OCR Roadmap

source

OCR preview stage basis: CNN and RNN

Video Preview CNN: convolutional neural network computer vision

• Knowledge Point 1: convolution neural network analysis
• Knowledge Point 2: over-fitting and random inactivation
• Knowledge Point 3: understanding the neural network convolution
• Knowledge Point 4: Typical network structure Detailed
• Actual project: the use of ResNet and inception to solve the general problem of image classification routines

RNN preview video: recurrent neural networks and natural language processing

• Knowledge Point 1: Recurrent Neural Networks
• Knowledge Point 2: the length of time-dependent problems with long memories network
• Knowledge Point 3: BPTT algorithm
• Actual project: the use of recurrent neural networks generate text poetry

The first stage character recognition and OCR technology Glance

The first lesson OCR Technology Overview

• Knowledge Point 1: What is OCR
• Knowledge Point 2: Overview of common applications, such as printed identification: document recognition, license plate recognition, license / business card / note recognition, video / image understanding, photo title search, as well as handwriting recognition: recognition online, offline recognition
• Knowledge Point 3: Basic process: preprocessing, layout analysis, positioning text lines, the character recognition, post-processing
• Knowledge Point 4: Common OCR tools: Tesseract, Abbyy, Baidu OCR API, IFLYTEK OCR API
• Real items: Tesseract OCR engine to build the first

Lesson single character segmentation and recognition

• Knowledge Point 1: Single character recognition (template matching, feature extraction + traditional classification, CNN model)
• Knowledge Point 2: 1 character segmentation: positional candidate dividing position (position of the projection candidate extraction rule, based on the candidate extraction position Model)
• Knowledge Point 3: 2 split font: split path selection (Viterbi algorithm / Beam Search, PCFG / 2D-PCFG)
• Real items: a single character recognition based on CNN and over-segmentation

Lesson character sequence recognition

• Knowledge Point 1: Overview RNN and LSTM + CTC, CRNN, RARE
• Knowledge Point 2: Overview of attention mechanisms Attention, and DRAM / DRAW, Transformer
• Real items: RNN and CNN combat the complex CRNN

The second stage: Master text positioning and text detection

The fourth lesson positioning the Bank

• Knowledge point 1: the conventional method (projection-based positioning text lines, text lines based on the positioning of the minimum spanning tree)
• 知识点2： 深度学习（Full-Page Text Recognition: Learning Where to Start and When to Stop、Learning Text-Line Localization with Shared and Local Regression Neural Networks、TextSnake）
• Actual project: actual Text Line Extraction Based on MST

Text Detection Lesson natural scene

• 知识点1： Reading Text in the Wild with Convolutional Neural Networks
• Knowledge Point 2: CTPN, RRPN, FTSN, DMPNet, EAST
• Knowledge Point 3: SegLink, PixelLink
• Knowledge Point 4: Textboxes, WordSup, FOTS
• Real items: CTPN combat scene text detection algorithm of

The third phase grasp another image problem

Lesson image quality enhancement and pre-processing

• Knowledge Point 1: image enhancement (to blur, super-resolution reconstruction)
• Knowledge Point 2: binarizing (global threshold and a local threshold: LOCAL light and noise, learning-based method)
• Knowledge Point 3: rotation / twist deformation (angle estimation and correction, and distortion restoration DocUNet)
• Actual project: actual generating confrontation GAN network variants SRGAN