Fig 1.
Samples of the rail defect dataset with 8 categories.
From left to right, the first row contains saplling, dirt, unknown, and gap, and the second row contains squat, crush, scratch, and crack.
Fig 2.
Structure of the proposed MBDA framework.
The input image is a detection sample of YOLO v5.
Fig 3.
Image augmentation methods within IAB.
Fig 4.
Feature augmentation methods within FAB.
Table 1.
Detection results of different methods over the rail defect datasets.
Fig 5.
Detection performances and validation losses.
The first row from left to right are YOLOv5s, YOLOv5s6, and YOLOv5m. The second row from left to right are FasterRCNN R50, FasterRCNN R101, and MBDA.
Fig 6.
From left to right are ground truth, MBDA, FasterRCNN, and YOLO.
Table 2.
Different MBDA structures.
Fig 7.
Number of backbones ensembled and their detection precisions.
Table 3.
Effectiveness of image augmentation operations.
N/A refers to taking no image augmentation method. OS and BS are short for object swap and layer swap, others refer to randomly selected image augmentation methods other than the two newly designed operations.
Table 4.
Effectiveness of feature augmentation operations.
LS, CS, SC are short for layer swap, channel swap, spot cover, respectively.
Table 5.
The impact of the combination of image augmentation and feature augmentation operations.