微型电脑应用2025，Vol.41Issue(11)：1-5,5.
基于改进YOLOX-s的多尺度目标检测算法
Multi-scale Object Detection Algorithm Based on Improved YOLOX-s
摘要
Abstract
With the rapid development of neural networks,object detection is greatly improved in terms of speed and accuracy,but multi-scale object detection is still a major problem in this field.This paper proposes a multi-scale adaptive object detection algorithm based on improved YOLOX for this problem.The receptive field block(RFB)module is integrated into the cross stage partial(CSP)structure of the backbone network to improve the multi-scale feature extraction ability of the network.The coordinate attention(CA)mechanism is introduced to better capture the spatial and semantic information of the target,and ef-fectively deal with the relationship between different target scales.The Neck structure of YOLOX is optimized,three cross-lay-er connection structures are added on the basis of path aggregation network(PANet)to make up for the loss of feature infor-mation,and then the conflict information is filtered in combination with the adaptive spatial feature fusion(ASFF)structure to increase the expression of key information in different scale branches,so that the detection head of each scale can obtain more accurate feature information for object detection at this scale.The proposed algorithm is verified on the VOC 2007 test dataset.Compared with the baseline algorithm,the mAP_5095 and mAP are improved by 2.0％and 0.9％,respectively,which effec-tively improves the performance of the multi-scale object detection algorithm.关键词
深度学习/多尺度目标检测/注意力机制/特征融合Key words
deep learning/multi-scale object detection/attention mechanisms/feature fusion分类
信息技术与安全科学引用本文复制引用
王庆麒,袁文翠,王梅,赵建民..基于改进YOLOX-s的多尺度目标检测算法[J].微型电脑应用,2025,41(11):1-5,5.基金项目
国家自然科学基金资助项目(51774090) （51774090）
教育部产学合作协同育人重点项目(202102486031) （202102486031）
