the aid of the LGIFM, the LGIFP enhances the feature representation of small object information, which contributes to dense connection across the multi-scale feature maps. Extensive experiments validate that our proposed method outperforms state-of-the-art performance. Specifically, MSNet achieves mean average precision (mAP) scores of 75.3%, 93.39%, 96.00%, and 95.62% on the DIOR, HRRSD, NWPU VHR-10, and RSOD datasets, respectively. Show more