Gabor feature extraction for handwritten character recognition

• gabor_feature_extraction.R: an R function that extracts 72 features from an input image using Gabor filters (detailed explanation below).

• spatial_domain_feature_extraction.R: an R function that extracts 12 simple features from the spatial domain of the input image.

• MNIST_example/feature_extraction.R: feature extraction example for a reduced MNIST dataset (4000 samples).

• MNIST_example/svm_classification.py: SVM classification example using extracted features.

Feature extraction

Gabor features

The input image is filtered with a set of 18 Gabor filters. Gabor filters are generated using 3 different wavelengths and 6 different orientations. 18 filtered images are obtained for each sample. The standard deviation, mean, kurtosis and skewness of the filtered images are computed. This forms a vector of 72 features for each sample image.

Input image should consist of the matrix corresponding to a single-channel digital image where each pixel is represented by a whole number ranging from 0 to 225.

Spatial domain features

Other simple features are extracted from the spatial domain of the sample images.

Results

Features are classified using support vector machines (SVM). These results are comparable to those obtained by using convolutional neural networks (CNN).

Dataset	Accuracy
MNIST	98%
Reduced MNIST (4000 samples)	95%