Image Processing

題外話

GUI的生成(PyQt)

先用 Qt Designer 製作想要的模板 Reference

Main Window
Dialog
Widget

完成之後在cmd下指令(切到目標dir)

pyuic5 -x filename.ui -o filename.py

python filename.py

成功的話這裡會跳出視窗

寫一個控制的主程式 ( main.py )

from PyQt5 import QtWidgets
from PyQt5.QtWidgets import *
from PyQt5.QtCore import *
from PyQt5.QtGui import *

import GUI as ui  #GUI.py為剛剛上面用GUI.ui產生出來的

class Main(QDialog, ui.Ui_Dialog):
  def __init__(self):
    super().__init__()
    self.setupUi(self)

if __name__ == '__main__':
  import sys
  app = QtWidgets.QApplication(sys.argv)
  window = Main()
  window.show()
  sys.exit(app.exec_())

之後要控制UI上的物件，例如:按鈕觸發的程式，也是加在這個主程式

GPU 掛接

如果沒有執行掛接，預設用CPU去跑

安裝 Cuda & Cudnn Cuda官網 cuDNN官網 ( cuDNN Archive )
- 尋找對應的版本
- [2021/12] 發現有 bug ： CUDA-11.5跟 cuDNN v8.3.0 對不上，需要載 v8.2.1 (for CUDA-11.x)才可以正常運行
- 「系統」變數、「環境」參數加上
打勾應該會自動加上

(這次原本裝在D槽，但安裝完成後，自動跳到C槽了)
(不確定是否必要)安裝tensorflow-gpu

pip3 install tensorflow-gpu

測試

import tensorflow as tf
print(tf.test.is_gpu_available())

會出現警訊和結果

照著警訊上的建議

import tensorflow as tf
print(tf.config.list_physical_devices('GPU'))

若不為空矩陣 => 成功 (之前根本暴殄天物，放著強大GPU不用)

HW1_1

1.1 Load Image

利用OpenCV讀取圖片 Reference

import numpy as np
import cv2

img = cv2.imread('image.jpg')

#check:
type(img)
img.shape

# 以灰階的方式讀取圖檔
img_gray = cv2.imread('image.jpg', cv2.IMREAD_GRAYSCALE)

# 顯示圖片
cv2.imshow('視窗名稱', img)

# 關閉特定視窗
cv2.destroyWindow('視窗名稱')

# 按下任意鍵則關閉所有視窗
cv2.waitKey(0)
cv2.destroyAllWindows()

# 讓視窗可以自由縮放大小
cv2.namedWindow('視窗名稱', cv2.WINDOW_NORMAL)

遇到問題: 已經設定環境參數，但是還是無法import cv2

解決辦法:
改成直接用cmd下載opencv

pip install opencv-python

1.2 Color Separation

利用openCV內建function(cv2.split(img))

img = cv2.imread("Path")

(imgB,imgG,imgR) = cv2.split(img)

#補零是為了讓圖片不變灰階
zero = np.zeros(img.shape[:2],dtype = "uint8")
cv2.imshow("B channel",cv2.merge([imgB,zero,zero]))
cv2.imshow("G channel",cv2.merge([zero,imgG,zero]))
cv2.imshow("R channel",cv2.merge([zero,zero,imgR]))

cv2.imshow()

RBG圖片的shape = (height, width, 3) 灰階圖片的shape = (height, width)

1.3 Color Transformation

Method 1 用openCV內建function(cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)) formula: 0.07B+0.72G+0.21*R

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q1_Image\Sun.jpg')
imgI1 = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
cv2.imshow("Perceptually weighted formula",imgI1)

Method 2 自訂比例(ex. (B+G+R)/3)

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q1_Image\Sun.jpg')
(imgB,imgG,imgR) = cv2.split(img)
imgI2 = imgB[:]//3+imgG[:]//3+imgR[:]//3
cv2.imshow("Average weighted formula",imgI2)

1.4 Blending

用openCV內建function

cv2.createTrackbar cv2.createTrackbar('參數名稱','視窗名稱',最小值,最大值,回傳函數) 用回傳函數來更新
cv2.addWeighted cv2.addWeighted(image1,alpha1,image2,alpha2,gamma) gamma是附加值(效果類似亮度)

imgS = cv2.imread('Dataset_OpenCvDl_Hw1\Q1_Image\Dog_Strong.jpg')
imgW = cv2.imread('Dataset_OpenCvDl_Hw1\Q1_Image\Dog_Weak.jpg')
def bar(x):
    global barPos
    barPos = cv2.getTrackbarPos('Blend','DOG')
    imgDog = cv2.addWeighted(imgS,(255-barPos)/255,imgW,barPos/255,0)
    cv2.imshow('DOG',imgDog)
cv2.namedWindow('DOG')
cv2.createTrackbar('Blend','DOG',0,255,bar)
cv2.setTrackbarPos('Blend','DOG',127)

HW1_2

2.1 Gaussian Blur

Smoothing(reference) using Gaussian Blur 直接call內建function(cv2.GaussianBlur())

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q2_Image\Lenna_whiteNoise.jpg')
blur = cv2.GaussianBlur(img,(5,5),0)
cv2.imshow('whiteNoise',img)
cv2.imshow('Gaussian',blur)

2.2 Bilateral filter

Smoothing using Bilateral filter 直接call內建function(cv2.bilateralFilter())

