一种融合上下文信息及自适应感受野的多尺度目标检测算法  

AN MULTI-SCALE OBJECT DETECTION ALGORITHM COMBINING CONTEXT INFORMATION AND ADAPTIVE RECEPTIVE FIELD

作　　者：张婷 兰时勇[1] Zhang Ting;Lan Shiyong(National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610064,Sichuan,China)

机构地区：[1]四川大学视觉合成图形图像技术国防重点学科实验室,四川成都610064

出　　处：《计算机应用与软件》2024年第10期314-318,共5页Computer Applications and Software

基　　金：四川省科技厅重点研发项目(2021YFG0300);视觉合成图形图像技术国防重点学科实验室开放研究项目。

摘　　要：目标检测在实际应用各类复杂场景中面临着诸多的挑战,如目标遮挡、光照变化、目标尺度变化等。为了提高多尺度目标检测的性能,提出一种改进的特征金字塔(FPN)的目标检测算法。以特征金字塔网络框架为基础引入上下文信息融合模块,充分利用目标对象与其周围环境的关联属性,增强宽动态尺度范围的目标对象的特征表征,提高不同尺度目标的辨识能力。此外,构建一个跨通道注意机制,自适应调整不同尺度目标特征的通道灵敏度,学习到适应目标尺度的感受野范围。该算法在Pascal VOC数据集训练验证,其平均精确率(mAP)比基准方法提高了3%。Object detection faces many challenges in the practical application of various complex scenes,such as object occlusion,illumination changes,and object size changes in the practical application.In order to improve the performance of multi-scale target detection,this paper proposes an improved feature pyramid network(FPN)target detection algorithm.Based on the FPN framework,the context information fusion was introduced to utilize the relevance of an object to its surrounding environment and enhance feature representation of objects for wide dynamic range images and to improve the ability of detection ability for different scales.In addition,a cross-channel attention mechanism was constructed to adaptively adjust the channel sensitivity of target features at different scales.Experiments on the Pascal VOC dataset show that the proposed method improves the detection performance by 3%compared with the baseline method in terms of mean average precision(mAP).

关 键 词：目标检测 上下文信息融合 跨通道注意力机制 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]