Study On Bio-inspired Compound Eye Array Image Stitching And Object Detection Technology

In recent years,large field of view,high resolution image and image target detection are in great demand in earth observation,medical imaging,industrial manufacturing,aerospace and other fields.The compound eye array vision system imitates the parallel compound eyes of insects,and fixes the miniature camera in the designed structure,so that the images with repeated field of view can be collected by the adjacent cameras,and then spliced into a large field of view,high resolution image.Compared with microlens array bionic compound eye,camera array bionic compound eye has larger resolution,clearer image and more perfect technology.In order to enrich the application of bionic compound eye array system,the movable small deep learning platform can be used to detect the object in the large field of view and high resolution image.In this way,a small bionic compound eye array target recognition system can be obtained,which promotes the continuous improvement of image information acquisition and processing technology.This article presents the design and construction of a bionic compound eye array target recognition system,and proposes an improved fast image stitching algorithm,which achieves the splicing of large field of view,high-resolution images,and the fast detection of targets.The research content mainly includes the following parts:Firstly,system design and construction of the compound eye array vision system.As aerial target detection requires a large field of view and high resolution in aerial images,conventional unmanned aerial vehicle(UAV)cameras cannot meet the requirements of both large field of view and high resolution simultaneously.This paper proposes a method of using a multi-camera compound eye array for visual system construction.The system mainly consists of a high-resolution camera,a compound eye array module,and an image processing system.The compound eye array module arranges high-resolution cameras at calculated angles to obtain the largest imaging field.The image processing system performs image stitching and target recognition on images captured by multiple cameras,achieving target recognition of the compound eye array system.Secondly,this thesis proposes a fast and efficient method for stitching multiple images with a large field of view based on the improved ORB-GMS-SPHP(Oriented FAST and Rotated BRIEF＿Grid-based Motion Statistics Shape-preserving half-projective)algorithm.This method has the advantages of fast feature point matching,high accuracy,and the ability to retain more correct matching points.In the 2736 pixels x 3648 pixels image,the feature point matching time is reduced to 6.463 s,and the image registration precision(RMSE)is reduced to 3.87.Experimental results show that this method has fast and accurate feature point matching,high stitching accuracy,and no significant chromatic aberration.Then,This paper proposes an improved YOLOv7 object detection model.In order to reduce the feature point loss caused by the network model processing,the MPConv module in the YOLOv7 model is subjected to feature separation processing,and experiments are designed to determine the optimal position for the improved MPConv module.In order to reduce the occurrence of missed detections,this paper uses the SIo U function instead of the CIo U function in YOLOv7,which reduces the degree of freedom of the loss function and improves the robustness of the original object detection network.In response to these improvements,this paper conducted comparative experiments on the public FLo W-Img subdataset.The results show that for images of large scales and dense targets in the dataset,the algorithm improved in this paper significantly reduced the missed detection rate compared to the original YOLOv7 algorithm,with m AP reaching 70.9%,which is 3.5% higher than the original model.Finally,this paper designs and constructs a compound eye array target recognition system,proposes an improved ultra-fast image stitching algorithm,and completes the stitching of multiple overlapping images within the system.Finally,an improved YOLOv7 algorithm is proposed,which improves recognition accuracy while ensuring target recognition speed.