Try   HackMD

Disentangling Writer and Character Styles for Handwriting Generation

Introduction

本篇論文主要針對中文手寫生成做優化
主要的概念在中文字並不是每個字都會用一樣的寫法
例如

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

所以本篇論文解離兩種style並用contrastive learning的方式來train它

  1. writer-wise styles
  2. characterwise styles

所以並不會產出不同style的手寫字，但是更會注意不同字的細節

Method

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

Writer loss就是找postive和negative pair 來train

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

glyph loss 用patch的方式來找postive和negative pair 來train

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

  • pen moving prediction loss
    真低看不太懂，有人可以解釋的嗎
  • pen state classification loss
    分類pan-down pan-up pan-end三種狀況

Image Not Showing Possible Reasons
  • The image was uploaded to a note which you don't have access to
  • The note which the image was originally uploaded to has been deleted
Learn More →

Last changed by 
 
履歷軌
0
301

Read more

研究內容

發展歷史

May 5, 2025
Multiple Object Tracking (MOT)

Introduction 物件追蹤包含兩個部分: <font color=#20639B>物件偵測</font> (Object detection)物件偵測在眾多算法百家爭鳴下， 其準確度已經高到一個境界，舉凡 YOLO, SSD, Retinanet, CenterNet, …都是很好的選擇，它的功用就是要抓到 image 內的 bounding box 以及物件classification <font color=#20639B>追蹤器</font> (tracker) 追蹤器要做的事呢基本上就是==判斷前後 frame 抓到的 object 是否屬於同一個==，若是則 assign 相同 ID，若否則 assign 新的 ID。如下圖，frame t 檢測到的 3 個 objects (黃、藍、紅)，frame t+1 檢測到 4 個(灰色框)，而要如何把前後 frame 的框關連起來就是物件追蹤要做的事 追蹤器 (tracker)

May 2, 2024
Week 2：模型壓縮 - 剪枝 (pruning)

PRUNING FILTERS FOR EFFICIENT CONVNETS 剪枝 - Filter & feature maps 數學符號與圖例說明 $n_{i}$ : 第 i 層 input channel 數 $F_{i,j}$ : 第 i 層的第 j 個 filter $X_{i}$ : 第 i 層的 feature maps 模型運算量計算 :question: 如果剪掉其中一個 filter $F_{i,j}$，會減少多少運算量呢？ 圖示如下：

Mar 26, 2024
HandWriting

AttentionHTR: Handwritten Text Recognition Based on Attention Encoder-Decoder Networks github: https://github.com/dmitrijsk/attentionhtr 很廢，就是拿Attention Model拿來fine tune連code都直接從clova那邊clone過來，augmentation都放在furture work的paper 但我想拿這篇來用... 用Iam和Imgur5K兩個手寫dataset來FineTune Transformer-based Optical Character Recognition with Pre-trained Models github: https://github.com/microsoft/unilm/tree/master/trocr

Feb 14, 2023
Read more from 履歷軌
Published on HackMD