Conditional TransGAN-Based Data Augmentation for PCB Electronic Component Inspection

Table 3

Experimental results (mAP) on the DeepPCB dataset.

Models	Settings	Open	Short	Mousebite	Spur	Copper	Pin-hole	mAP
Faster R-CNN ResNet101	w/o augmentation	94.8	95.7	98.5	98.8	98.9	99.5	97.7
	W/IPAug	95.6	96.3	98.2	98.7	98.8	99.4	97.8
	w/TransGAN	95.4	96.1	97.9	98.4	99.1	99.6	97.8
	w/CGAN	96.2	97.6	98.2	98.3	98.9	99.5	98.1
	w/cTransGAN	98.1	98.3	99.2	98.9	99.1	99.2	98.8
YOLOV3 DarkNet	w/o augmentation	91.2	92.4	94.8	92.3	96.4	93.2	93.4
	W/IPAug	92.4	93.2	94.6	92.5	96.8	93.7	93.9
	w/TransGAN	92.8	93.6	94.9	92.4	97.1	94.5	94.2
	w/cGAN	93.5	93.5	94.7	92.6	97.8	94.1	94.4
	w/cTransGAN	95.2	95.6	96.1	94.3	98.1	97.2	96.1
SCNet ResNet101	w/o augmentation	93.6	93.8	96.2	96.9	97.5	99.2	96.2
	W/IPAug	94.5	95.2	98.1	97.3	97.9	99.1	97.1
	w/TransGAN	94.7	95.9	97.6	96.9	97.9	99.2	97.0
	w/CGAN	95.1	96.4	98.4	97.2	97.8	99.3	97.4
	w/cTransGAN	97.5	96.9	98.9	98.1	98.4	99.3	98.2
SOTA [33]	w/max pooling	98.5	98.5	99.1	98.2	98.5	99.4	98.7

Table 3 gives the metrics for the best models trained by the Faster R-CNN ResNet101, YOLOV3 DarkNet, and SCNet ResNet101 models using different enhancements on the two dataset tasks, respectively, where the metrics include the AP value and the mean AP value (mAP) for each subtarget. The range of AP values is from 0% to 100%, with higher values demonstrating better detection of the target. As can be seen in Tables 3, the detection results using cTransGAN are almost always optimal.