基于语义采样和检测框优化的目标检测算法

Object Detection Based on Semantic Sampling and Localization Refinement

样本采样和检测框优化是目标检测任务中的两项重要技术。为了解决正负样本分配不合理的问题,获取更优的图像分类特征和检测框,提出一个精确且高效的单阶无锚框目标检测算法,算法由基于语义的定位、自适应特征增强和高效的检测框优化3个模块组成。首先,定位模块提出基于语义的样本采样方法,根据目标的语义特征区分前/背景区域,合理选择正样本和负样本,优先选择语义信息量较大的前景区域作为正样本;其次,特征增强模块利用目标语义概率图和检测框偏移逐像素调整图像分类特征,增大前景特征所占比重,根据目标大小自适应调整特征编码范围;最后,采用并联的方式优化检测框,对优化前后的检测框计算分类损失,几乎无成本地提升了定位性能,保证了特征对齐性和一致性。在MS COCO数据集下,提出的目标检测算法取得了平均精度为42.8%的检测精度,单张图像的检测时间达到78 ms,实现了检测精度与速度的平衡。

The two important techniques for object detection are training samplers and localization refinement.To solve the problem of unreasonable distribution of positive and negative samples,and get better image classification features and localizations,this study presented an accurate and effective single step anchorfree algorithm for object detection.The algorithm consists of three modules:semantic based positioning,adaptive feature enhancement,and efficient localization refinement.Firstly,the positioning module proposes a semantic based sampling method,which distinguishes the front/background regions according to the semantic characteristics of the object,reasonably selects positive samples and negative samples,and preferentially selects the foreground region with large amount of semantic information as the positive samples.Secondly,the feature enhancement module uses the target semantic probability map and detection frame offset to adjust the image classification features pixel by pixel,increases the proportion of foreground features,and adaptively adjusts the feature coding range according to the object size.Finally,the localizations are optimized in parallel,and the classification loss is calculated for the localizations before and after optimization,which improves the positioning performance almost without cost,and ensures the feature alignment and consistency.In the MS COCO dataset,the proposed algorithm achieves 42.8%in average precision,the detection time of a single image reaches 78 ms,realizing the balance between detection accuracy and speed.