Earf-YOLO: An Efficient Attention Receptive Field Model for Recognizing Symbols of Zhuang Minority Patterns

Table 1

Comparison of AP of Earf-YOLO with that of other latest models on the datasets of Zhuang pattern symbols.

Methods	The backbone network	AP(%)	AP50(%)	AP75(%)	APS(%)	APM(%)	APL(%)
YOLOv4 640(baseline) [11]	ResNet-101	36.8	56.2	39.2	20.4	39.8	46.4
YOLOv5s8	ResNet-101	37.7	56.8	40.4	21.1	40.8	47.6
Libra RetinaNet [39]	ResNet-101	36.5	55.7	39.1	21.0	40.6	46.3
RetinaNet w/AugFPN [40]	ResNet-101	37.2	55.4	40.2	20.2	40.3	47.1
Our model	ResNet-101	39.1	58.5	41.7	22.3	42.3	48.9
RetinaNet w/SABL [41]	CSPDarkNet53	36.3	59.2	39.1	18.3	39.1	48.3
LRF [42]	CSPDarkNet53	38.3	60.4	41.8	20.2	41.2	50.3
Faster RCNN [43]	CSPDarkNet53	37.4	58.6	39.8	19.8	41.9	50.2
YOLOv4 640 [11]	CSPDarkNet53	38.2	58.6	40.9	21.3	42.6	48.3
RDSNet [44]	CSPDarkNet53	40.6	59.4	41.9	21.1	41.7	50.4
YOLOX [45]	CSPDarkNet53	40.7	59.9	42.5	22.5	43.2	51.6
Our model	CSPDarkNet53	41.0	61.7	43.8	24.3	44.4	51.8