.png){kind=icon}
.png){kind=icon}
Return
CSPBoTNet-YOLOv8: a lightweight and optimized architecture for small-object detection in complex aerial imagery
.png){kind=icon}
PRE
AI
Q
G
L
S
DOI：10.1117/1.JEI.35.2.023034.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object detection constitutes a fundamental task in the field of computer vision. However, drone-based aerial detection remains challenging, as small targets within complex backgrounds frequently cause missed detections, false positives, and low detection accuracy. To address these issues, we present CSPBoTNet-YOLOv8, an application-driven and lightweight detection framework derived from YOLOv8s, specifically designed for drone-based small-object detection. The proposed method adopts a coordinated architectural redesign that jointly optimizes lightweight feature extraction, global context modeling, attention-guided feature fusion, and robust localization. First, we replace the backbone's original C2f modules with CSPHeT. This lightweight structure integrates heterogeneous convolution, dual convolution, and CSPBottleneck, significantly decreasing both the number of parameters and the model's size while preserving detection accuracy. Second, we design the dual-path bottleneck transformer block module and integrate it into the backbone to enhance global receptive capability and strengthen feature fusion, improving the small-target detection. Third, we incorporate ResCBAM into the neck to strengthen the model's capacity for focusing on critical information and promote a more efficient flow of gradient information. Finally, we employ WIoUv3 as the bounding box regression loss, which significantly enhances localization accuracy by a dynamically weighting mechanism to emphasize the importance of medium-quality samples. Extensive experiments conducted on the VisDrone2019 dataset demonstrate that CSPBoTNet-YOLOv8 achieves significant performance gains over YOLOv8s, with improvements of 15.4% in mAP0.5 and 11.9% in mAP0.5:0.95, while simultaneously reducing the number of parameters by 13.2%. These results validate the effectiveness of the proposed system-level optimization strategy and provide practical insights for balancing detection accuracy and computational efficiency in resource-constrained drone-based imaging scenarios. (c) 2026 SPIE and IS&T
Keywords:
small-object detection
drone aerial imagery
YOLOv8
CSPHeT
DP-BoTNet
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers