Research And Implementation Of Image Stitching Algorithm Based On Deep Learning

In recent years,image stitching technique has become a research hotspot in the field of computer vision.In the face of images with weak or no texture,such as forest,sky,sea and other scenes,the traditional image stitching technology will lead to poor stitching effect due to the lack of feature detection ability.In the field of deep learning,the supervised method is difficult to be popularized due to the high labor cost and time overhead of data set labeling.However,due to the parallax problem between images,the stitching result of unsupervised method is often not ideal in the face of large parallax images.In order to stitching forest images better,this paper proposes a new,unsupervised deep learning image stitching algorithm,which has better image stitching effect in forest images with fewer features and different parallax problems.The algorithm is mainly divided into two parts:transformation and image stitching.(a)In the transformation network based on transformer,based on the powerful feature extraction capability of CNN,the transformation method is adopted.After obtaining the transformation matrix between images,the feature mapping is transformed through the monoresponse matrix using transformer.Thus,the image to be stitched with large parallax is converted into the image to be stitched with small parallax,which not only ensures the integrity of the image content,but also reduces the parallax problem between the images to be stitched.(b)In the image stitching network based on SE attention,the VGG-16 network was used as the basis to stitch the images with small parallax first,and then the CNN and SE at tention were combined to solve the problems of color inconsistency and color distortion after initial stitching.Through experimental,in terms of subjective vision,the algorithm proposed in this paper has a natural stitching,fewer stitching artifacts and a low color distortion rate in various forest scene stitching tasks.In terms of quantitative analysis,the PSNR of the image stitching algorithm is improved by 1.991 d B compared with SIFT algorithm,11.7921 d B compared with SFE algorithm,and 0.803 d B compared with DHN algorithm.And the SSIM value of the image stitching algorithm is improved by 7.37% compared with SIFT algorithm,8.68% compared with SFE algorithm and 2.32% compared with DHN algorithm.