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q2_Image\Lenna_whiteNoise.jpg')
blur = cv2.bilateralFilter(img,9,90,90)
cv2.imshow('whiteNoise',img)
cv2.imshow('Bilateral',blur)

2.3 Median filter

Smoothing using Median filter 直接call內建function(cv2.medianBlur())

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q2_Image\Lenna_pepperSalt.jpg')
blur3 = cv2.medianBlur(img,3)
blur5 = cv2.medianBlur(img,5)
cv2.imshow('pepperSalt',img)
cv2.imshow('Median(3x3)',blur3)
cv2.imshow('Median(5x5)',blur5)

HW1_3

3.1 Gaussian Blur

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q3_Image\House.jpg')
imgGray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
Ginit = [(-1,1),(0,1),(1,1),
         (-1,0),(0,0),(1,0),
         (-1,-1),(0,-1),(1,-1)]
Gfilter = Ginit
SUM = 0
for i in range(9):
    (x,y) = Ginit[i]
    Gfilter[i] = np.exp(-(x**2+y**2)/(2*0.5))/(2*np.pi*0.5) #取σ^2=0.5
    SUM += Gfilter[i]
Gnorm = np.array(Gfilter)/SUM
global Gblur
Gblur = np.array(imgGray).copy()
for i in range(1,Gblur.shape[0]-1):
    for j in range(1,Gblur.shape[1]-1):
        temp = 0
        for a in range(i-1,i+2):
            for b in range(j-1,j+2):
                temp += imgGray[a][b]*Gnorm[(a-i+1)*3+(b-j+1)]
        Gblur[i][j] = temp
cv2.imshow('Gaussian Blur',Gblur)

灰階圖套Gaussian發現，python的 '=' 用在matrix的時候屬於call by reference的概念

例如:

(assign) B=A ，若改變 B ， A會跟著變

所以要用 B=np.array(A).copy() 或 B=A[:]

3.2 Sobel X

Xfilter = [-1,0,1,
           -2,0,2,
           -1,0,1]
Xnorm = np.array(Xfilter)
global sobX
sobX = np.array(Gblur).copy()/255
for i in range(1,Gblur.shape[0]-1):
    for j in range(1,Gblur.shape[1]-1):
        temp = 0
        for a in range(i-1,i+2):
            for b in range(j-1,j+2):
                temp += Gblur[a][b]*Xnorm[(a-i+1)*3+(b-j+1)]
        sobX[i][j] = temp/255
cv2.imshow('Sobel X',sobX)

跟上面Gaussian差不多，但是 Sobel Matrice 的值介於 [0,1] (?)

3.3 Sobel Y

Yfilter = [-1,0,1,
           -2,0,2,
           -1,0,1]
Ynorm = np.array(Yfilter)
global sobY
sobY = np.array(Gblur).copy()/255
for i in range(1,Gblur.shape[0]-1):
    for j in range(1,Gblur.shape[1]-1):
        temp = 0
        for a in range(i-1,i+2):
            for b in range(j-1,j+2):
                temp += Gblur[a][b]*Ynorm[(a-i+1)*3+(b-j+1)]
        sobY[i][j] = temp/255
cv2.imshow('Sobel Y',sobY)

3.4 Magnitude

sobXnp = np.array(sobX)
sobYnp = np.array(sobY)
Mag = ((sobXnp**2+sobYnp**2))**(0.5)
cv2.imshow('Magnitude',Mag)

HW1_4

4.1 Resize

幾何變化1 幾何變化2

img = cv2.imread('Dataset_OpenCvDl_Hw1\Q4_Image\SQUARE-01.png')
reHeight = 256
reWidth = 256
global imgRe
imgRe = cv2.resize(img,(reHeight,reWidth),interpolation=cv2.INTER_AREA)
cv2.imshow('img_Resized',imgRe)

interpolation 選擇

4.2 Translation

global winX,winY
winX,winY=400,300

global Xmov,Ymov
Xmov,Ymov = 0,60

m = np.float32([[1,0,Xmov],[0,1,Ymov]])

global imgTr
imgTr = cv2.warpAffine(imgRe,m,(winX,winY))
cv2.imshow('img_Resized',imgRe)
cv2.imshow('img_Translation',imgTr)

warpAffine的第三個參數是視窗大小，如果設跟圖片一樣大，圖片會跑出框

變化矩陣 m 要用 float32 的格式

4.3 Rotation, Scaling

angle = 10
scale = 0.5

m = cv2.getRotationMatrix2D((imgTr.shape[0]/2,imgTr.shape[1]/2),10,0.5)

global imgRo
imgRo = cv2.warpAffine(imgTr,m,(winX,winY))
cv2.imshow('img_Rotation,Scaling',imgRo)

getRotationMatrix2D

第一個參數: 旋轉中心

第二個參數: 旋轉角度

第三個參數: 縮放比例

4.4 Shearing

oldL = [[50,50],
        [200,50],
        [50,200]]
newL = [[10,100],
        [200,50],
        [100,250]]

m = cv2.getAffineTransform(np.float32(oldL),np.float32(newL))

imgSh = cv2.warpAffine(imgRo,m,(winX,winY))
cv2.imshow('img_Shearing',imgSh)

三種變化用不同的function產生變化矩陣m

KYX2013/Image-Processing