面向少样本网状结构体的候选区域自适应检测方法

Proposal adaptive detection method for small sample reticular structures

在仅有少量标签数据的图像网状结构检测任务中,需要大量训练数据的目标检测模型,检测性能大幅下降。基于区域候选的目标检测模型在预测时,检测目标越多,检测时间越长。若基于区域候选的目标检测模型产生候选框的数量固定不变,而不同图像中网状结构目标数量不同,造成目标检测中额外的时间消耗。针对该问题,通过对训练样本中网状结构目标在图片中的密度分析以及根据网状结构体在图片中的特征分布,提出一种面向少样本网状结构体的候选区域自适应检测方法。该方法通过基于二值标签图标注方法得到大量训练样本,由候选区域自适应方法选取合理的候选框数量。与未改进的模型相比,在几乎不损失准确率的情况下,其加快了检测速度,尤其在目标数量稀少的数据中优势更为明显。

The detection performance of the object detection model requiring a large amount of training data is greatly reduced in the reticular structures detection task with a small amount of labeled data. The detection model which based on region proposals will spend longer time with larger outputs during predicting. And the number of reticular structures are different in different images. It will result in extra time consumption in object detection if the number of proposals is fixed in different images.In view of this problem,this paper proposed a proposal adaptive detection method for small sample reticular structures by analyzing the density of the reticular structures in the training samples and according to the characteristic distribution of the reticular structures in the images. This method obtained a large number of training samples by the method of binary-value labeled map marking,and selected a reasonable number of proposals by the proposal adaptive method. Compared with the unimproved model,the detection speed of proposed method is accelerated without loss of accuracy,especially in the data with few objects.