The document summarizes two machine learning projects: 1. Face image compression using a multilayer perceptron neural network. The network takes in 16x16 pixel blocks and compresses them to lower dimensions while minimizing distortion. 2. Handwritten digit recognition using support vector machines. Images are preprocessed into 200-dimensional feature vectors capturing gradient information at different orientations. The SVM is trained on 70,000 samples and achieves 98.4% accuracy on a test set of 30,000 images.

1. MLP & Machine Learning Presentation 2003.12.10 정규준 민혜진 Shang, Y.L. Face Image Compression and Handwritten data recognizer

2. Overview Face Image Compression Problem using MLP Handwritten Digit Recognition

3. Face Image Compression Problem(1/2) Get training vectors from face images Discrete Cosine Transform (DCT) image per 8 by 8 block People can not easily recognize the distortion of the signal because of sensitiveness of the low frequency criteria Neural Network Structure Tool Matlab training function - Resilient back propagation, performance goal - 0.002

4. Face Image Compression Problem(2/2) Neural Network Structure - Multilayer neural network ▶ Number of Layers Input layer , Output layer: M=16 Combiner, Decombiner: M=16 Compressor: 1 ▶ Number of Neurons Input layer, Output layer: p 2 =64 Combiner, Decombiner: N h =32 Compressor: Q=16 ▶ Quantization of outputs 8bit(128/16) quantization in each output of compressor layer’s neurons

5. Handwritten Digit Recognition(1/2) Training method – Support Vector Machine svmlight ( http://svmlight.joachims.org ) Kernel function - polynomial Feature Vector 8 directional gradient features On each of 8 gradient maps, 5x5 Gaussian mask is imposed  200 dimension (x * x’) d where d is 4

6. Feature extraction outline Original image (various size) Normalized image (35 x 35) Sobel mask Mean filtering Image normalization 8 Gradient maps (35x35x8) Gaussian mask Feature vector (200x1)

7. Feature extraction step1 Mean filtering remove salt noise take the point the summation of whose neighbors is more than 5, otherwise remove it. Image normalization move original image in the center of square canvas shrink or enlarge square canvas  size of normalized image is 35x35 2 . . . 2 2 2 2

8. Feature extraction step2 Sobel mask in order to get 8 gradient feature maps compute the gradient components in two axes get direction vector and strength and record each map Gaussian mask split gradient feature maps and apply Gaussian mask  25 dimension per feature map 8 Gradient maps (35x35x8)

9. Training & Test Result Training – 70000 data Test – 30000 data Performance – 98.4% accuracy