Innovative research on multispectral image fusion technology for defect detection of composite insulators in ultra-high voltage transmission lines

Figure 1.

Figure 2.

Figure 3.

Figure 4.

Comparison of matching performance on data set

Algorithm	Evaluation index	Training set	Test set
SIFT	Average registration point pairs	412	107
	err	0.54	0.71
	recall	91%	87%
PIIFD	Average registration point pairs	98	13
	err	0.67	0.74
	recall	82%	77%
MatchNet	Average registration point pairs	57	0
	err	0.94	—
	recall	62%	0%
RF-Net	Average registration point pairs	69	0%
	err	0.91	—
	recall	67%	0%
IORB	Average registration point pairs	546	387
	err	0.32	0.38
	recall	100%	100%

Comparison of algorithm Metrics on the test set(Scale transformation)

Algorithm	Evaluation index	The index of the algorithm in different scaling multiple
		1	0.8	0.6	0.4
ORB	Registration point pairs	242	103	17	0
	err	0.68	0.71	0.77	—
	recall	91%	88%	82%	0%
IORB	Registration point pairs	387	358	199	163
	err	0.34	0.39	0.47	0.51
	recall	100%	100%	100%	100%

Comparison of algorithm on training set(Rotational scale transformation)

Algorithm	Evaluation index	Index
ORB	Registration point pairs	0
	err	—
	recall	0%
IORB	Registration point pairs	59
	err	0.47
	recall	100%

Comparison of algorithm Metrics on the test set(Rotation transformation)

Algorithm	Evaluation index	The index of the algorithm in different rotation angles
		0°	90°	180°	270°
ORB	Registration point pairs	134	128	13	4
	err	0.89	0.92	0.74	0.83
	recall	74%	71%	83%	72%
IORB	Registration point pairs	487	462	451	436
	err	0.34	0.37	0.39	0.42
	recall	100%	100%	100%	100%

Comparison of algorithm Metrics on the training set(Scale transformation)

Algorithm	Evaluation index	The index of the algorithm in different scaling multiple
		1	0.8	0.6	0.4
ORB	Registration point pairs	362	117	23	3
	err	0.59	0.63	0.71	0.84
	recall	98%	93%	87%	74%
IORB	Registration point pairs	392	363	204	172
	err	0.31	0.37	0.42	0.46
	recall	100%	100%	100%	100%

Results statistics of defect identification channels

Channel number	1	2	3	4	5
Normal	0.892	0.934	0.831	0.726	0.648
Spontaneous detonation	0.873	0.952	0.817	0.708	0.625
Crack	0.869	0.948	0.803	0.714	0.639
Filth	0.882	0.955	0.813	0.734	0.628

Comparison of algorithm Metrics on the training set(Rotation transformation)

Algorithm	Evaluation index	The index of the algorithm in different rotation angles
		0°	90°	180°	270°
ORB	Registration point pairs	271	167	32	11
	err	0.84	0.81	0.86	0.73
	recall	82%	89%	91%	76%
IORB	Registration point pairs	541	632	489	607
	err	0.31	0.33	0.29	0.36
	recall	100%	100%	100%	100%