Automatic Classification of Blue and White Porcelain Sherds Based on Data Augmentation and Feature Fusion

Figures & data

Figure 1. Data preprocessing.

Figure 2. Augmentation of the traditional image transformation.

Figure 3. FCutMix result.

Figure 4. Class activation map.

Figure 5. Feature map.

Figure 6. FFCNet structure diagram.

Figure 7. Examples of the archeological ceramic pieces: (a) lotus pattern, Yuan; b) decoration with interlock branch lotus, Ming; c) chrysanthemum pattern, Ming; d) peach grain, Qing; e) grape pattern, Qing; f) pattern with fold branch peony, Yuan.

Figure 8. Accuracy comparison.

Table 1. Results of multiple tests

	Cut-Rate	Recall	Kappa	F1 score	Accuracy	Std-acc	Mean-acc
Base	100%	0.690	0.641	0.684	0.679	0.015	0.698
		0.605	0.658	0.648	0.698
		0.597	0.677	0.651	0.717
	90%	0.543	0.650	0.609	0.693	0.050	0.634
	80%	0.519	0.613	0.581	0.660
	70%	0.516	0.609	0.577	0.656
	60%	0.485	0.565	0.542	0.613
	50%	0.444	0.494	0.490	0.547
FCutMix	100%	0.661	0.731	0.708	0.762	0.019	0.758
		0.707	0.702	0.720	0.734
		0.654	0.749	0.711	0.778
	90%	0.690	0.745	0.729	0.774	0.042	0.727
	80%	0.678	0.724	0.714	0.755
	70%	0.628	0.691	0.668	0.731
	60%	0.619	0.691	0.668	0.726
	50%	0.514	0.605	0.574	0.651
FFCNet	100%	0.678	0.757	0.728	0.785	0.007	0.795
		0.715	0.772	0.754	0.797
		0.704	0.777	0.750	0.802
	90%	0.688	0.748	0.729	0.776	0.017	0.745
	80%	0.637	0.715	0.687	0.746
	70%	0.600	0.709	0.664	0.741
	60%	0.640	0.702	0.683	0.733
	50%	0.650	0.696	0.687	0.728

Figure 9. Results of multiple tests.

Figure 10. The mean value of test results.

Table 2. Accuracy of each model

	Model 1	Model 2	Model 3	Integrated Learning (Voting strategy)	Integrated Learning (Average one-hot)	Integrated Learning (Weighted average one-hot)
Accuracy	79.31%	75.00%	80.17%	81.47%	81.90%	83.19%
Recall	66.76%	67.42%	70.36%	71.70%	72.88%	74.11%
F1 Score	0.7249	0.7101	0.7495	0.7627	0.7713	0.7839
Kappa	0.7663	0.7176	0.7767	0.7905	0.7955	0.8103

Figure 11. Confusion matrix.