xwMOOC 딥러닝

이미지 주석/라벨(annotation) 달기

Tidyverse Korea

1 캡챠 데이터셋¹²

1.1 헬로월드

먼저 이미지 데이터와 라벨 데이터를 다운로드 받는다.

library(tidyverse)
library(magick)

one_img <- image_read("data/captcha_tutorial/9.jpeg")

one_img

1.2 이미지 + 라벨

캡챠 이미지 데이터와 라벨데이터를 결합 시켜보자.

library(glue)

filename <- list.files("data/captcha_tutorial/", pattern = "*.jpeg")

file_names_seq <- basename(filename) %>% 
  str_remove(pattern = ".jpeg") %>% 
  as.integer()

train_labels <- read_csv("data/captcha_tutorial/labels.txt", col_names = FALSE) %>% 
  rename(label = X1)

captcha_df <- tibble(filename = filename,
  filename_seq = file_names_seq) %>% 
  arrange(file_names_seq) %>% 
  slice(1:5000) %>% 
  bind_cols(train_labels) %>% 
  mutate(image = glue("data/captcha_tutorial/{filename_seq}.jpeg"))

captcha_df %>% tail

# A tibble: 6 x 4
  filename  filename_seq label  image                          
  <chr>            <int> <chr>  <glue>                         
1 4995.jpeg         4995 511236 data/captcha_tutorial/4995.jpeg
2 4996.jpeg         4996 144833 data/captcha_tutorial/4996.jpeg
3 4997.jpeg         4997 590583 data/captcha_tutorial/4997.jpeg
4 4998.jpeg         4998 891249 data/captcha_tutorial/4998.jpeg
5 4999.jpeg         4999 075695 data/captcha_tutorial/4999.jpeg
6 5000.jpeg         5000 209554 data/captcha_tutorial/5000.jpeg

1.3 캡챠 갤러리

library(trelliscopejs)

captcha_sample <- captcha_df %>%
  sample_n(100) %>%
  mutate(label = as.character(label),
         image = as.character(image),
         panel = img_panel_local(image))

trelliscope(captcha_sample, name = "captcha_tutorial", 
            nrow = 3, ncol = 7,
            path="captcha_tutorial"
            )

---

1. Rahul Agarwal (2019), “Deep Learning Based OCR for Text in the Wild”, Nanonets↩︎

2. “Captcha Breaking Tutorial”↩︎