Getting familiar with the data

In this notebook, we are going to do some basic exploration of the dataset. We shall observe some samples from each dataset and number of image samples in them. As images are obtained from multiple sources, the quality of the segmentation and the dimension of the images varies a lot. We are going use data analysis toolkit pandas to perform our exploration. If you are not familiar with this library check out their 10 Minutes to pandas short introduction.

In [262]:
# Importing necessary libraries
import numpy as np
import os
import glob
import cv2
import matplotlib.pyplot as plt
import pandas as pd
In [99]:
# Declare fontsize variables which will be used while plotting the data
FS_AXIS_LABEL=14
FS_TITLE=17
FS_TICKS=12
FIG_WIDTH=16

While writing the codes, files and folder was organized in the following way

  • OpenBengali
    • code
    • Final_DB

The code folder contains this jupyter notebook, and the Final_DB folder has the raw image datasets.

In [203]:
project_dir='..'

Let's see what is in the Final_DB folder.

In [204]:
os.listdir(os.path.join(project_dir,'Final_DB'))
# Notice that I am using the os.path.join() function to create the filepaths instead of writing them down explicitly with a 
# filepath separator ('\\' for windows '/' for linux). This allows us to run this notebook both in windows and linux 
# environment without manually changing the filepath separator
Out[204]:
['testing-a',
 'testing-a.csv',
 'testing-b',
 'testing-b.csv',
 'testing-e',
 'testing-e.csv',
 'training-a',
 'training-a.csv',
 'training-b',
 'training-b.csv',
 'training-e',
 'training-e.csv']

Our dataset has images from sources A, B, and E. Tha datasets are already split into train and test sets. The labels of each image is saved in the corresponding .csv files. Let's see some samples of the images.

Setup the path variables

All the images have .png extension. We are going get all the filepaths that have .png extensions by using the glob.glob() function

In [43]:
paths_train_a=glob.glob(os.path.join(project_dir,'Final_DB','training-a','*.png'))
paths_train_b=glob.glob(os.path.join(project_dir,'Final_DB','training-b','*.png'))
paths_train_e=glob.glob(os.path.join(project_dir,'Final_DB','training-e','*.png'))
paths_test_a=glob.glob(os.path.join(project_dir,'Final_DB','testing-a','*.png'))
paths_test_b=glob.glob(os.path.join(project_dir,'Final_DB','testing-b','*.png'))
paths_test_e=glob.glob(os.path.join(project_dir,'Final_DB','testing-e','*.png'))
path_label_train_a=os.path.join(project_dir,'Final_DB','training-a.csv')
path_label_train_b=os.path.join(project_dir,'Final_DB','training-b.csv')
path_label_train_e=os.path.join(project_dir,'Final_DB','training-e.csv')
path_label_test_a=os.path.join(project_dir,'Final_DB','testing-a.csv')
path_label_test_b=os.path.join(project_dir,'Final_DB','testing-b.csv')
path_label_test_e=os.path.join(project_dir,'Final_DB','testing-e.csv')

Some Utility Functions

In [205]:
def get_img(path,mode=cv2.IMREAD_GRAYSCALE):
    # read image (if no read mode is defined, the image is read in grayscale)
     return cv2.imread(path,mode)   
def imshow_group(paths,n_per_row=10):
    # plot multiple digits in one figure, by default 10 images are plotted per
    n_sample=len(paths)
    j=np.ceil(n_sample/n_per_row)
    fig=plt.figure(figsize=(20,2*j))
    for i, path in enumerate(paths):
        img=get_img(path)
        plt.subplot(j,n_per_row,i+1)
        plt.imshow(img,cmap='gray')  
        plt.title(img.shape)
        plt.axis('off')
    return fig
def get_key(path):
    # separate the key from the filepath of an image
    return path.split(sep=os.sep)[-1]

Check a few samples from dataset A

We are going to randomly choose a few image filepaths from trainnig set A. Then load them in grayscale and plot them.

