Senin, 04 Maret 2019

Python-Scene Text detection

Scene text detection merupakan suatu proses untuk mengkonversi region text  yang memungkinkan dibaca oleh sebuah komputer, ada perbedaan yang mendasar mengenai recognition vs scene text detection seperti berikut
  1. clean background vs. cluttered background
  2. regular font vs. various fonts
  3. plain layout vs. complex layouts
  4. monotone color vs. different colors

Tantangan utama dari Scene text detection yaitu
  1. Diversity of scene text: different colors, scales, orientations, fonts, languages
  2. Complexity of background:  elements like signs, fences, bricks, and grasses are virtually indistinguishable from true text
  3. Various interference factors:  noise, blur, non-uniform illumination, low resolution, partial occlusion
Metode Konvensional untuk text detection yaitu
MSER:
  1. extract character candidates using MSER (Maximally Stable Extremal Regions), assuming similar color within each character
  2. robust, fast to compute, independent of scale
  3. limitation: can only handle horizontal text, due to features and linking strategy

SWT
  1. extract character candidates with SWT (Stroke Width Transform), assuming consistent stroke width within each character
  2. robust, fast to compute, independent of scale
  3. limitation: can only handle horizontal text, due to features and linking strategy
Deep Learning menggunakan EAST yaitu Efficient and Accurate Scene Text
https://arxiv.org/abs/1704.03155

Kode: https://github.com/argman/EAST
  1. main idea: predict location, scale and orientation of text with a single model and multiple loss functions (multi-task training)
  2. advantages: (a). accuracy: allow for end-to-end training and optimization;  (b). efficiency: remove redundant stages and processings
Lagi malas buat terjemahkan (:)

Tidak ada komentar: