Many Local Pattern Texture Features: Which Is Better for Image-Based Multilabel Human Protein Subcellular Localization Classification?

Table 4

Five evaluation index comparisons based on single-label and entire dataset by using BR model fed into different combinations of local and db6 global features.

Evaluation index	Single label samples (258 proteins)				Entire dataset (348 proteins)
	SLFs	SLFs_LBP	SLFs_CLBP	SLFs_LTrP	SLFs	SLFs_LBP	SLFs_CLBP	SLFs_LTrP
Subset accuracy	0.5058	0.5575	0.5872	0.5745	0.3825	0.4140	0.4555	0.4405
Accuracy	0.5141	0.5713	0.6153	0.5975	0.4586	0.4956	0.5451	0.5296
Recall	0.3855	0.4159	0.4640	0.4198	0.3528	0.3831	0.4240	0.4141
precision	0.3697	0.3944	0.4706	0.4610	0.3555	0.3802	0.4233	0.4552
Label accuracy	0.8353	0.8492	0.8508	0.8452	0.8020	0.8124	0.8360	0.8302

258 proteins correspond to single-label samples in our benchmark dataset, and 348 proteins denote entire dataset involved with single-label and multilabel samples.
All columns correspond to average of 2-fold on db6. SLFs denote global features involved with Haralick features and DNA-protein overlap features. Label accuracy denotes the average prediction accuracy of six labels.