| 摘 要: 针对现有小目标检测中的模型过大导致实时工作能力不足的问题,提出一种基于多尺度特征结合的轻量化小目标检测算法 GBF-YOLO(Global-localBiFPNFusionYOLO)。首先,引入改进高效多尺度卷积模块(Efficient Multi-Scale Convolution Plus,EMSCP),提取多尺度特征信息;其次,提出双向特征金字塔网络BiFPN GLSA并替换原颈部网络,减少位置信息的丢失;最后,使用了Soft-NMS优化相邻检测目标的保留判断。实验结果表明,本文算法相较于原模型,参数量和计算量分别减少了34.5%和15.7%,mAP@0.5相较于YOLOv8n提升了12.1个百分点,能够在实现模型轻量化的同时有效提升小目标图像检测性能。 |
| 关键词: 小目标检测 BiFPN 特征融合 EMSCP |
|
中图分类号:
文献标识码: A
|
| 基金项目: 中共甘肃省委宣传部2023年农村电影公益放映第三方监测服务项目(2023zfcg00315) |
|
| Lightweight Small Object Detection Algorithm Based on Multi-Scale Feature Fusion |
|
SUN Wenyun,ZHU Yifan
|
(School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing 210044, China)
wenyunsun@nuist.edu.cn; 202312490767@nuist.edu.cn
|
| Abstract: To address the issue of insufficient rea-l time performance caused by excessively large models in existing small object detection methods, a lightweight small object detection algorithm named GBF-YOLO (Global-local BiFPN Fusion YOLO) based on mult-i scale feature fusion is proposed. Firstly, an improved Efficient Mult-i Scale Convolution Plus(EMSCP) is introduced to extract multi-scale feature information. Secondly, a Bidirectional Feature Pyramid Network-Global to Local Spatial Aggregation Module (BiFPN-GLSA) is proposed to replace the original neck network, reducing the loss of positional information. Finally, Sof-t NMS is employed to optimize the retention judgment of adjacent detected objects. Experimental results demonstrate that the proposed algorithm reduces the number of parameters and computational cost by 34.5% and 15.7% , respectively, compared to the original model. Meanwhile, the mAP@0.5 is improved by 12.1 percentage points compared to YOLOv8n. The algorithm effectively enhances the detection performance for small object images while achieving model lightweighting. |
| Keywords: small object detection BiFPN feature fusion EMSCP |
