Image enhancement based on deep learning

Image enhancement ： Enhance the useful information in the image , It can be a distortion process , The purpose is to improve the visual effect of the image .

In the context of deep learning , Image enhancement is to enrich the sample features of the image , Make model generalization more powerful . Common ways are ： rotate 、 tailoring 、 Horizontal or vertical rotation 、 Change the brightness 、 Gauss filtering 、 blurred .

The image enhancement technology involved in this paper is through opencv Realization .

import random
import pickle
import os
import cv2
import time
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image, ImageDraw
import os.path as osp


def img_show(img_raw):
    
    cv2.namedWindow("Image")
    cv2.imshow('Image',img_raw)
    cv2.waitKey (0)

    cv2.destroyAllWindows()
    return img_raw

def rota_img(raw_img,degree):
    """ opencv Realization   Random arbitrary angle of the image (0~360 degree ) rotate  degree: Rotation Angle  :return:  Image after rotation  """
    h, w, c = raw_img.shape#(768,480,3)
    center_x = w // 2
    center_y = h // 2
    center = (center_x, center_y)
    M= cv2.getRotationMatrix2D(center, -degree, 1.)
    image = cv2.warpAffine(raw_img, M, (w, h))
    #mode=Image.BICUBIC
    #random_angle = np.random.randint(1, 360)
    #img_rota=raw_img.rotate(random_angle, mode)
    return image

def reflect_img(raw_img):
    '''  Affine transformation ： p1: Three points in the original picture  p2: Three points after the radiation transformation   The three points correspond to the upper left corner 、 Upper right corner 、 The lower left corner  '''
    h, w, c = raw_img.shape#(768,480,3)
    p1 = np.float32([[0,0], [w-1,0], [0,h-1]])
    p2 = np.float32([[0,h*0.3], [w*0.8,h*0.2], [w*0.15,h*0.7]])
    M = cv2.getAffineTransform(p1, p2)
    img = cv2.warpAffine(raw_img, M, (w,h))
    return img
    
def random_flip(raw_img,min_ratio=0.6, max_ratio=1.0):
    '''  The picture to achieve random 0.6~1.0 Ratio clipping , Keep the aspect ratio unchanged  '''
    h, w, c = raw_img.shape
    ratio = random.random()
    scale = min_ratio + ratio * (max_ratio - min_ratio)
    new_h = int(h*scale)
    new_w = int(w*scale)
    y = np.random.randint(0, h - new_h)
    x = np.random.randint(0, w - new_w)
    image = raw_img[y:y+new_h, x:x+new_w, :]
    return image
    
def random_resize_img(raw_img,size=128,bias=5):
    '''  Press size Size clipping  '''
    h, w, c = raw_img.shape
    size_min = min(w,h)
    size_max = max(w,h)
    min_size = size + np.random.randint(1, bias)
    scale = float(min_size) / float(size_min)
    image = cv2.resize(raw_img, dsize=(0, 0), fx=scale, fy=scale)
    return image
    
def random_crop_and_resize(image, size=128):
    '''  Cut to size  '''

    image = random_resize_img(image)

    h, w = image.shape[:2]

    y = np.random.randint(0, h-size)
    x = np.random.randint(0, w-size)

    image = image[y:y+size, x:x+size, :]

    return image
   
    
def img_noise(image,flag):
    ''' Noise() '''
    if flag==1:
        # median filtering 
        img_result=cv2.medianBlur(image,5)
        font=cv2.FONT_HERSHEY_SIMPLEX
    # Mean filtering 
    if flag==2:
        img_result=cv2.blur(image,(5,5))
        
    # Gauss filtering  
    if flag==3:
        img_result=cv2.GaussianBlur(image,(7,7),0)
 
    # Gaussian bilateral filtering 
    if flag==4:
        img_result=cv2.bilateralFilter(image,40,75,75)

    return img_result