From: Feature analysis for classification of trace fluorescent labeled protein crystallization images
Research paper | Image categories | Feature extraction | Classification method | Classification accuracy |
---|---|---|---|---|
Zuk and Ward (1991) [7] | NA | Edge features | Detection of lines using Hough transform and line tracking | Not provided |
Walker et al. (2007) [22] | 7 | Radial and angular descriptors from Fourier Transform | Learning vector quantization | 14 - 97% for different categories |
Xu et al. (2006) [23] | 2 | Features from multiscale Laplacian pyramid filters | Neural network | 95% accuracy |
Wilson (2002) [24] | 3 | Intensity and geometric features | Naive Bayes | Recall 86% for crystals, 77% for unfavourable objects |
Hung et al. (2014) [26] | 3 | Shape context, Gabor filters and Fourier transforms | Cascade classifier on naive Bayes and random forest | 74% accuracy |
Spraggon et al. (2002) [17] | 6 | Geometric and texture features | Self-organizing neural networks | 47 to 82% for different categories |
Cumba et al. (2003) [8] | 2 | Radon transform line features and texture features | Linear discriminant analysis | 85% accuracy with roc 0.84 |
Saitoh et al. (2004) [20] | 5 | Geometric and texture features | Linear discriminant analysis | 80 - 98% for different categories |
Bern et al. (2004) [15] | 5 | Gradient and geometric features | Decision tree with hand crafted thresholds | 12% FN and 14% FP |
Cumba et al. (2005) [9] | 2 | Texture features, line measures and energy measures | Association rule mining | 85% accuracy with ROC 0.87 |
Zhu et al. (2004) [10] | 2 | Geometric and texture features | Decision tree with boosting | 14.6% FP and 9.6% FN |
Berry et al. (2006) [11] | 2 | NA | Learning vector quantization, self organizing maps and bayesian algorithm | NA |
Pan et al. (2006) [12] | 2 | Intensity stats, texture features, Gabor wavelet decomposition | Support vector machine | 2.94% FN and 37.68% FP |
Yang et al. (2006) [14] | 3 | Hough transform, DFT, GLCM features | Hand tuned thresholds | 85% accuracy |
Saitoh et al. (2006) [16] | 5 | Texture features, differential image features | Decision tree and SVM | 90% for 3-class problem |
Po and Laine (2008) [13] | 2 | Multiscale Laplacian pyramid filters and histogram analysis | Genetic algorithm and neural network | Accuracy: 93.5% with 88% TP and 99% TN |
Liu et al. (2008) [21] | Crystal likelihood | Features from Gabor filters, integral histograms, and gradient images | Decision tree with boosting | ROC 0.92 |
Cumba et al. (2010) [18] | 3 and 6 | Basic stats, energy, Euler numbers, Radon-Laplacian, Sobel-edge, GLCM | Multiple random forest with bagging and feature subsampling | Recall 80% crystals, 89% precipitate, 98% clear drops |
Sigdel et al. (2013) [28] | 3 | Intensity and blob features | Multilayer perception neural network | 1.2% crystal misses with 88% accuracy |
Sigdel et al. (2014) [25] | 3 | Intensity and blob features | Semi-supervised | 75% - 85% overall accuracy |
Dinc et al. (2014) [27] | 3 and 2 | Intensity and blob features | 5 classifiers, feature reduction using PCA | 96% on non-crystals, 95% on likely-leads |
Yann et al. (2016) [19] | 10 | Deep learnining on grayscale image | Deep CNN with 13 layers | 90.8% accuracy |