index.md (3767B)
1 +++ 2 title = 'Classification' 3 +++ 4 # Classification 5 a pattern is an entity vaguely defined, that could be given a name. 6 recognition is identification of pattern as member of a category 7 8 ## Types of pattern recognition (classification) systems 9 10 - speech recognition: 11 12 1. PC card converts analog waves from mic into digital format 13 2. acoustical model breaks the word into phonemes 14 3. language model compares phonemes to words in built-in dictionary 15 4. software decides on what spoken word was and displays best match 16 17 - brain-computer interface that acquires signals directly from the brain 18 - gesture recognition using acceleration magnitude from watch 19 - image recognition 20 21 ## Classification (known categories) 22 23 - given a few classes, each item belongs to one class 24 - objects are described by features 25 - system needs a training set (both positive and negative examples) 26 - if a new item comes, its features are measured and the system decides which class it belongs to 27 28  29 30 Components: 31 32 - Sensing module 33 - Preprocessing mechanism 34 - Feature extraction mechanism 35 - Classifier 36 - Training set of already classified examples 37 38 ### Building a pattern recognition system 39 40 1. Choose features, define classes (e.g. coins 10 cent, 20 cent, 50 cent, 1$, 2$) 41 42 - features need to have discriminative power 43 - not too many, but enough to reliably separate classes based on them 44 - e.g. coins colour and diameter 45 - algorithms 46 - simple: rule-based activity recognition (If…And/Or…Then) 47 - complicated: machine learning decision trees, HMM, neural networks 48 49 2. Extract features 50 51 - image recognition 52 - shape decriptors 53 - form factor (round object has 1, others smaller) 54 - Euler number (number of objects minus number of holes in objects) 55 - perimeter, area, roundness ratio… 56 - preprocessing 57 - binarisation, morphological operators, segmentation 58 - extract features (e.g. area, coordinates of centre of mass) 59 - optical character recognition (OCR) 60 - converts image into machine readable text 61 - uses statistical moments (total mass, centroid, elliptical parameters, etc.) 62 - invariant moments of Hu 63 - sound recognition 64 - features 65 - frequency spectrum 66 - spectrograms 67 - Mel cepstrum coefficients — FFT to Log(|x|) to IFFT results in cepstrum 68 - vowels recognition: second formant vs first formant frequency for vowels (significant freqs) 69 70 3. Train the classifier 71 4. Evaluate the performance of classification 72 73 ### Classifiers 74 #### Rule-based: if-then-else 75 76 - exhaustive, mutually exclusive rules 77 - works well if there aren’t too many features 78 79 #### Template-matching: a set of reference patterns is available, match an unknown using nearest-neighbour 80 81 get a fingerprint for a specific signal, using FFT (freq. spectrum) or Mel cepstrum coefficients 82 83 train with various words, store fingerprints, and then apply 84 two approaches: 85 86 - maximum correlation 87 - minimum error — calculate Euclidian distance between vectors 88 89 #### Neural networks: 90 synapses are weights 91 92 output is binary, depends on comparison between weighted sum of inputs and threshold θ 93 94 a neuron has: 95 96 - set of weighted inputs — dendrites+synapses 97 - an adder — soma 98 - an activation function to decide whether or not the neuron fires 99 100 a neuron cannot learn, but a perceptron can. by changing the weights which are adjustable. 101 102 neural networks are collections of artificial neurons, and have hidden layers. 103 104 they learn by testing output against desired output and adjusting weights accordingly. 105 106 