面对雾天天气下,目标检测网络精度不高,鲁棒性差等问题,本文提出了一种结合多尺度特征融合去雾算法的可变形大核卷积雾天目标检测网络.雾天目标网络基于YOLOv8改进,骨干网络添加特征融合去雾模块(Feature Fusion Dehazing Module,FFDM),颈部网络引入可变形大核卷积注意力构造新的特征融合网络可变大卷积核特征金字塔(Deformable Large Convolution Feature Pyramid,DL-FPN),有效去雾的同时增强了目标检测网络对不规则目标和远小目标的网络检测能力.为了验证实验结果的有效性,本文在真实雾天数据集上进行了一系列实验,实验结果表明,与其他方法相比,本文的方法可以获得更高的检测精度.
DLK-YOLO:Deformable Large Kernel Convolution Enhances Foggy Target Detection
In the face of issues such as low accuracy and poor robustness of target detection networks under foggy weather conditions,this paper presents a deformable large-kernel convolution foggy target detection network that combines a multi-scale feature fusion dehazing algorithm.The foggy target network is an improvement based on YOLOv8,with a feature fusion dehazing module(Feature Fusion Dehazing Module,FFDM)added to the backbone,and a deformable large-kernel convolution attention introduced in the neck network to construct a new feature fusion network,namely the deformable large convolution kernel feature pyramid(Deformable Large Convolution Feature Pyramid,DL-FPN).This effectively dehazes while enhancing the network's detection ability for irregular and distant small targets.To validate the validity of the experimental results,a series of experiments were carried out on real foggy datasets in this paper.The experimental results indicate that,compared with other methods,the proposed method can achieve higher detection accuracy.
deformable large kernel convolutionmulti-scale feature fusion dehazingfog detectionattention mechanism
万方数据
维普
