A FINE-GRAINED OBJECT DETECTION MODEL FOR AERIAL IMAGES BASED ON YOLOV5 DEEP NEURAL NETWORK

Abstract

This research aims to solve the common task of highly precise object detection in remote sensing images,
which existing approaches that were developed for natural scenes are not suitable for. In this process we are using
Circular Smooth Label (CSL), which does angle regression into a class-type, reducing losses due to angle
periodicity. YOLOv5 will be taken as the original model for which we add the CSL and an attention mechanism
module to strive for higher detection accuracy for small objects with arbitrary orientations. On the FAIR1M dataset,
our YOLOv5-CSL resulting in an average 0.72 mAP. Besides that, examining the types such as YOLOv5x6 yields a
measuring points, with the mAP increasing by 0.80% or above. This note allows to assert that a remote sensing
object detection enhancement certainly belongs to a kind of the research topics in the future.