Fig 1.
Complex defective wafer patterns.
Notes: (a) Edge-ring, (b) Donut, (c) Scratch.
Fig 2.
Local defective wafer patterns.
Fig 3.
Methodology of the proposed hybrid system.
Fig 4.
Distribution of Defect Classes in the Dataset.
Fig 5.
Defect and non-defect wafer map patterns before histogram equalization.
Notes: (i) Center (ii) Donut (iii) Edge local (iv) Edge ring (v)Local (vi) Near full (vii) Random (viii) Scratch (ix) None.
Fig 6.
Defect and non-defect wafer map patterns after histogram equalization.
Notes: (i) Center (ii) Donut (iii) Edge local (iv) Edge ring (v)Local (vi) Near full (vii) Random (viii) Scratch (ix) None.
Fig 7.
Histogram of equalized and un-equalized defective images.
Notes: (a) histogram of unequalized Center defect image; (b) histogram of equalized Center defect image; (c) histogram of unequalized Edge-Local defect image; (d) histogram of equalized Edge-Local defect image; (e) histogram of unequalized Edge-Ring defect image; (f) histogram of equalized Edge-Ring defect image; (g) histogram of the unequalized Local defect image; (h) histogram of equalized Local defect image; (i) histogram of unequalized Scratch defect image; (j) histogram of equalized Scratch defect image.
Fig 8.
Structure of original MobileNet [28].
Fig 9.
Graphical representation of ReLU and Swish activation functions.
Fig 10.
Structure of modifiedMobileNet [28].
Fig 11.
The multi-head attention mechanism process.
Fig 12.
The detailed architecture of the proposed hybrid model.
Fig 13.
The block diagram of the ECOC coding matrix.
Table 1.
Experimental Settings.
Table 2.
Performance analysis of proposed hybrid model configurations on the imbalanced WM-811k dataset (image size:32 × 32).
Fig 14.
Analysis of the proposed model using several optimizers.
Fig 15.
Contribution of the Swish activation function to the proposed model.
Fig 16.
Contribution of the multi-head attention to the proposed model.
Fig 17.
Contribution of the ECOC-SVM classifier to the proposed model.
Fig 18.
Impact of histogram equalization on ‘Local’ defect class.
Fig 19.
Overall performance evaluation of the proposed model based on the histogram equalization process.
Fig 20.
Confusion matrix of the proposed model.
Table 3.
Performance analysis of the proposed hybrid model based on different classifiers.
Fig 21.
Comparative analysis based on testing accuracy.
Fig 22.
Comparative analysis based on MFLOPs.
Fig 23.
Comparative analysis based on total parameters.
Table 4.
Comparison of the proposed hybrid approach with other methods based on average values of performance metrics.
Table 5.
Comparison of the proposed hybrid approach with other systems based on the performance metrics.
Table 6.
Comparison of the proposed hybrid model and the base MobileNet model based on dataset balancing.
Table 7.
Comparison of the proposed hybrid approach with other developed methods- based on the classification accuracy of defect classes on the WM-811k dataset.
Fig 24.
Comparison of the proposed hybrid approach with other methods based on accuracy.