Image Classification

This is recipe of image classification, also contains general knowledge about techniques.
이미지 인식에 대한 딥러닝 레시피입니다. 특정 기술들에 대한 설명도 포함합니다.

0. Import Libraries

I think in the image classification, there are special libraries which have almost always used.
Pandas를 이용하여 data를 load하고, numpy로 간단한 linear algebra들을 사용하며, seaborn과 matplotlib으로 visualization을 해줍니다.

# General Libraries
import pandas as pd
import seaborn as sn
import numpy as np
import matplotlib.pyplot as plt

scikit-learn, tensorflow, keras, pytorch 등등 여러가지 framework을 이용하여 Machine Learning과 Deep Learning을 할 수 있습니다.
정답은 없으며, 개인적인 생각으로는 keras가 building model이 매우 쉽습니다. pytorch는 python coding에 있어서 굉장히 친숙하게 느껴집니다.
필요한 라이브러리가 있다면, jupyter notebook 환경안에서 간단하게 설치합니다.

!pip install nameOfLibrary

import something framework, anything...

# scikit-learn style
import sklearn
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix

# keras style
from keras.models import Sequential
from keras.utils.np_utils import to_categorical
from keras.layers import Conv2D, MaxPool2D, Dense, Dropout, Flatten,
from keras.optimizers import RMSprop
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ReduceLROnPlateau

# tensorflow style
import tensorflow as tf

# pytorch style
import torch.nn as nn

1. Load dataset

Pandas를 이용하여 손쉽게 Load할 수 있습니다. 또는 computer vision을 이용하여 data를 사용할 수 있습니다.

# PANDAS style, import pandas as pd
# .csv
dataset = pd.read_csv("path/dataset.csv", axis=1)

# .txt
dataset = pd.read_csv("path/dataset.txt", usecols=())

# Computer Vision style, import cv2
def loadData():
    result = []
    datasets = ['trainPath', 'testPath']
   
    for dataset in datasets:
        directory = dataset
        images = []
        labels = []
        
        for folder in os.listdir(directory):
            currentLabel = folder
       
            for file in os.listdir(directory + "/" + folder):
                imgPath = directory + "/" + folder + "/" + file
                currentImg = cv2.imread(imgPath)
                currentImg = cv2.resize(currentImg, (120, 120))
                images.append(currentImg)
                labels.append(currentLabel)
                
        images, labels = shuffle(images, labels)
        result.append((images, labels))
            
    return result

(trainImages, trainLabels), (testImages, testLabels) = loadData()    

2. Visualize and Explore the images

Visualization can give you a insight, we don’t need to know about all of method to visualize,
but there are also specifically good techniques. Presentation이나 Discussion을 위하여 data를 시각화할 수 있어야 합니다. 용도에 맞게 가장 간단하고, 효과적인 방법 몇가지를 소개합니다.

(1) univariate analysis, i.e, target variable.
단독변수 분석을 할 때 입니다. 종속변수를 예를 들자면, 종속변수의 class 분포, class별로 몇개씩 있는지가 중요할 수 있습니다.

# count plot
sns.countplot(variable) # Draw a bar plot of each class of the feature.
variable.value_counts() # This show you a number of each class.

# distribution plot

# pie plot, in case of the categorical feature
classes = np.bincount(train)
plt.pie(classes, labels=['className1','className2','className3'], autopct='%1.2f%%')
plt.title('The title whatever you want')
plt.show()

때때로 image가 어떻게 생겼는지 직접 확인하고 싶을 때도 있습니다.

# 1 image
idx = np.random.randint(trainImages.shape[0])
plt.figure()
plt.imshow(trainImages[idx])
plt.title('This image is #{} '.format(trainLabels[idx]))
plt.show()

# many images
fig = plt.figure(figsize=(10,10))
fig.suptitle('Sample images of the dataset', fontsize=12)
for i in range(9):
    plt.subplot(3, 3, i+1)
    plt.xticks([])
    plt.yticks([])
    plt.imshow(trainImages[i])
    plt.xlabel(train_labels[i])
plt.show()

3. Preprocessing images

Scaling

4. Modeling

5. Image Data Augmentation

6. Hyper-parameter optimization

7. Classify the images