Deep Learning Technology for Weld Defects Classification Based on Transfer Learning and Activation Features

Table 9

Comparison of the transfer learning-based pretrained DCNN models with works.

Study	Method	Probability of classification (%)
(23) (Kaftandjian et al., [43])	Region-based + reference + morphological geometrical fuzzy logic	80% at 0% false alarms
(29) (Valavanis et al., [44])	Graph-based (region growing + connectivity similarities); Geometrical, texture SVM, ANN, k-NN	46%–85%
(28) (Zapata, vilar, Ruiz, [42])	Region edge growing; geometrical + PCA multiple-layer ANN, adaptive network-based fuzzy inference system (ANFIS)	79%–83%
(26) (Kumar, et al. [40])	Edge, region growing, watershed texture, geometrical ANN	87%
(19) (Lim, Ratnam [46])	Thresholding, geometrical, multiple-layer ANN	97%
(27) (Zahran et al., [41])	Morphological + wavelet + region growing; Mel-frequency cepstral coefficients (MFCCs) and power density spectra ANN	75%