图像处理入门100题（一）

图像处理入门100题（一）

本篇记录了自己写的GIthub上的ImageProcessing 100 Wen的问题1-10的答案，注释里包含了一些自己的感悟。为了方便，注释是用英文写的。

问题序号	题目
1	通道交换
2	灰度化
3	二值化
4	大津二值化算法（最佳阈值）
5	HSV与RGB的变换
6	减色处理
7	平均池化
8	最大池化
9	高斯滤波
10	中值滤波

Python答案如下，结尾包含了问题的验证。

import cv2.cv2
import numpy as np


# Question One
# change channels
# from RGB to BGR
# please note that the images are stored by BGR
# Answer

def RGB2BGR(img):
    out = img.copy()
    out[:, :] = out[:, :, (2, 1, 0)]  # from 0:B 1:G 2:R <= 2:R 1:G 0:B
    return out


# Question Two
# gray scale of images
# Y = 0.2126 * R + 0.7152 * G + 0.0722 * B
# Answer

def RGB2GRAY(img):
    out = 0.2126 * img[:, :, 2] + 0.7152 * img[:, :, 1] + 0.0722 * img[:, :, 0]
    out = out.astype(np.uint8)
    return out


# Question Three
# Binary image
# threshold: if y < t then y := 0 else y := 1
# Answer

def BINARY(img, th=128):
    out = RGB2GRAY(img)

    out[out < th] = 0  # please not this usage, out < th return a matrix which has the same dimension with out
    out[out >= th] = 255  # out < th is consisted of True and False

    return out


# Question Four
# find best threshold for binary image
# Otsu's method:
# there are two classes: class0 that is less than threshold and class1 that is greater and equal than threshold
# w0: the rate of number of pixels of class0, the same as w1
# M0: the mean value of pixel value of class0, the same as M1
# objective function: make Sb^2 = w0 * w1 * (M0 - M1)^2 biggest
# Answer

def BESTTH(img):
    best_t = 0
    max_sb = 0
    H = img.shape[0]
    W = img.shape[1]
    gray = RGB2GRAY(img)

    for t in range(1, 255):

        class0 = gray[gray < t]
        class1 = gray[gray >= t]

        w0 = len(class0) / (H * W)
        w1 = len(class1) / (H * W)

        M0 = np.mean(class0) if len(class0) > 0 else 0
        M1 = np.mean(class1) if len(class1) > 0 else 0

        sb = w0 * w1 * np.square(M0 - M1)

        if sb > max_sb:
            best_t = t
            max_sb = sb

    return best_t


# Question Five
# HSV to RGB, RGB to HSV
# Answer
def RGB2HSV(img):
    hsv = np.zeros_like(img, dtype=np.float)

    img = img / 255.0  # RGB in [0, 1]
    max = np.max(img, axis=2)  # max = max(R, G, B)
    min = np.min