Fig 1.
Common electronic panel surface defects.
Fig 2.
A kind of supervised target detector applied to the detection of surface defects in electronic panels.
It includes Backbone, Neck, and Double T-head with features extracted by T-deformable convolution module, which will be described in detail in Chapter 3.
Fig 3.
A more detailed network structure.
Fig 4.
The deformable convolution module of the fusion convolution self-attention module.
The left side is the feature extraction branch, and the right side is the offset learning branch. The offset is learned by the CBAM module.
Fig 5.
The difference between Standard convolution, deformable convolution, and T-deformable convolution.
The left, middle, and right sides represent standard convolution, deformable convolution, and T-deformable convolution respectively. The upper, middle, and lower sides represent the activation unit on the feature map, 3 ×3 filter, sampling position.
Fig 6.
Double T-head structure, in addition to head1, head2 is also designed to enhance the detection of small target defects.
Fig 7.
SE module structure.
Fig 8.
Training results of our designed network.
Fig 9.
Test results of each trained network on the four categories of TNDSR1, TNDDR0, TNPLR0, TNNPE0.
Table 1.
Comparison of the proposed method with Faster- RCNN, Retinanet, SSD, YOLOV3, YOLOV5, and YOLOX on the same dataset for each metric.
Fig 10.
Training results of our network on the MSD dataset.
Table 2.
Comparison experiments between the proposed method and each comparison method on the MSD data set.
Table 3.
Map metrics and recall metrics for each network in the ablation experiment.