融合双重注意力机制的目标检测模型研究

Research on Object Detection Model Integrating Dual Attention Mechanism

针对Single Shot MultiBox Detector(SSD)算法未充分利用不同特征层之间的语义关系以及获取语义信息和位置信息能力不够好的问题,提出了一种Dual Attention Single Shot MultiBox Detector(DA-SSD)改进算法。为了建立浅层特征层与中间层之间的语义关系,采用并行残差多尺度特征提取网络,从而增强浅层特征层的语义信息和中间层的上下文信息。为了提高各特征层对语义信息和空间位置信息的获取能力,使用双重注意力机制加强对关键信息的学习。经实验验证,DA-SSD算法在PASCAL VOC2007测试集的检测精度相较于原始SSD算法提高了1.8%,有利于对自然场景的目标进行准确检测。

For the problem that Single Shot MultiBox Detector(SSD)algorithm does not fully utilize semantic relationship between different feature layers and the ability to obtain semantic information and position information is not good enough,an improved Dual Attention Single Shot MultiBox Detector(DA-SSD)algorithm is proposed.In order to establish semantic relationship between the shallow feature layer and the intermediate layer,a parallel residual multi-scale feature extraction network is used to enhance semantic information of the shallow feature layer and contextual information of the intermediate layer.In order to improve the ability of each feature layer to acquire semantic information and spatial position information,a dual attention mechanism is adopted to strengthen the learning of key information.After experiments,the detection accuracy of DA-SSD algorithm in PASCAL VOC2007 test set is improved by 1.8%in comparison with original SSD algorithm,which is conducive to accurate detection of objects in natural scenes.