Algorithm Design And Development Of Computer Vision-based Aircraft Maintenance Inspection System

In modern society,airplanes have been widely accepted as a convenient and high-speed means of transportation,and the safety of airline flights has always been a focus of attention.Pre-flight safety checks for airplanes,as the last safety measure before takeoff,are generally visually inspected by aviation maintenance personnel according to prescribed procedures on important parts of the plane.This work not only requires a lot of manpower and material resources but also can result in misjudgments and omissions due to the negligence of maintenance personnel.With the rapid development of computer vision technology,this problem can be solved in a more intelligent way.This article proposes an aircraft safety maintenance inspection system deployed on an aircraft inspection robot,which can conduct safety inspections on various flight components of the airplane through images taken during the inspection process by the inspection robot,and locate and record faults and potential hazards.This article focuses on object detection tasks,improves and optimizes the YOLOv7-tiny-based object detection algorithm to reduce the system’s use of computing resources and improve recognition accuracy,and covers several main research areas.The first part mainly introduces the research background of computer vision,summarizes the development history of computer vision,the current research status of object detection algorithms and the research status of neural network model lightweighting.It introduces the basic working principles of mainstream single-stage object detection algorithms and model lightweighting,and compares and analyzes classical object detection models.Finally,YOLOv7-tiny,a lightweight model more suitable for deployment on inspection robots,is selected as the algorithm model.The second part proposes an improvement to the YOLOv7-tiny algorithm,combining GhostNet network and depthwise separable convolution to create a lightweight YOLO_GDW model that reduces parameter and computational complexity while maintaining the overall framework of the original algorithm.This makes it more suitable for deployment on edge devices like limited computing power inspection robots to perform object detection tasks despite varying target sizes and detection difficulties.Meanwhile,Coordinate Attention and SPD structures are used to optimize the YOLO_GDW model and improve its backbone network’s feature extraction capability,thus enhancing detection accuracy.Experimental results show that the improved SCA_YOLO_GDW model achieves 98.0%accuracy on the pre-flight safety maintenance inspection dataset,a 1.5%increase compared to the YOLOv7-tiny model.In terms of parameters,the YOLOv7-tiny model has 1.45 times more parameters than the SCA YOLO GDW model,and in terms of computation load,the YOLOv7-tiny model is 1.43 times heavier than the SCA_YOLO_GDW model,meeting the requirement for model lightweighting.The third part of the paper designs a pre-flight safety maintenance inspection system based on the improved SCA_YOLO_GDW algorithm for airplane inspection robots.The system includes image acquisition,object detection,and network modules,and its functions and performance are tested.Experimental results show that all modules in the system operate correctly,and after improvement and optimization of the object detection module algorithm,the system’s efficiency significantly improves.The system’s single-frame operation time is about 21.42ms,and it can test 46 to 47 frames per second,meeting real-time requirements.