Zobrazit v EDS

Improving UAV Image Target Detection: A Novel Approach Using OptiDETR with Swin Transformer.

Uloženo v:
Podrobná bibliografie
Název: Improving UAV Image Target Detection: A Novel Approach Using OptiDETR with Swin Transformer.
Autoři: Wenlong Ma1 2098862229@qq.com, Weisheng Liu2 succman@163.com
Zdroj: IAENG International Journal of Computer Science. Mar2025, Vol. 52 Issue 3, p771-780. 10p.
Témata: Object recognition (Computer vision), Transformer models, Process capability, Data augmentation, Algorithms
Abstrakt: In the analysis of drone aerial images, object detection tasks are particularly challenging, especially in the presence of complex terrain structures, extreme differences in target sizes, suboptimal shooting angles, and varying lighting conditions, all of which exacerbate the difficulty of recognition. In recent years, the DETR model based on the Transformer architecture has eliminated traditional post-processing steps such as NMS(Non-Maximum Suppression), thereby simplifying the object detection process and improving detection accuracy, which has garnered widespread attention in the academic community. However, DETR has limitations such as slow training convergence, difficulty in query optimization, and high computational costs, which hinder its application in practical fields. To address these issues, this paper proposes a new object detection model called OptiDETR. This model first employs a more efficient hybrid encoder to replace the traditional Trans- former encoder. The new encoder significantly enhances feature processing capabilities through internal and cross-scale feature interaction and fusion logic. Secondly, an IoU ( Intersection over Union) aware query selection mechanism is introduced. This mechanism adds IoU constraints during the training phase to provide higher-quality initial object queries for the decoder, significantly improving the decoding performance. Additionally, the OptiDETR model integrates SW-Block into the DETR de- coder, leveraging the advantages of Swin Transformer in global context modeling and feature representation to further enhance the performance and efficiency of object detection. To tackle the problem of small object detection, this study innovatively employs the SAHI algorithm for data augmentation. Through a series of experiments, It achieved a significant performance improvement of more than two percentage points in the mAP (mean Average Precision) metric compared to current mainstream object detection models. Furthermore, there is a noticeable reduction in computation and memory consumption, demonstrating the excellent performance and practical value of OptiDETR in object detection tasks. [ABSTRACT FROM AUTHOR]
Databáze: Supplemental Index
Popis
Abstrakt:In the analysis of drone aerial images, object detection tasks are particularly challenging, especially in the presence of complex terrain structures, extreme differences in target sizes, suboptimal shooting angles, and varying lighting conditions, all of which exacerbate the difficulty of recognition. In recent years, the DETR model based on the Transformer architecture has eliminated traditional post-processing steps such as NMS(Non-Maximum Suppression), thereby simplifying the object detection process and improving detection accuracy, which has garnered widespread attention in the academic community. However, DETR has limitations such as slow training convergence, difficulty in query optimization, and high computational costs, which hinder its application in practical fields. To address these issues, this paper proposes a new object detection model called OptiDETR. This model first employs a more efficient hybrid encoder to replace the traditional Trans- former encoder. The new encoder significantly enhances feature processing capabilities through internal and cross-scale feature interaction and fusion logic. Secondly, an IoU ( Intersection over Union) aware query selection mechanism is introduced. This mechanism adds IoU constraints during the training phase to provide higher-quality initial object queries for the decoder, significantly improving the decoding performance. Additionally, the OptiDETR model integrates SW-Block into the DETR de- coder, leveraging the advantages of Swin Transformer in global context modeling and feature representation to further enhance the performance and efficiency of object detection. To tackle the problem of small object detection, this study innovatively employs the SAHI algorithm for data augmentation. Through a series of experiments, It achieved a significant performance improvement of more than two percentage points in the mAP (mean Average Precision) metric compared to current mainstream object detection models. Furthermore, there is a noticeable reduction in computation and memory consumption, demonstrating the excellent performance and practical value of OptiDETR in object detection tasks. [ABSTRACT FROM AUTHOR]
ISSN:1819656X