基于特征金字塔的多尺度特征融合网络

特征金字塔网络（FPN）是CNN网络对图像信息进行表达输出的一种有效方法，在目标检测网络中广泛应用．然而，FPN没有充分地将浅层的细节信息传递到深层的语义特征，存在特征融合不足的缺陷，因而只能依靠深层语义信息来进行预测，从而忽略了网络低层细节信息，对各种视觉学习的效果造成了一定的影响．针对FPN存在的以上问题，本文提出基于特征金字塔的多尺度特征融合网络模型，在FPN主干网络的基础上，设计了混合特征金字塔和金字塔融合模块，并结合注意力机制，对特征金字塔进行了多尺度的深度融合．本文在PASCAL VOC2012和MS COCO2014数据集上，以Faster R-CNN作为基础检测器进行实验，验证了MFPN对特征融合的有效性．

Muti-scale Feature Fusion Network Based on Feature Pyramid Model

Feature pyramid network (FPN) is an enhanced method for CNN network to express and output image information. It has been widely used in object detection network and has achieved significant effect improvement. The traditional feature pyramid model can not fully transfer the shallow details to the deep semantic features, which leads to inadequate feature fusion. It can only rely on the deep semantic information to make predictions, but ignore the underlying location information of the network. In terms of the above problems, we proposed a muti-scale feature fusion network based on feature pyramid model. Based on the FPN backbone, a mixed feature pyramid and a pyramid fusion module are designed. Based on the attention mechanism, multi-scale deep fusion of the feature pyramid is performed. We carry out the experiments on the PASCAL VOC2012 and MS COCO2014 datasets, and verify the effectiveness of MSFFN for feature fusion.