Experimental Design and Implementation of Image Stitching Based on Unsupervised Deep Learning
In the application fields of intelligent transportation,virtual reality and remote sensing monitoring,image stitching can present images obtained from multiple cameras.Aiming at the lack of teaching cases related to image stitching in electronic information courses such as digital image processing and pattern recognition,as well as the problems of artifacts and distortion in large parallax image stitching,an unsupervised deep learning image stitching method based on depth-color image fusion is proposed.The depth of field map corresponding to the color image is used to fuse images,the homography transformation is calculated to obtain the roughly aligned image,and the low resolution branch and high resolution branch are used to obtain the fine aligned images,and the channel attention and dilated convolution are used to improve the stitching effect.The effect of parallax image stitching is tested in different data sets and multiple scenes,and it is verified that the method can effectively complete large parallax image stitching and multiple pictures stitching in real scenes.
image stitchinghomography estimationunsupervised deep learningscene depth
万方数据
维普
