BAFPN: An Optimization for YOLO

Abstract

Object detection is essential in Computer Vision and is widely applied in all areas. This paper proposes a method called BAFPN. BAFPN is a new bidirectional Feature Pyramid Network that constructs accurate object detection networks based on YOLOv4 by implementing Adaptively Spatial Feature Fusion. Besides, Exponential Moving Average is used to improve the network performance. The developed network not only maintains high computing speed but also enhances the mAP by 4.3% when testing with the MS COCO dataset and when comparing it to the original YOLOv4. To further improve the performance, the trained model was pruned using the Batch Normalization layer's scaling factor, achieving an 18% reduction in size and improving object detection speed.

YOLO based deep learning on needle-type dashboard recognition for autopilot maneuvering system

“…It was used to describe the relationship between the power of signal and background noise. Its definition can be written as equation (3).…”

“…In Li et al, 3 the BAFPN improved the detection ability by replacing the Path Aggregation Network (PANet) 4 structure in the original You Only Look Once version 4 (YOLOv4). This has increased both the calculation speed and accuracy, and it showed a great performance in both mAP and FPS by the Microsoft Common Objects in Context (MS COCO) dataset.…”