基于Double-Head的雾天图像目标检测

Object detection in foggy image based on Double-Head

雾天环境下的图像对比度低,图像中的目标较为模糊并且其特征提取存在一定难度。现有的目标检测方法对于雾天图像的检测准确率偏低。针对上述问题,本文在Double-Head框架上基于图像的特征提取部分和预测头部进行改进。首先,在提取的深层特征图上添加通道和空间双维度的复合注意力机制,提高网络关注显著目标的能力;其次,将原始图像经过改进的暗通道先验以及处理后得到的先验矩阵和特征图进一步融合,获取更全面的雾天图像特征信息;最后,在预测头部引入可分离卷积,使用解耦合预测头对目标进行最终的分类和回归。该方法在RTTS数据集上的mAP为49.37%,在合成数据集S-KITTI和S-COCOval数据集上的AP值分别为66.7%和57.7%。与其他主流算法相比,本文算法具有更高的目标检测精度。

Image contrast in the foggy environment is low,and the object is fuzzy so that it is difficult to extract features in images.The existing object detection methods has a low accuracy for detecting objects in foggy images,and the objects is fuzzy and is difficult to extract features.To solve these problems,the feature extraction and prediction head are improved on the Double-Head framework.Firstly,multi-scale salient and effective features of objects in the image are carried out by adding channel attention to the feature maps extracted from the backbone network.Secondly,the prior matrix and fea-ture maps from the original image processing by dark channel prior method with image processing are fused to get more comprehensive feature information in foggy images.Finally,the separable convolution is introduced into the prediction head and the effective decoupled head is used to complete the classification and regression tasks.The proposed method has the mAP of 49.37%on the RTTS dataset,and the AP of 66.7%and 57.7%on the S-KITTI and S-COCOval dataset.Compared with other mainstream algorithms,this algorithm has higher object detection accuracy.