Cargando…

Real-time airplane detection using multi-dimensional attention and feature fusion

The remote sensing image airplane object detection tasks remain a challenge such as missed detection and misdetection, and that is due to the low resolution occupied by airplane objects and large background noise. To address the problems above, we propose an AE-YOLO (Accurate and Efficient Yolov4-ti...

Descripción completa

Detalles Bibliográficos
Autores principales: Li, Li, Peng, Na, Li, Bingxue, Liu, Hao
Formato: Online Artículo Texto
Lenguaje:English
Publicado: PeerJ Inc. 2023
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10280687/
https://www.ncbi.nlm.nih.gov/pubmed/37346692
http://dx.doi.org/10.7717/peerj-cs.1331
Descripción
Sumario:The remote sensing image airplane object detection tasks remain a challenge such as missed detection and misdetection, and that is due to the low resolution occupied by airplane objects and large background noise. To address the problems above, we propose an AE-YOLO (Accurate and Efficient Yolov4-tiny) algorithm and thus obtain higher detection precision for airplane detection in remote sensing images. A multi-dimensional channel and spatial attention module is designed to filter out background noise information, and we also adopt a local cross-channel interaction strategy without dimensionality reduction so as to reduce the loss of local information caused by the scaling of the fully connected layer. The weighted two-way feature pyramid operation is used to fuse features and the correlation between different channels is learned to improve the utilization of features. A lightweight convolution module is exploited to reconstruct the network, which effectively reduce the parameters and computations while improving the accuracy of the detection model. Extensive experiments validate that the proposed algorithm is more lightweight and efficient for airplane detection. Moreover, experimental results on the airplane dataset show that the proposed algorithm meets real-time requirements, and its detection accuracy is 7.76% higher than the original algorithm.