作者:李成严，赵帅，车子轩

Authors:LI Cheng-yan,ZHAO Shuai,CHE Zi-xuan
摘要:摘要:为了提升SSD(single shot multiBox detector)的检测精度，提出一种基于Anchor-object匹配的A-SSD(anchor-object SSD)目标检测算法。在算法的特征提取部分，使用并行卷积和空洞卷积构成感受野模块，增大特征图的感受野，获得多尺度的特征信息；将含有纹理、边缘等细节信息的浅层特征与含有丰富语义信息的深层特征融合。在算法的检测器部分，采用Anchor-object匹配方法联合SSD多层特征图为每个检测目标构建相应的Anchor包，通过选择-抑制优化策略，选择置信度较高的Anchor去更新模型，为每个Anchor评估得分。通过迭代学习，模型的参数、Anchor的位置坐标与分类置信度不断优化。在PASCAL VOC数据集上A-SSD算法的mAP达到了80.7，在车间行人数据集上A-SSD算法的漏检率为3.5%，准确率为91.5%。
Abstract:Abstract:In order to improve the detection accuracy of SSD (Single Shot MultiBox Detector), an A-SSD (Anchor-object SSD) object detection algorithm based on Anchor-object matching is proposed. In the feature extraction part of the algorithm, parallel convolution and hole convolution are used to form the receptive field module, which increases the receptive field of the feature map and obtains multi-scale feature information, combining the shallow features containing texture, edge and other detailed information with rich fusion of deep features of semantic information. In the detector part of the algorithm, the Anchor-object matching method combined with the SSD multi-layer feature map is used to construct the corresponding Anchor package for each detection target. Through the selection-suppression optimization strategy, the Anchor with higher confidence is selected to update the model. Anchor evaluation score, in continuous iterative learning, the model parameters, Anchor position coordinates and classification confidence are continuously optimized. The mAP of the A-SSD algorithm on the PASCAL VOC data set reached 80.7, and the missed detection rate of the A-SSD algorithm on the workshop pedestrian data set was 3.5%, and the accuracy rate was 91.5%.

PDF全文下载地址:

可免费Download/下载PDF全文