注意力优化RetinaNet的多尺度目标检测算法  

Multi-scale Target Detection Algorithm Based on Attention-optimized RetinaNet

作　　者：吴宇震 胡朋 赵一帆 丁洪伟[1] 杨志军[1] WU Yuzhen;HU Peng;ZHAO Yifan;DING Hongwei;YANG Zhijun(School of Information Science&Engineering,Yunnan University,Kunming 650500,China;School of Electrical and Information Engineering,Yunnan Minzu University,Kunming 650500,China)

机构地区：[1]云南大学信息学院,云南昆明650500 [2]云南民族大学电气信息工程学院,云南昆明650500

出　　处：《无线电工程》2023年第7期1725-1733,共9页Radio Engineering

基　　金：国家自然科学基金(61461053)。

摘　　要：针对目标检测算法RetinaNet在多尺度物体检测任务中存在利用特征上下文信息和多尺度特征融合不充分及边界框回归不够快速精准的问题,提出了一种注意力优化RetinaNet的多尺度目标检测算法。在特征提取模块嵌入无参数的3D注意力机制,来充分利用特征上下文信息;同时,构建了特征融合细化模块,实现多尺度融合特征的细化和增强;使用距离交并比(Distance Intersection over Union,DIoU)损失函数优化定位损失,提升边界框回归精准度。为了论证该方法的有效性,分别在PASCAL VOC数据集和MS COCO数据集上进行实验。改进模型的检测精度分别达到了82.1%、52.3%,其中,小目标、中目标和大目标的检测精度相比原算法分别提升了1.9%、1.1%和1.4%。In the multi-scale object detection task,the object detection algorithm RetinaNet has the problems that fusion of features from contexts and multiple scales is not enough,and the bounding box regression is not fast and accurate enough.A multi-scale target detection algorithm based on attention-optimized RetinaNet is proposed.Firstly,a parameter-free 3D attention mechanism is embedded in the feature extraction module to make full use of feature context information;while a feature fusion refinement module is constructed to achieve the refinement and enhancement of multi-scale fusion features;finally,the Distance Intersection over Union(DIoU)loss function is used to optimize the location loss and improve the accuracy of bounding box regression.In order to demonstrate the effectiveness of the proposed method,experiments were carried out on PASCAL VOC dataset and MS COCO dataset respectively.The detection accuracy of the improved model reaches 82.1%and 52.3%,respectively.Compared with the original algorithm,the detection accuracy of small target,medium target and large target is improved by 1.9%,1.1%and 1.4%,respectively.

关 键 词：目标检测 特征融合细化模块 注意力机制 RetinaNet 距离交并比 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]