COCO-Text 场景文本识别 数据集裁剪代码及裁剪完毕版本下载

我们知道，要下COCO-Text这个数据集需要下载12GB的COCO2014数据集：http://mscoco.org/dataset/#download

然后前往COCO-Text官网：https://bgshih.github.io/cocotext/#h2-download

下载API。下面附上裁剪用代码 根据SARS的论文，我推断他就使用了易读的图片

这里coco使用的是比较新的cocotextv2。如此裁剪后有14k张图像用作验证集，59k张图像用作训练集

此外，处理好的GT和数据集可以点击这里下载：https://download.csdn.net/download/zhaominyiz/12407232

裁剪示例： 可以发现COCO-T这个数据集难度还是挺高的

    

代码：

import coco_text
import numpy as np
import skimage.io as io
import matplotlib.pyplot as plt
from PIL import Image
import pylab
import math
import traceback
import os
import html

pylab.rcParams['figure.figsize'] = (10.0, 8.0)

dataDir='F:\dataset\COCO'
dataType='train2014'

ct = coco_text.COCO_Text('../cocotext.v2.json')
ct.info()
imgIds = ct.getImgIds(imgIds=ct.train,catIds=[('legibility','legible')])
# return
# img = ct.loadImgs(imgIds[np.random.randint(0,len(imgIds))])[0]
# print("???",'%s/%s/%s'%(dataDir,dataType,img['file_name']))
# I = io.imread('%s/%s/%s'%(dataDir,dataType,img['file_name']))
# print(I.shape)
# print ('/%s/%s'%(dataType,img['file_name']))
# plt.figure()
f = open("./COCO-TEXT/gt_train.txt","w",encoding="utf-8")

print("???",len(imgIds))
# os._exit(233)
cnt = 0
for idx in range(0,len(imgIds)):
    img = ct.loadImgs(imgIds[idx])[0]
    I = io.imread('%s/%s/%s'%(dataDir,dataType,img['file_name']))
    # plt.imshow(I)
    # plt.show()
    # os._exit("233")
    if idx%100==0:
        print(idx,"/",len(imgIds))
    # print('%s/%s/%s'%(dataDir,dataType,img['file_name']),"___",I.shape,len(I.shape))
    annIds = ct.getAnnIds(imgIds=img['id'])
    anns = ct.loadAnns(annIds)
    # print("ANNS=",anns)
    # ct.showAnns(anns)
    for GT in anns:
        if "utf8_string" not in GT or GT["language"]!="english":
            continue
        # print("GT=",GT)
        # os._exit(233)
        label = GT["utf8_string"]
        label = label.strip()
        if label =="" or label ==" ":
            continue
        label=html.unescape(label)
        bb = GT["bbox"]
        x = math.floor(bb[0])
        y = math.floor(bb[1])
        dx = math.ceil(bb[2])
        dy = math.ceil(bb[3])
        xx = x+dx
        yy = y+dy
        # print(xx,yy)
        if xx>=I.shape[1]:
            xx = I.shape[1]-1
        if yy>=I.shape[0]:
            yy = I.shape[0]-1
        if len(I.shape)==2:
            croppedimg = I[y:yy,x:xx]
        else:
            croppedimg = I[y:yy,x:xx,:]
        # print(label)
        cnt+=1
        url = './COCO-TEXT/train/'+str(cnt)+".jpg"
        gt_url = "./train/"+str(cnt)+".jpg"
        try:
            io.imsave(url,croppedimg)
            f.write(url+" "+label+"\n")
        except Exception as ex:
            cnt-=1
            # traceback.print_exc()
            print(ex)
        # plt.imshow(croppedimg)
        # plt.show()

f.close()