Lightweight Target Detection for Computationally Resource-Constrained and Low-cost UAV Platforms Using Dual-Weights Attention Technology

• Cuimei Liu,
• Chuan Lin,
• Guangjie Han,
• Zhongxiang Wang,
• Zhenyu Wang,

Vol. 19, No. 7, July 31, 2025
10.3837/tiis.2025.07.007, Download Paper (Free):

• Resource-constrained and low-cost UAV
• Lightweight target detection
• Depthwise dimension separation bottleneck
• Attention Decoupling Head
• Convolutional Block Attention Module

Abstract

Computationally resource-constrained and low-cost UAV platforms have become pivotal in advancing industrial automation, safety monitoring, and infrastructure inspection, offering transformative socio-economic value. These platforms enable cost-effective, real-time data acquisition in industries such as logistics, energy, and agriculture, driving operational efficiency, predictive maintenance, and risk mitigation. However, deploying sophisticated target detection algorithms on resource-limited UAVs remains a critical challenge, as traditional models often compromise accuracy for efficiency. Addressing this gap, this study proposes Dual-Weights Attention Technology (DWAT), a novel lightweight detection framework tailored for UAV platforms. The methodological innovation of DWAT lies in three key components: (1) A lightweight depthwise dimension separation bottleneck (BottleNeck- DDW), which replaces conventional convolutions with multi-branch parallel processing to reduce parameters while retaining feature richness. Using grouped convolutions and attention-driven pooling, it minimizes computational costs without sacrificing feature extraction. (2) A CBAM-enhanced Neck network that integrates spatial-channel attention to prioritize salient features during multi-scale fusion, enhancing detection robustness in cluttered environments. (3) An Attention Decoupling Head (AD-Head) that decouples classification and localization tasks via task-specific attention mechanisms, addressing spatial misalignment and boosting prediction accuracy. Experimental validation on VOC, VisDrone, HIT-UAV, and FAIR1M2.0 datasets confirms DWAT’s superiority. Compared to leading lightweight models, DWAT achieves 81.6% mAP_0.5 (3% higher than YOLOv11-tiny) and 59.4% mAP_0.5:0.95 (5.8% improvement) with 54.3 FPS inference speed, outperforming YOLOv4-tiny, Drone-YOLO, and MobileNet-YOLO variants. Ablation studies highlight DWAT’s modular synergy: integrating BottleNeck-DDW reduced parameters by 23%, while AD-Head boosted mAP by 4.2%. Visual results further demonstrate DWAT’s capability to detect small, occluded objects with fewer false positives, even in UAV-captured aerial scenes. This work bridges the accuracy-efficiency gap in resource-constrained and low-cost UAV platforms, enabling reliable real-time detection for automated quality control, safety surveillance, and remote monitoring. By balancing minimal computational overhead with high performance, DWAT empowers scalable deployment of UAV technologies in resource-constrained settings, advancing intelligent manufacturing and smart industry ecosystems.

Statistics

Cite this article