In [269]:
paths=np.random.choice(paths_train_a,size=40)
fig=imshow_group(paths)
fig.suptitle('Samples from {} training images in dataset A'.format(len(paths_train_a)), fontsize=FS_TITLE)
plt.show()

The digits do not fill up the entire image and most of them are not centered.

Shape statistics of dataset A

Next, we are going to observe if images in dataset A have different shapes. To do this, we are going to put the image shapes in a pandas series object and use its .value_counts() attribute to obtain the counts of its unique values.

Train A

In [208]:
shapes_train_a_sr=pd.Series([get_img(path).shape for path in paths_train_a])
In [209]:
shapes_train_a_sr.value_counts()
Out[209]:
(180, 180)    19702
dtype: int64

All the images have a fixed shape of 180 x 180

Test A

In [210]:
shapes_test_a_sr=pd.Series([get_img(path).shape for path in paths_test_a])
In [211]:
shapes_test_a_sr.value_counts()
Out[211]:
(180, 180)    3489
dtype: int64

Class distribution statistics of dataset A

Let's observe the frequency of each digit in dataset A.

Train A

Let's read the .csv file which contains the labels of dataset A. We are using read_csv() function from the pandas library which will return the content of the .csv file in a dataframe.

In [219]:
df_train_a=pd.read_csv(path_label_train_a)
df_train_a.head() # Observe first five rows
Out[219]:
filename original filename scanid digit database name original contributing team database name
0 a00000.png Scan_58_digit_5_num_8.png 58 5 BHDDB Buet_Broncos training-a
1 a00001.png Scan_73_digit_3_num_5.png 73 3 BHDDB Buet_Broncos training-a
2 a00002.png Scan_18_digit_1_num_3.png 18 1 BHDDB Buet_Broncos training-a
3 a00003.png Scan_166_digit_7_num_3.png 166 7 BHDDB Buet_Broncos training-a
4 a00004.png Scan_108_digit_0_num_1.png 108 0 BHDDB Buet_Broncos training-a

Next, we are going to replace the numerical index of the dataframe with the filename column which will give us more convenient access to the dataframe elements.

In [213]:
df_train_a=df_train_a.set_index('filename')
df_train_a.head() # Observe first five rows
Out[213]:
original filename scanid digit database name original contributing team database name
filename
a00000.png Scan_58_digit_5_num_8.png 58 5 BHDDB Buet_Broncos training-a
a00001.png Scan_73_digit_3_num_5.png 73 3 BHDDB Buet_Broncos training-a
a00002.png Scan_18_digit_1_num_3.png 18 1 BHDDB Buet_Broncos training-a
a00003.png Scan_166_digit_7_num_3.png 166 7 BHDDB Buet_Broncos training-a
a00004.png Scan_108_digit_0_num_1.png 108 0 BHDDB Buet_Broncos training-a

As we can see the labels of each digit is located under the digit column. We are going to put the labels of each digit in a series object and use the .value_counts() attribute to get the count of each digit.

In [214]:
labels_train_a_sr=pd.Series([df_train_a.loc[get_key(path)]['digit'] for path in paths_train_a])
In [264]:
labels_train_a_sr_vc=labels_train_a_sr.value_counts()

The pandas series object has a .plot() attribute which is useful to plot the contents of the series. We are going to use it to make a plot of the frequency of each digit.

In [215]:
plt.figure(figsize=(FIG_WIDTH,5))
labels_train_a_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Train A\nMean frequency of digits per class: {}, Standard Deviation: {:.4f} '.format(labels_train_a_sr_vc.mean(),labels_train_a_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

Test A

Let's see the class distribution statistics in the test set.

