Fig 1.
(a) shows the original grouped convolution, (b) shows different channels being shuffled to different positions after grouped convolution, and (c) shows the result after channel shuffling is completed.
Fig 2.
SH-encoder module, DW stands for depthwise separable convolution.
Fig 3.
UPC-SimAM module, DW stands for depthwise separable convolution.
Fig 4.
(a) Channel-wise attention, (b) Spatial-wise attention, (c) Full 3-D weights for attention.
Fig 5.
SH-DETR architecture.
Table 1.
Experimental hardware and software configuration.
Fig 6.
Examples of each category in NEU-DET.
Fig 7.
Distribution of categories in the NEU-DET dataset.
Fig 8.
Examples of each category in GC10-DET.
Fig 9.
Distribution of categories in the GC10-DET dataset.
Table 2.
The detection performance of different categories on NEU-DET dataset.
Fig 10.
Loss function curve and detection precision experimented on NEU-DET dataset.
Fig 11.
Confusion matrix of the detection results on NEU-DET dataset.
Fig 12.
Loss function curve and detection precision experimented on GC10-DE dataset.
Table 3.
The detection performance for different dataset.
Table 4.
The detection performance for different categories on GC10-DE dataset.
Table 5.
The detection performance for different categories on welding defect dataset.
Fig 13.
Examples of images processed with different data augmentation methods.
Fig 14.
Loss curves and detection precision for processed images on the GC10-DE dataset.
Fig 15.
Histogram of model accuracy training results.
Table 6.
Comparison of our proposed model with other state-of-the-art models on NEU-DET dataset.
Table 7.
Comparison of detection methods by defect category on NEU-DET dataset.
Table 8.
Comparison of detection methods by defect category on GC10-DET dataset.
Table 9.
Impact of each module on the model in ablation studies.
Fig 16.
Visualization results of different modules in ablation studies.