Stereo Matching And Image Segmentation Based On Deep Learning For Low-light Vision

Stereo matching network based on deep learning has developed rapidly in recent years and has been widely used in automatic driving,robot navigation,3D reconstruction and other fields.The existing stereo matching algorithms mainly study how to obtain a good disparity map under the ideal illumination environment.In the weak-light conditions such as cloudy days and nights,the matching accuracy of the current stereo matching algorithm based on deep learning is not so good.Therefore,this paper takes the low-light image as the research object,first studies the stereo matching algorithm for low-light images,and then discusses the image segmentation as the application direction.In this paper,a noise-resistant pyramid stereo matching network(NR-PSMNet)integrating imaging denoising and stereo matching is proposed for low-light images.The image reconstruction module is added to the classical stereo matching network to achieve low-light images denoising.Based on the idea of multi-task learning,combining stereo matching module with image reconstruction module can lead the two modules to interact with each other in the training process,and to work together to update the network parameters of feature extraction part,so that the image reconstruction module can assist low-light images to obtain more accurate disparity maps in the stereo matching module.The proposed method has been evaluated and achieves good performance on synthesized low-light images from public stereo datasets,KITTI2015 and Middlebury respectively,and also on real-world images captured with low-light cameras in dark environment.The low-light images show the characteristics of low contrast,low signal-to-noise ratio(SNR),complex and severe noise and the uneven illumination.When the images are segmented,the object contour cannot be accurately obtained,and getting good segmentation results.Thus,the low-light images and the disparity maps generated above are combined to introduce depth information into the RGB images and complement information,so as to obtain better segmentation results.In this paper,Attention Complementary Network(ACNet)is used to selectively extract image information layer by layer from low-light images,disparity maps and fusion features of each layer,then perform feature fusion to obtain higher quality features,and last,do upsampling to achieve image segmentation.The algorithm has beentested on the synthesized low-light images from public stereo datasets,NYUDv2,and also on real-world low images.Compared with the image segmentation that only uses the low-light image as the model input,a better segmentation effect is obtained.