We studied the data annotation process of wind turbine blade images to reduce inaccuracies in defect detection,and benchmark the performance of patch based detection frameworks on the latest deep learning architecture.In the study,the biggest difficulty in the detection task was identified,which was caused by the extreme aspect ratio of bounding boxes in annotations.Experiments on two additional annotation sets showed that sets with changed cuboid aspect ratios can improve overall defect detection accuracy,especially for classes containing cuboids with very small aspect ratios.A large number of class results were also provided through visual examples highlighting the problem.
维普
万方数据
