基于优化预测定位的单阶段目标检测算法

针对单阶段多边框检测(SSD)算法中存在目标定位不准确和小目标检测精度不高的问题，提出基于优化预测定位的单阶段目标检测算法EL-SSD. 通过双向加权特征金字塔将原SSD预测特征图特征融合，对输出特征图进行特征位置信息解码后进行特征通道权重再分配，提升了特征语义信息，捕获了跨通道位置信息. 通过构建分类置信度及额外的定位置信度级联聚类对预测框进行非极大值抑制，提高在检测阶段对选择目标的定位精度. 实验结果表明，EL-SSD算法在PASCAL VOC2007上的平均检测均值达到79.8%，比原SSD算法提高了2.6%. 在COCO数据集上的精度达到29.4%，比原SSD算法提高了3.5%，在检测图片上的目标定位效果及小目标检测效果明显优于SSD, 适用于需要高定位性能的实时应用场景.

Single-stage object detection algorithm based on optimizing position prediction

A single-stage object detection algorithm named EL-SSD based on optimizing position prediction was proposed aiming at the problem of inaccurate target positioning and low accuracy of small object detection in the single shot multi-box detector (SSD) algorithm. The prediction feature maps from original SSD were decoded according to feature location information after feature fusion by bi-directional feature pyramid network. Then the weights of feature channels were redistributed. The feature semantic information was improved, and the cross-channel location information was captured. The classification confidence and additional fixed position confidence cascade clustering were constructed to non-maximum suppress prediction bounding box at the detection stage, which improved the positioning accuracy of the selected target. The experimental results showed that the mean average precision value of the EL-SSD on the PASCAL VOC2007 dataset was 79.8%, which was 2.6% higher than that of the original SSD algorithm. The mean average precision value of the EL-SSD on the COCO dataset was 29.4%, which was 3.5% higher than that of the original SSD algorithm. The improved SSD algorithm has a better location performance and small target detection performance, which is suitable for application scenarios that require high localization performance.