In [220]:
df_test_a=pd.read_csv(path_label_test_a)
df_test_a=df_test_a.set_index('filename')
df_test_a.head()
Out[220]:
original filename scanid digit database name original contributing team database name
filename
a00000.png Scan_178_digit_4_num_2.png 178 4 BHDDB Buet_Broncos test-a
a00001.png Scan_206_digit_9_num_1.png 206 9 BHDDB Buet_Broncos test-a
a00002.png Scan_175_digit_3_num_5.png 175 3 BHDDB Buet_Broncos test-a
a00003.png Scan_96_digit_0_num_2.png 96 0 BHDDB Buet_Broncos test-a
a00004.png Scan_131_digit_4_num_4.png 131 4 BHDDB Buet_Broncos test-a
In [221]:
labels_test_a_sr=pd.Series([df_test_a.loc[get_key(path)]['digit'] for path in paths_test_a])
In [218]:
labels_test_a_sr_vc=labels_test_a_sr.value_counts()
plt.figure(figsize=(FIG_WIDTH,5))
labels_train_a_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Test A\nMean frequency of digits per class: {}, Standard Deviation: {:.4f} '.format(labels_test_a_sr_vc.mean(),labels_test_a_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

The digit classes are well balanced both in the test and train set. We are going to repeat the same steps for dataset B and E.

Check a few samples from dataset B

In [254]:
paths=np.random.choice(paths_train_b,size=40)
fig=imshow_group(paths)
fig.suptitle('Samples from {} training images in dataset B'.format(len(paths_train_b)), fontsize=FS_TITLE)
plt.show()

Similar to dataset A, most of the digits in B do not fill the entire image. Also most of them are not centered.

Shape statistics of dataset B

Train B

In [222]:
shapes_train_b_sr=pd.Series([get_img(path).shape for path in paths_train_b])
In [223]:
shapes_train_b_sr_vc=shapes_train_b_sr.value_counts()
In [224]:
plt.figure(figsize=(FIG_WIDTH,5))
shapes_train_b_sr_vc.plot(kind='bar')
plt.xlabel('Image Shapes', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.xticks(fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.title('Train B\nNo. of unique shapes: {}'.format(shapes_train_b_sr_vc.count()),fontsize=FS_TITLE)
plt.show()

Test B

In [57]:
shapes_test_b_sr=pd.Series([get_img(path).shape for path in paths_test_b])
In [58]:
shapes_test_b_sr_vc=shapes_test_b_sr.value_counts()
In [143]:
plt.figure(figsize=(FIG_WIDTH,5))
shapes_test_b_sr_vc.plot(kind='bar')
plt.xlabel('Image Shapes', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.xticks(fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.title('Test B\nNo. of unique shapes: {}'.format(shapes_test_b_sr_vc.count()),fontsize=FS_TITLE)
plt.show()

The image shapes in dataset B varies a lot.

Class distribution statistics of dataset B

Train B

In [225]:
df_train_b=pd.read_csv(path_label_train_b)
df_train_b=df_train_b.set_index('filename')
df_train_b.head()
Out[225]:
original filename scanid digit database name original contributing team database name
filename
b00000.png 0_0005.png 5 0 B101DB Shongborton training-b
b00001.png 2_0038.png 38 2 B101DB Shongborton training-b
b00002.png 5_0037.png 37 5 B101DB Shongborton training-b
b00003.png 2_0039.png 39 2 B101DB Shongborton training-b
b00004.png 5_0028.png 28 5 B101DB Shongborton training-b
In [226]:
labels_train_b_sr=pd.Series([df_train_b.loc[get_key(path)]['digit'] for path in paths_train_b])
In [227]:
plt.figure(figsize=(FIG_WIDTH,5))
labels_train_b_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Train B\nMean frequency of digits per class: {}, Standard Deviation: {}'.format(labels_train_b_sr_vc.mean(),labels_train_b_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

