基于密集连接的FPN多尺度目标检测算法

图像中目标对象的多尺度问题一直以来都是目标检测领域的主要难点之一,尤其是极端尺度对象的检测。研究发现,目标检测网络模型的深层语义特征有利于对象的识别,而浅层空间特征对对象的边界框回归很有帮助。DC-FPN使用密集连接代替FPN网络模型中的横向连接,能够从多层特征层中获取目标检测所需的特征信息,其中密集连接可以融合FPN自底向上传输模块中的所有特征层的特征信息,使FPN自顶向下传输模块的预测特征层能从中获取不同尺度对象检测所需的特征信息。实验表明,DC-FPN目标检测算法能够显著地提升模型的多尺度目标检测性能,使用MS COCO数据集训练和测试,其检测准确度(AP)能达到43.1%。

FPN MULTI-SCALE OBJECT DETECTION ALGORITHM BASED ON DENSE CONNECTIVITY

The multi-scale of objects in images has always been one of the main challenges in the object detection field,especially for the extreme-scale object detection.The study shows that the semantic features of the deeper layers in the object detection network are beneficial to the object recognition,while the spatial features of the shallow layers are helpful in the bounding box regression of the object.The DC-FPN which used a dense connection instead of the lateral connection in FPN could extract feature information required for object detection from multiple layers.The dense connection could fuse the feature information of all layers in the bottom-up path of FPN so that the feature layers in the top-down path of FPN could obtain the feature information required for the detection of different scale objects.Experimental results show that the multi-scale object detection performance can be improved significantly by the DC-FPN detection algorithm.The DC-FPN is further investigated on MS COCO detection set,and the detection accuracy(AP)can reach 43.1%.