Research On Ground Image Stitching Technology Based On Multiple Cameras

Image stitching technology is a significant research area in the field of computer vision,which has practical applications in assessing the degree of ground changes before and after live-fire training,as well as evaluating the damage performance of weapon strikes in combat training.However,ground images captured on the range often lack distinctive features and textures,resulting in inaccurate evaluation results and high error rates.To address these limitations,this article proposes a ground image stitching technique based on multi-camera imaging,which enables the restoration of a wide range of ground images.The main research focus of this study is to:(1)To address the problem of over-reliance on image features in traditional image alignment,an image alignment algorithm without significant features in the local overlap region is proposed.The overlapping area of the camera field of view is pre-determined to determine the approximate location of the overlapping area between the images acquired by the camera,and the overlapping area is masked and segmented with the help of image masking and segmentation techniques.Experiments have shown that the method improves the alignment efficiency by 28% while reducing the mis-matching rate between images with insignificant features.(2)This thesis proposes an unsupervised stitching algorithm that utilizes VGG-19 as the backbone network to address issues of poor stitching quality in large range ground scenes with large parallax and few features.To achieve alignment,a deep single-response network is used to estimate the single-response between images,and the input image is distorted through the stitching domain transformer layer.A feature-to-pixel method is also employed to learn the deformation rules of image stitching and improve the resolution of the reconstructed image.Experimental results demonstrate that the proposed unsupervised image stitching algorithm outperforms traditional stitching approaches and is better suited for the research scenario presented in this paper.Moreover,the method effectively addresses common issues such as artifacts and distortions in large range ground scenes.(3)An unsupervised image stitching and fusion method is proposed in this study to address the problem of structural distortion and irregular boundaries that may arise in the overlapping regions of images when using an unsupervised stitching algorithm.Specifically,the feature extraction layer of the deep single-response network within the unsupervised alignment module is deepened from conv＿block4 to conv＿block5,and a convolutional layer is introduced into the high-resolution branch of the reconstructed network to prevent the loss of low-level information as the network depth increases,thereby mitigating the issue of distortion in the overlapping regions.Moreover,a one-stage image correction deep learning scheme is developed to rectify the distorted stitched images.The scheme entails creating a rigid target network for the input image,predicting the initial network via a convolutional network with progressive regression of residuals,forming a transformation model between the input image and the rigid target network,and utilizing a comprehensive objective function to enforce high fidelity content constraints on the corrected image.The experimental results confirm the efficacy of the optimized method,as it produces natural-looking images with appreciable improvements and high feasibility.(4)This thesis conducts experiments on four different groups of scenes from the Gobi,desert,crop field,and forest,and evaluates the stitched images using both objective and subjective evaluations.The results show that the improved multi-camera based large-range ground image stitching method adopted in this paper considers both the large field of view and high-resolution features,exhibiting excellent stitching performance,improved migration,and robustness.These findings are of significant research value for the restoration of large-range range ground scenes.