INSTANCE SEGMENTATION ALGORITHM FOR DETECTING AND CLASSIFYING DEFECTS IN AIRCRAFT FUSELAGE SKINS WITH CONVOLUTIONAL NEURAL NETWORKS

The use of computer vision in visual inspection automated to identify noncon- formities in the Manufacturing and Maintenance of products has become well established in various industrial sectors. When employing Deep Learning techniques with Computer Vision, these sectors have achieved remarkable results, continually improving quality and reducing inspection times compared to traditional manual methods. However, some industries, such as Aeronautics faces challenges in implementing these technologies due to the critical nature of their final products. This study addressed the specific challenges of applying such technologies for product inspections in the Aerospace Industry. To this end, experiments were developed using state-of-the-art algorithms with convolutional neural networks to create instance segmentation models to detect and classify nonconformities in aircraft fuselage, with the aim of understanding and comparing the results obtained with the application to a set of data created for the research, containing cracks, dents, scratches and dirt on aircraft fuselages. The analyzes showed good results for the task that analyzed the performance of models created with YOLOv8 and YOLOv9, identifying a better result from the latter and evidence of points of common difficulties such as the detection of small defects.