Test B

In [228]:
df_test_b=pd.read_csv(path_label_test_b)
df_test_b=df_test_b.set_index('filename')
df_test_b.head()
Out[228]:
original filename scanid digit database name original contributing team database name
filename
b00000.png 0_0036.png 36 0 B101DB Shongborton testing-b
b00001.png 7_0024.png 24 7 B101DB Shongborton testing-b
b00002.png 9_0016.png 16 9 B101DB Shongborton testing-b
b00003.png 2_0024.png 24 2 B101DB Shongborton testing-b
b00004.png 4_0027.png 27 4 B101DB Shongborton testing-b
In [229]:
labels_test_b_sr=pd.Series([df_test_b.loc[get_key(path)]['digit'] for path in paths_test_b])
In [230]:
labels_test_b_sr_vc=labels_test_b_sr.value_counts()
In [231]:
plt.figure(figsize=(FIG_WIDTH,5))
labels_test_b_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Test B\nMean frequency of digits per class: {}, Standard Deviation: {}'.format(labels_test_b_sr_vc.mean(),labels_test_b_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

Checking samples from dataset E

In [252]:
paths=np.random.choice(paths_train_e,40)
fig=imshow_group(paths)
fig.suptitle('Samples from {} training images in dataset E'.format(len(paths_train_e)), fontsize=FS_TITLE)
plt.show()

The images are cropped well and they have minimum non digit area.

Shape statistics of dataset E

Train E

In [115]:
shapes_train_e_sr=pd.Series([get_img(path).shape for path in paths_train_e])
In [121]:
shapes_train_e_sr.nunique()
Out[121]:
1515

There are 1515 unique shapes in training dataset E.

In [122]:
shapes_train_e_sr_vc=shapes_train_e_sr.value_counts()

Plotting 50 most frequently occuring shapes.

In [146]:
plt.figure(figsize=(FIG_WIDTH,5))
shapes_train_e_sr_vc.iloc[:50].plot(kind='bar')
plt.xticks(fontsize=10)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Image Shapes', fontsize=FS_AXIS_LABEL)
plt.ylabel('Count', fontsize=FS_AXIS_LABEL)
plt.title('Train E\nNo. of unique shapes: {}\nPlot of 50 most frequently occuring shapes'.format(shapes_train_e_sr_vc.count()),fontsize=FS_TITLE)
plt.show()

Test E

In [265]:
shapes_test_e_sr=pd.Series([get_img(path).shape for path in paths_test_e])
In [152]:
shapes_test_e_sr.nunique()
Out[152]:
555

There are 555 unique shapes in test dataset E.

In [149]:
shapes_test_e_sr_vc=shapes_test_e_sr.value_counts()
In [271]:
plt.figure(figsize=(FIG_WIDTH,5))
shapes_test_e_sr_vc.iloc[:50].plot(kind='bar')
plt.xticks(fontsize=10)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Image Shapes', fontsize=FS_AXIS_LABEL)
plt.ylabel('Count', fontsize=FS_AXIS_LABEL)
plt.title('Test E\nNo. of unique shapes: {}\nPlot of 50 most frequently occuring shapes'.format(shapes_test_e_sr_vc.count()),
          fontsize=FS_TITLE)
plt.show()

Class distribution statistics of dataset E

Train E

In [154]:
df_train_e=pd.read_csv(path_label_train_e)
df_train_e=df_train_e.set_index('filename')
df_train_e.head()
Out[154]:
original filename districtid institutionid gender age datestamp scanid digit database name original database name
filename
e00000.png 01_0001_0_17_0916_0193_55.png 1 1 0 17 916 193 4 BanglaLekha-Isolated training-e
e00001.png 02_0002_0_23_1016_0858_57.png 2 2 0 23 1016 858 6 BanglaLekha-Isolated training-e
e00002.png 02_0002_0_14_1016_1673_51.png 2 2 0 14 1016 1673 0 BanglaLekha-Isolated training-e
e00003.png 02_0002_0_26_1016_1828_56.png 2 2 0 26 1016 1828 5 BanglaLekha-Isolated training-e
e00004.png 02_0002_1_19_1016_1336_59.png 2 2 1 19 1016 1336 8 BanglaLekha-Isolated training-e
In [156]:
labels_train_e_sr=pd.Series([df_train_e.loc[get_key(path)]['digit'] for path in paths_train_e])
In [266]:
labels_train_e_sr_vc=labels_train_e_sr.value_counts()
In [160]:
plt.figure(figsize=(FIG_WIDTH,5))
labels_train_e_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Train E\nMean frequency of digits per class: {}, Standard Deviation: {:.4f}'.format(labels_train_e_sr_vc.mean(),labels_train_e_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

