This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Enhancing Satellite Object Localization with Dilated Convolutions and Attention-aided Spatial Pooling
Created by
Haebom
Author
Seraj Al Mahmud Mostafa, Chenxi Wang, Jia Yue, Yuta Hozumi, Jianwu Wang
Outline
This paper addresses the fact that object localization in satellite imagery is particularly challenging due to the diversity of objects, low spatial resolution, interference of noise such as clouds and city lights, and interference of key features. This study focuses on three satellite datasets: upper-atmosphere gravity wave (GW), mesosphere bore, and ocean eddy (OE) to address the unique challenges of each dataset. These challenges include the variability in the scale and shape of key object patterns, where the size, shape, and feature coverage of the objects of interest can vary significantly. To address these challenges, this paper presents YOLO-DCAP, a novel improved version of YOLOv5 designed to improve object localization in complex scenarios. YOLO-DCAP integrates a multi-scale dilated residual convolution (MDRC) block that captures multi-scale features with different dilation rates, and an attention-assisted spatial pooling (AaSP) module that improves feature selection by focusing on globally relevant spatial regions. These structural improvements help to better locate objects in satellite imagery. Experimental results show that YOLO-DCAP significantly outperforms both the YOLO baseline model and the state-of-the-art approaches, achieving an average of 20.95% improvement in mAP50 and 32.23% improvement in IoU over the baseline model, and 7.35% and 9.84% improvement over the state-of-the-art alternatives, respectively. The consistent performance improvements across three satellite datasets highlight the robustness and generalizability of the proposed approach. The code is available at https://github.com/AI-4-atmosphere-remote-sensing/satellite-object-localization .
We present a novel YOLO-DCAP model that improves object localization performance in satellite images.
◦
Consistently excellent performance across a variety of satellite data sets.
◦
Experimentally verifying the effectiveness of MDRC blocks and AaSP modules.
◦
Improving accessibility through open source code disclosure.
•
Limitations:
◦
Limited to performance evaluation on a specific satellite data set. Further verification of generalizability to other types of satellite image data is needed.
◦
Further analysis is needed on the model's computational complexity and real-time processing potential.
◦
Further research is needed on the robustness of the model to various noises and interferences.
