图像分类实操--分析数据集

分析数据集的好处：能够清楚的认识到自己所以处理的内容是什么，目前存在的问题有哪些，以及采用何种办法来解决。

1)将数据集分成训练集和测试集

刚开始打算用PIL里面的Image.save来保存图像，但没有成功，所有采用了Opencv来读取图像和保存图像。

Code:

importos
import random
import cv2

def split_dataset(datapath):
    # input: data path
    # output: data split result includegenerate train image, test image
    files = os.listdir(datapath)                                   # all classes name
    for file in files:                                             # each class name
        class_path = os.path.join(datapath, file)                  # each class path
        image_sequence =os.listdir(class_path)                   # each class imagename
        each_class_number = len(image_sequence)                    # number of each class image
        train_number = int(each_class_number/2)                    # the ratio of train set
        train_image_sequence =random.sample(image_sequence, train_number)  #random sample to compose train set
        train_image_list = generate_image_path(class_path, train_image_sequence)  #generate train image full path
        generate_file(InputImagePathList=train_image_list, OutputImagePath='../AID50/train')    # generate trainimage file
        test_image_sequence = list(set(image_sequence).difference(set(train_image_sequence)))    #test image set
        test_image_list =generate_image_path(class_path, test_image_sequence)    #generate test image full path
        generate_file(InputImagePathList=test_image_list, OutputImagePath='../AID50/test')      # generate testimage file


def generate_image_path(RootPath, ImageNameSequence):
    # input : root path
    # input : image name sequence
    # output : return image path sequence
    FullImagePath = []
    for imagename in ImageNameSequence:
        fullimagepath =os.path.join(RootPath, imagename)
       FullImagePath.append(fullimagepath)
    return FullImagePath

def generate_file(InputImagePathList,OutputImagePath):
    # input: input image path
    # input: output image path
    # output: generate image file
    imagenumber = 0           # each class imageindex
    for imagepath in InputImagePathList:
        image = cv2.imread(imagepath)
        imagenumber = imagenumber + 1
        class_name = imagepath.split('\\')[1]   # from Input Image Path sparse class name
        ImageSavePath =os.path.join(OutputImagePath, class_name)
        if not os.path.exists(ImageSavePath):
            os.makedirs(ImageSavePath)
        imagename = ImageSavePath + '\\' + class_name + '_' + str(imagenumber) + '.jpg'
        cv2.imwrite(imagename, image)

def main():
    AID_Path = "../AID"
    split_dataset(datapath=AID_Path)

if __name__ == "__main__":
    main()

2)统计训练数据集中每个类别的数量

importos
import matplotlib.pyplot as plt
import copy
import numpy as np

def plot_classes_number(InputPath):
    files = os.listdir(InputPath)
    all_classes_number = len(files)  #the number of all classes
    class2number = {}
    for file in files:
        class_path =os.path.join(InputPath, file) # each class path
        picture_name_sequence =os.listdir(class_path)  # each class name list
        each_class_number = len(picture_name_sequence)  #each class number
        class2number[file] = each_class_number  #dict to store class number
    list_classname = []
    list_classnumber = []
    for file in files:
        list_classname.append(file)
       list_classnumber.append(class2number[file])
    plt.figure(figsize=(11, 4))
    width = 0.7
    x = np.arange(len(list_classname))
    y = np.array(list_classnumber)
    plt.bar(x, y, width, align='center')  # plot a bar
    plt.ylabel("Number")
    # plt.xlabel("Class name")
    # plt.title("Number to eachclass")
    # set x scale
    plt.xticks(x, list_classname, size='small', rotation=270)
    # set digital label
    for a, b in zip(x, y):
        plt.text(a, b + 1, '%.0f' % b, ha='center', va='bottom', fontsize=7)
    plt.show()

AID_Train_Path = "../AID50/train"
plot_classes_number(AID_Train_Path)

可以每个类别的数目是一样的，所以存在样本不平衡的问题。

3)人眼观察一下每个类别图像的具体内容

fromPIL import Image
import matplotlib.pyplot as plt
import os
import random

def acquire_sample_picture(PictureSequencePath, Number):
    # input: PictureSequencePath stand for allpictures of each class in an fold
    # input: Number is return sample pictures'number
    # output: return sample pictures'name list
    picture_name_sequence =os.listdir(PictureSequencePath) #each class name list
    relative_sample_picture_list =random.sample(picture_name_sequence, Number)
    absolutive_sample_picture_list = []
    for relative_sample_picture in relative_sample_picture_list:
        absolutive_sample_picture =os.path.join(PictureSequencePath, relative_sample_picture)
       absolutive_sample_picture_list.append(absolutive_sample_picture)
    return absolutive_sample_picture_list

def PlotFigure(AID_Path, ColNumber, EachClassSampleNumber):
    # input: AID_Path; every picture inAID_Path/class name/picture name
    # ColNumber: Number of picturesdisplay in a col
    # EachClassSampleNumber: number of samplesfor each class
    # output: show a figure
    files = os.listdir(AID_Path)              # all classes name list
    Classes_Number = len(files)               # number of classes
    Col_Number = ColNumber                    # figure col number
    Sample_Number = EachClassSampleNumber     #each class sample number
    Row_Number =(Sample_Number*Classes_Number)/Col_Number # figure row number
    sample_picture_number = 0                 # figure image index
    for file in files:
        class_path =os.path.join(AID_Path, file)
        sample_picture_list =acquire_sample_picture(PictureSequencePath=class_path, Number=Sample_Number)
        for picturepath in sample_picture_list:
            image =Image.open(picturepath)
            sample_picture_number =sample_picture_number + 1  #col index
            plt.subplot(Row_Number, Col_Number, sample_picture_number)
            plt.imshow(image)
            plt.axis("off")                  # shut down axis
    plt.subplots_adjust(wspace=0, hspace=0)  #no gap between nearby images
    plt.show()

def main():
    AID_Path = "../AID50/train"
    PlotFigure(AID_Path, ColNumber=10, EachClassSampleNumber=2)

if __name__ == "__main__":
    main()

按行看，每类图像展示了两张，可以大致看出相同类别存在差异性，比如第一行的第五张和第六张图像属于相同的类别，但肉眼很难判断出是同一类；还有不同类别间的相似性，比如第二行的第六张和第七张，分属不同类别但具有很强的相似性。