A Study on Fog and Haze Weather Object Detection Method Based on DeblurGANv2 and YOLOv4

Abstract

To address the issue of low detection accuracy in object detection under foggy and hazy conditions, a novel defogging object detection method that integrates DeblurGANv2 and YOLOv4 is proposed. This approach incorporates an image enhancement algorithm from the DeblurGANv2 generative adversarial network into the preprocessing module of YOLOv4. This integration aims to preprocess foggy images, preserving high-quality texture and color information. Additionally, the method replaces the CSPDarkNet53 backbone network in YOLOv4 with the lightweight ShuffleNet V2 neural network to enhance model detection speed. Experimental results demonstrate that this proposed method effectively reduces issues of significant color discrepancy and fog residue, achieving a mean average precision (mAP) of 86.56% on the test dataset. These results indicate the method's efficacy in real-world defogging object detection scenarios.