Lightweight Object Detection in Complex Scenarios via Structural Reparameterization and Multi-Scale Adaptive Fusion
Due to adverse factors such as dense target distributions, mutual occlusion, illumination variations, and complex background clutter, existing object detection networks still face challenges in balancing detection accuracy and model complexity within real-world scenarios. To address these issues, this paper explores a lightweight multi-scale object detection architecture based on structural reparameterization and dynamic feature fusion by adapting and integrating several complementary mechanisms into the YOLO11s baseline. First, a two-column lightweight reversible backbone network (Dbb_RevCol) is constructed by combining Reversible Column Networks (RevCol) with the Diverse Branch Block (DBB). This design facilitates the preservation of shallow spatial details during forward propagation while compressing inference parameters via structural reparameterization. Second, Depthwise Separable Convolutions (DSConv) and the Content-Aware ReAssembly of FEatures (CARAFE) upsampling operator are incorporated into a streamlined four-level Bidirectional Feature Pyramid Network (DS-BiFPN) to reduce multi-scale fusion overhead and alleviate detail loss during spatial upsampling. Finally, a high-resolution detection branch (160 × 160) cascaded with the Multi-scale Attention (EMA) module is introduced to enhance feature discrimination for small and partially occluded targets in cluttered scenes. Experimental evaluations on the PASCAL VOC 2012 benchmark demonstrate that the proposed model achieves a Precision of 78.1% and an [email protected] of 78.8%, representing improvements of 0.7% and 2.9% over the baseline YOLO11s, respectively, though its strict-threshold [email protected]:0.95 (50.8%) remains comparable to the baseline (51.0%). Meanwhile, the model requires 6.4 M parameters and 24.6 GFLOPs, reducing parameter count by 42.3% and computational cost by 12.8% relative to YOLO11s. The proposed architectural integration provides a practical trade-off between model compression and detection sensitivity at standard IoU thresholds.
Authors
- 靳俊杰 Jin Junjie
- Jinghui Qiao (ORCID: https://orcid.org/0000-0002-7093-1826)
- Linan Zhang (ORCID: https://orcid.org/0009-0000-7655-2862)
Institutions
- Shenyang University of Technology (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-17
- DOI
- https://doi.org/10.3390/app162211404
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00