![](https://csdnimg.cn/release/download_crawler_static/89118428/bg4.jpg)
4 L. ANG, SKNA. RAHIM, R. HAMZAH, R. AMINUDDIN, AND G. YOUSHENG
Figure 2: The overall process flowchart.
2. RELATED WORK
2.1. Feature Pyramid Network. Feature Pyramid Network (FPN) [LDG
+
17] is a feature
fusion method commonly used for object detection, which is a network model for extracting
pyramid feature representations. It is usually used in the feature fusion stage of object
detection. After performing a bottom-up feature extraction operation on the backbone
network, the FPN is connected to the corresponding layer’s feature maps from top to bottom
and horizontally, sequentially combining the two adjacent layers in the backbone network’s
feature hierarchy to construct a feature pyramid. Although FPN is simple and effective,
some aspects still have shortcomings. Before the feature fusion at each layer, there is a
semantic gap between different layers, and direct fusion will have a negative impact on the
representation ability of multi-scale features. During feature fusion, the feature information
at the high-level of the pyramid network will be lost during the upsampling process.
Path Aggregation Network (PANet) [LQQ
+
18] structure has been improved based on
FPN and is extensively employed in the YOLO object detection frameworks and its variant
frameworks. This network has two feature fusion paths, namely top-down and bottom-up.
This approach reduces the fusion distance between deep and shallow features. optimize the