MNIST data classification using ML approach
• Feature Extraction – We use two methods for feature extraction: Feature extraction using HOG feature detection and LBPH method. – Histogram of Oriented Gradients (HOG) feature detection to extract the feature of MNIST dataset. – LBPH algorithm is the combination of Local Binary Patterns (LBP) and Histograms in given radius with number of points used for his- togram bins.
• Feature Learning We use the logistic regression model to train the ex- tracted features with the label and predict the output for each test image features extracted.
• Classification accuracy – Acuuracy with LBPH feature extraction As mentioned the two variations resulted in below accuracy : With 64 bins and 16 radius - 54.68% accuracy With 64 bins and 2 radius - 58.63% accuracy – Acuuracy with HOF feature extraction We experimented with two different cell size for HOG feature extraction: 7 x 7 and 14 x 14. With the above cell size we got 94.22% accuracy with 7 x 7 cell size and 83.43% accuracy with 14 x 14 cell size. This indicates the importance of feature extraction. With 7 x 7 cell size we got 72 features for each sample while with 14 x 14 cell size we got 36 features for each sample. – Impact of features extracted is clearly visible in the accuracy ob- tained.
MNIST data classification using DL approach
• Conv2D A filter (3x3) slides over the input image (28x28) with stride=1and com- putes dot product to give a ‘feature map (26x26)’. (size of input image – size of filter + 1) = 28 – 3 + 1 = 26 For RGB image, 3 distinct filters are stacked to produce 3 feature maps (depth=3). In case of MNIST images, depth of feature map is 1 as their pixel values are grayscale. For 3 convolutional layers, the output feature map is 22x22. • MaxPooling Largest element from the feature map within a window of 2x2 is consid- ered, thereby resulting in an output image of size 11x11. This reduces the spatial size of the input and also computations in the network. • Flatten This layer converts the 2D input image (13x13) to a 1-D vector (11x11x24 = 2904) • Fully Connected (FC) First FC layer contains 200 neurons with rectifier activation function (ReLu). This layer takes the inputs from the feature analysis and ap- plies weights to predict the correct label. Final FC layer containing 10 neurons with softmax gives the final prob- abilities for each class label. The neuron with a maximum probability is the output for the model.
Maximum accuracy obtained with this model is 98.71% with 3 convlutional layers, 0.5 dropout, 200 neurons and 25 epochs.