| Image stitching is the conversion and stitching of two or more images into a large field of view panoramic image,wihch is widely used in various fields,such as image monitoring in industrial production,intelligent manufacturing field,stereo analysis of organ images in medical field,and generation of wide field of view remote sensing maps in military field.However,there are still many technical difficulties in the image stitching task.The traditional single-strain matrix estimation method,which relies heavily on manual extraction of feature points,is not sufficient for extracting feature point information of weak texture and low-light images,leading to artifacts in the stitching of such images;when convolutional neural networks are used for image stitching,the convolutional operation field is small and lacks global information,and there is a stitching misalignment problem in the large parallax stitching.In this paper,we introduce CA attention mechanism to solve the problem of image stitching.In this paper,CA attention mechanism is introduced to solve the problem of inadequate extraction of image feature point information,and non-local mechanism is introduced to solve the problem of stitching artifacts caused by small perceptual field.The main research content and innovation points are:1.A homography estimation network algorithm based on deep learning is proposed.Existing algorithms use feature pyramids for singleresponse matrix estimation,and although the features at the top and bottom layers are combined and utilized,the feature information extraction at the same layer is not sufficient.By introducing the CA attention mechanism to encode the precise location information with the channel relationship and long-term dependency,the attention of the network can be enhanced and the image feature representation can be improved.2.An image fusion network algorithm based on super-resolution reconstruction is proposed.After the images to be stitched are deformed by the above single-strain matrix estimation network,they are to be finally stitched by the image fusion network.To address the problem of small perceptual field of convolution operation in the network,a non-local module is introduced to increase the perceptual field of the convolutional neural network to obtain more useful information in the image to be stitched,and to improve the stitching misalignment problem caused by excessive parallax.In order to better learn the stitching process by seam masking,a new seam line loss function is designed,combined with the Mix loss function,which is a combination of MS-SSIM and L1 function,and can better maintain the contrast in high frequency regions.And preserves the color and luminance and local structure.3.Complete the design of image stitching module,and use back-end development technology to encapsulate and integrate the algorithm.At the same time,design API for web call,as a micro-service module of industrial Internet intelligent detection system,deployed on the server,to verify the feasibility of the algorithm. |
