李宇琦, 赵海涛. 基于红外和可见光图像逐级自适应融合的场景深度估计[J]. 应用光学, 2020, 41(1): 24-32. DOI: 10.5768/JAO202041.0101004
引用本文: 李宇琦, 赵海涛. 基于红外和可见光图像逐级自适应融合的场景深度估计[J]. 应用光学, 2020, 41(1): 24-32. DOI: 10.5768/JAO202041.0101004
LI Yuqi, ZHAO Haitao. Depth estimation based on adaptive fusion of infrared and visible light images progressively[J]. Journal of Applied Optics, 2020, 41(1): 24-32. DOI: 10.5768/JAO202041.0101004
Citation: LI Yuqi, ZHAO Haitao. Depth estimation based on adaptive fusion of infrared and visible light images progressively[J]. Journal of Applied Optics, 2020, 41(1): 24-32. DOI: 10.5768/JAO202041.0101004

基于红外和可见光图像逐级自适应融合的场景深度估计

Depth estimation based on adaptive fusion of infrared and visible light images progressively

  • 摘要: 从图像中恢复场景的深度是计算机视觉领域中的一个关键问题。考虑到单一类型图像在深度估计中受场景不同光照的限制,提出了基于红外和可见光图像逐级自适应融合的场景深度估计方法(PF-CNN)。该方法包括双流滤波器部分耦合网络、自适应多模态特征融合网络以及自适应逐级特征融合网络。在双流卷积中红外和可见光图像的滤波器部分耦合使两者特征得到增强;自适应多模态特征融合网络学习红外和可见光图像的残差特征并将两者自适应加权融合,充分利用两者的互补信息;逐级特征融合网络学习多层融合特征的结合,充分利用不同卷积层的不同特征。实验结果表明:PF-CNN在测试集上获得了较好的效果,将阈值指标提高了5%,明显优于其他方法。

     

    Abstract: Recovering the depth of scenes from images is a key issue in the field of computer vision. Considering that the single type images were limited by different illumination of scenes in depth estimation, a method of depth estimation based on the progressively fusion convolution neural network of infrared and visible light images (PF-CNN) was proposed. This method includes the two-stream filter partially coupled network (TFCNet), the adaptive multi-model feature fusion network (AMFNet) and adaptive progressively feature fusion network (APFNet). The filters of infrared and visible light images are partially coupled in the TFCNet to enhance the features of them. The AMFNet learns the residual features of infrared and visible light images and fused them adaptively to fully utilize the complementary information. The APFNet learns the combination of multi-layer fusion features to make full use of the different characteristics of different convolutional layers. The experimental results show that the proposed PF-CNN obtains best performances on the data set and increases the threshold accuracy by 5%, which is better than other methods significantly.

     

/

返回文章
返回