.png){kind=icon}
.png){kind=icon}
Return
Small object detection for UAV images based on deeper poly kernel inception block and space-to-depth convolution
.png){kind=icon}
PRE
AI
X
H
江
X
DOI：10.1007/s10586-026-06432-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To deeply integrate features from multi-scale convolution kernels and enhance their complementarity, a deeper Poly Kernel Inception Block (D-PKIBlock) is designed with a Deeper Parallelized Patch-Aware Attention (DPPA) mechanism. In view of the dense distribution of small objects in UAV images, SPDConv has been employed in the backbone and the neck to enhance the network’s ability to extract fine features. The traditional upsampling operation is replaced by Dynamic Sampling (Dysample) to retain the fine feature of small objects. To improve computational efficiency, the convolution block attention module (CBAM) is replaced with a simple parameter-free attention module (SimAM) in YOLO framework. The experiments on the VisDrone2019 dataset show that the proposed module increases the value of mAP0.5:0.95 by 3.28% and mAP0.5 by 4.06%. The designed network outperforms the existing similar networks in terms of detection accuracy.
Keywords:
UAV images
Small object detection
YOLO
Multi-scale feature extraction
Journal
C
IF:
4.1
Papers:
4.8K
Citations:
7.5K
Organization
Cited Papers