42 lectures + 3 for exams

模式识别

模式描述:image -> RGB(100*100*3) vector (other repr exists) 特征选择:-> discard unused features for the above vector

task

  • preprocessing
    • video, img, signal processing
  • pattern separation
    • location detection / background / pattern-pattern separation
  • motion analysis
    • object tracking / movement pattern
  • pattern desp and classification
    • feature extraction/selection, classification, clustering, ML
  • application (on specific system)

topics in pattern classification

… blabla …

multi-modal

semi-supervised class.. / clustering distance metric learning weak supervised ???

pattern recognition formalization

example: human face recognition problem

模式描述 -> sample, label -> pattern classification

  • 2-class / multi-class
  • classifier design: ML
  • related: feature extraction / selection

模式表示: 特征表示 $x=[x_1,x_2, …, x_d]^T$

分类器表示: (???)

  • 类别模型: $M_i=M(x, \theta_i)$
  • 判别函数: $y_i=f(x, w_i)$ (discriminant function) 有的没有判别函数 本门课程大多可以写出

识别(分类):

  • distance metric(similarity): $\min_i d(x, M_i)$
  • decision region: $R_i=\arg\max_i f(x,w_i)$

fish example

firstly, data collection and fish feature distribution analsis

small dataset -> non-linear huge dataset(network / graph) -> linear (due to the huge amount and very high dimension, linear is good enough)

model complexity: string -> curve -> k-nearest-neighbor

systematic flow

data acquisition -> layout analysis -> character segmentation -> recognition -> post-preprocessing

?? recall: use multi-vector rather than a integer (compact embedding) ( [1,0,0]..[0,0,1] over 1, 2, 3 )

method category

statistical leanring / structual

statistical: parameterized / non-parameterized / … / ensemble / … structual: graph / …

other classification

??? blaaaaaaa

by generative / discriminative

generative: how to generative 找到内控因素,每个类分别学习,parameters are continuous

  • $p(\bf{x} \vert C) = f(\bf{x}, \theta)$ parameterized model
  • bayesian direct graph
  • (template-based classifier ???)
  • HMM

discriminative: given sample and label, unaware about how it is generated,所有类同时学习

  • Artificial NN
  • SVM
  • Boosting
  • CRF (in Ph.D. course)