Test E

In [161]:
df_test_e=pd.read_csv(path_label_test_e)
df_test_e=df_test_e.set_index('filename')
df_test_e.head()
Out[161]:
original filename districtid institutionid gender age datestamp scanid digit database name original database name
filename
e00000.png 01_0001_1_14_1016_1508_54.png 1 1 1 14 1016 1508 3 BanglaLekha-Isolated testing-e
e00001.png 02_0002_0_22_1016_1610_52.png 2 2 0 22 1016 1610 1 BanglaLekha-Isolated testing-e
e00002.png 02_0002_0_24_1016_1836_54.png 2 2 0 24 1016 1836 3 BanglaLekha-Isolated testing-e
e00003.png 02_0002_0_25_1016_1716_59.png 2 2 0 25 1016 1716 8 BanglaLekha-Isolated testing-e
e00004.png 01_0001_1_17_0916_0010_58.png 1 1 1 17 916 10 7 BanglaLekha-Isolated testing-e
In [165]:
labels_test_e_sr=pd.Series([df_test_e.loc[get_key(path)]['digit'] for path in paths_test_e])
In [267]:
labels_test_e_sr_vc=labels_test_e_sr.value_counts()
In [166]:
plt.figure(figsize=(FIG_WIDTH,5))
labels_test_e_sr_vc.plot(kind='bar')
plt.xticks(rotation='horizontal',fontsize=FS_TICKS)
plt.yticks(fontsize=FS_TICKS)
plt.xlabel('Digits', fontsize=FS_AXIS_LABEL)
plt.ylabel('Frequency', fontsize=FS_AXIS_LABEL)
plt.title('Test E\nMean frequency of digits per class: {}, Standard Deviation: {}'.format(labels_test_e_sr_vc.mean(),labels_test_e_sr_vc.std()),
         fontsize=FS_TITLE)
plt.show()

Summary

The classes in each dataset are well balanced. The images in dataset A and B should be further processed to extract a more focused crop of the digit. Although the shapes of dataset B and E varies a lot, they have a square or nearly square aspect ratio. Therefore, if resizing to a fixed to shape is necessary (e.g., feeding the images into a fixed input neural network), it should not create much distortion in the digit shape. Below is a summary statistic of the datasets.

In [268]:
df_summary=pd.DataFrame(data={
    'Avg. samples per digits (Train)':[labels_train_a_sr_vc.mean(),labels_train_b_sr_vc.mean(),labels_train_e_sr_vc.mean()],
    'Avg. samples per digits (Test)':[labels_test_a_sr_vc.mean(),labels_test_b_sr_vc.mean(),labels_test_e_sr_vc.mean()],
    'Train Samples':[labels_train_a_sr_vc.sum(),labels_train_b_sr_vc.sum(),labels_train_e_sr_vc.sum()],
    'Test Samples':[labels_test_a_sr_vc.sum(),labels_test_b_sr_vc.sum(),labels_test_e_sr_vc.sum()],
    'Total':[labels_train_a_sr_vc.sum()+labels_test_a_sr_vc.sum(),
             labels_train_b_sr_vc.sum()+labels_test_b_sr_vc.sum(),
             labels_train_e_sr_vc.sum()+labels_test_e_sr_vc.sum()]},
                        index=['A','B','E'],
                        columns=['Avg. samples per digits (Train)',
                                 'Avg. samples per digits (Test)', 
                                 'Train Samples',
                                 'Test Samples',
                                 'Total']                    
            )
df_summary
Out[268]:
Avg. samples per digits (Train) Avg. samples per digits (Test) Train Samples Test Samples Total
A 1970.2 348.9 19702 3489 23191
B 35.9 6.9 359 69 428
E 1677.7 297.1 16777 2971 19748

There are additional information in the .csv file of dataset E. We are going to focus on that in the next notebook.