Institute of Microelectronics, Chinese Academy of Sciences, No. 3 Beitucheng West Road, Chaoyang District, Beijing 100029, China.
University of Chinese Academy of Sciences, No. 19(A) Yuquan Road, Shijingshan District, Beijing 100049, China.
Sensors (Basel). 2021 Oct 27;21(21):7136. doi: 10.3390/s21217136.
Feature Pyramid Network (FPN) is used as the neck of current popular object detection networks. Research has shown that the structure of FPN has some defects. In addition to the loss of information caused by the reduction of the channel number, the features scale of different levels are also different, and the corresponding information at different abstract levels are also different, resulting in a semantic gap between each level. We call the semantic gap level imbalance. Correlation convolution is a way to alleviate the imbalance between adjacent layers; however, how to alleviate imbalance between all levels is another problem. In this article, we propose a new simple but effective network structure called Scale-Equalizing Feature Pyramid Network (SEFPN), which generates multiple features of different scales by iteratively fusing the features of each level. SEFPN improves the overall performance of the network by balancing the semantic representation of each layer of features. The experimental results on the MS-COCO2017 dataset show that the integration of SEFPN as a standalone module into the one-stage network can further improve the performance of the detector, by ∼1AP, and improve the detection performance of Faster R-CNN, a typical two-stage network, especially for large object detection APL∼2AP.
特征金字塔网络（FPN）被用作当前流行的目标检测网络的颈部。研究表明，FPN 的结构存在一些缺陷。除了由于通道数量减少而导致的信息丢失之外，不同层次的特征尺度也不同，并且不同抽象层次的相应信息也不同，从而导致每个层次之间存在语义差距。我们称这种语义差距为层次不平衡。相关卷积是缓解相邻层之间不平衡的一种方法；然而，如何缓解所有层次之间的不平衡又是另一个问题。在本文中，我们提出了一种新的简单而有效的网络结构，称为尺度均衡特征金字塔网络（SEFPN），它通过迭代融合每个层次的特征来生成多个不同尺度的特征。SEFPN 通过平衡每个特征层的语义表示来提高网络的整体性能。在 MS-COCO2017 数据集上的实验结果表明，将 SEFPN 作为独立模块集成到单阶段网络中，可以进一步提高探测器的性能，约为 1AP，并且提高了典型的两阶段网络 Faster R-CNN 的检测性能，特别是对于大目标检测 APL∼2AP。
