【摘要】 针对现有图像能见度检测算法准确率低的问题。提出一种基于通道差分先验的能见度检测算法。首先利用统一全景分割网络(UPSNet)分割图像，提取感兴趣区域(ROI)并计算其最小外接矩阵，其次，采用颜色恒常先验法计算最小外接矩阵的大气光值，使用通道差分先验法计算其透射率，并通过加权插值和引导滤波相结合的方法优化透射率。最后，计算被测图像和参考图像感兴趣区域的透射率比值，通过非线性拟合模块反演能见度。实验结果表明，所提算法的判定系数(R~2)提高0.064～0.287,平均相对误差(MRE)较低，并且在曝光图像上具有良好的鲁棒性。更多还原

【Abstract】 Aiming at the problem of low accuracy of existing image-based visibility detection algorithms, a visibility detection algorithm based on channel difference prior is proposed. Firstly, a unified panoptic segmentation network(UPSNet)is used to segment images, the ROI and the minimum circumscribed matrix of image are calculated. Secondly, the atmospheric light of minimum circumscribed matrix is calculated by adopting color constancy prior, the channel difference prior is used to calculated the transmittance, then the rough transmission map is refined through weighted interpolation and guided filtering. Finally, the transmittance ratio of the ROI of the measured image and reference image is calculated, and the visibility is calculated by using the nonlinear fitting module. The results show that the coefficient of determination(R~2)is increased by 0.064～0.287,the mean relative error(MRE)is low, and the proposed method has good robustness on exposure image dataset.更多还原