死宅一枚。爬取5000张二次元妹子的图片，生成了头图。
接下来看看怎么实现的：
- 使用 Scrapy 框架爬取5000张二次元图
- 使用 opencv 批量格式化图片
- 将图片按照RGB值的均方根排序，实现效果
一、安装环境
1.安装 Scrapy 爬虫框架
pip install Scrapy
windows 安装可以点击此处
2. 推荐使用 wheel 来安装 opencv 点击此处
3.安装 numpy 科学计算库
pip install numpy
4. 初始化一个 Scrapy 项目 acg
scrapy startproject acg
二、爬取图片
以下代码主要实现操作：
- 中间裁剪
- 统一大小
- 下载图片
- 重复抓取
/image.py
import scrapy
import urllib.request,urllib.parse
import numpy as np
import cv2
class acgimages(scrapy.Spider):
"""docstring for acgimages"""
name = 'images'
start_urls = [
"http://m.52dmtp.com/tupiandaquan/index_2.html"
]
count = 1
page = 2
def parse(self,response):
def imageSave(item,path):
try:
maxsize = 512
res = urllib.request.urlopen(item).read()
image = np.asarray(bytearray(res),dtype="uint8")
image = cv2.imdecode(image,cv2.IMREAD_COLOR)
height,width = image.shape[:2]
if height > width:
scalefactor = (maxsize*1.0) / width
res = cv2.resize(image,(int(width * scalefactor),(int(height * scalefactor))),interpolation = cv2.INTER_CUBIC)
cutImage = res[0:maxsize,0:maxsize]
if width >= height:
scalefactor = (maxsize*1.0) / height
res = cv2.resize(image,(int(width * scalefactor), int(height*scalefactor)), interpolation = cv2.INTER_CUBIC)
center_x = int(round(width*scalefactor*0.5))
cutImage = res[0:maxsize,int(center_x - maxsize/2):int(center_x + maxsize/2)]
cv2.imwrite(path,cutImage)
print('image is save in ' + path)
except:
print('image save error')
image_url = response.xpath("//div[@class='grid-wrap']//img/@src").extract()
for item in image_url:
item = item.split('?')[0]
item = urllib.parse.quote(item,safe='/:?=.')
if 'jpg' in item:
self.count = self.count + 1
path = 'img/'+ str(self.count) + ".jpg"
imageSave(item,path)
if 'png' in item:
self.count = self.count + 1
path = 'img/'+ str(self.count) + ".png"
imageSave(item,path)
if self.page < 1180:
self.page = self.page + 1
next_url = "http://m.52dmtp.com/tupiandaquan/index_%d.html"%self.page
yield scrapy.Request(next_url,callback = self.parse)
三、生成图片
软件下载地址：Welcome | FMEdda | Foto-Mosaik-Edda
1.创建一个图片数据库
2.导入图片
3.导入过程中
4.创建一个马赛克风格图片
5.打开原图
6.选择已经上传的数据库
7.生成完成
简单高效的平民玩法到此结束。
2017年12月10日 本来今天开开心心。被迫更新
一、安装依赖
1.安装图像处理库 pillow
pip install pillow
二、使用 python 做到马赛克拼图效果
写在前面，有人私信说 hsv（颜色空间）做效果会更好。大家可以尝试一下，这里的话只是对rgb的对比。主要效果差上面提到的程序一点，答主也在尽可能的优化代码，有更好的方法，欢迎私信评论。
核心部分：求像素平均值，尝试过了方差和均方根，个人对整体优化并不明显。
def get_avg_color(img):
width, height = img.size
pixels = img.load()
data = []
for x in range(width):
for y in range(height):
content = pixels[x, y]
data.append(content)
r = 0
g = 0
b = 0
count = 0
for x in range(len(data)):
r += data[x][0]
g += data[x][1]
b += data[x][2]
count += 1
rAvg = r / count
gAvg = g / count
bAvg = b / count
return (rAvg, gAvg, bAvg)
核心部分：对比图片块颜色与马赛克块平均颜色差值。很重要的一步就是，求出近似值数组后，取一个随机数，有效的解决马赛克块重复问题（不是不重复使用，因为这样需要的图片集会很大）。
def find_closiest(color, list_colors):
diff = 10000
cur_closer = []
arr_len = 0
for cur_color in list_colors:
n_diff = abs(color[0] - cur_color[0]) + abs(color[1] - cur_color[1]) + abs(color[2] -cur_color[2])
if n_diff < diff:
diff = n_diff
if len(cur_closer) <= 5:
cur_closer.append(cur_color)
else:
cur_closer[arr_len] = cur_color
arr_len += 1
if arr_len > 5:
arr_len = 0
index = random.randint(0,len(cur_closer) - 1)
return cur_closer[index]
核心部分： 图片分割并且按图片集名字对比就近的图片
def make_puzzle(img, color_list):
width, height = img.size
print("Width = {}, Height = {}".format(width,height))
background = Image.new('RGB', img.size, (255,255,255))
total_images = 0
for y1 in range(0, heigth, SLICE_SIZE):
for x1 in range(0, width, SLICE_SIZE):
y2 = y1 + SLICE_SIZE
x2 = x1 + SLICE_SIZE
new_img = img.crop((x1, y1, x2, y2))
color = get_avg_color(new_img)
close_img_name = find_closiest(color, color_list)
close_img_name = OUT_DIR + str(close_img_name) + '.jpg'
paste_img = Image.open(close_img_name)
total_images += 1
print("%s images \r" % total_images),
background.paste(paste_img, (x1, y1))
return background
全部代码 /main.py
import os
from PIL import Image,ImageOps
import argparse
import time
# 多进程，
from multiprocessing import Pool
import random
import math
# 少年，你不试试看hsv吗，据说效果更好
from colorsys import rgb_to_hsv
SLICE_SIZE = 60 # 马赛克大小
OUT_SIZE = 5000 # 图片大小（注意：要做到清晰，必须要大）
IN_DIR = "database/"
OUT_DIR = "output/"
def get_avg_color(img):
width, height = img.size
pixels = img.load()
data = []
for x in range(width):
for y in range(height):
content = pixels[x, y]
data.append(content)
r = 0
g = 0
b = 0
count = 0
for x in range(len(data)):
r += data[x][0]
g += data[x][1]
b += data[x][2]
count += 1
rAvg = r / count
gAvg = g / count
bAvg = b / count
return (rAvg, gAvg, bAvg)
def find_closiest(color, list_colors):
diff = 10000
cur_closer = []
arr_len = 0
for cur_color in list_colors:
n_diff = abs(color[0] - cur_color[0]) + abs(color[1] - cur_color[1]) + abs(color[2] -cur_color[2])
if n_diff < diff:
diff = n_diff
if len(cur_closer) <= 5:
cur_closer.append(cur_color)
else:
cur_closer[arr_len] = cur_color
arr_len += 1
if arr_len > 5:
arr_len = 0
index = random.randint(0,len(cur_closer) - 1)
return cur_closer[index]
def make_puzzle(img,color_list):
width, height = img.size
print("Width = {}, Height = {}".format(width,heigth))
background = Image.new('RGB', img.size, (255,255,255))
total_images = 0
for y1 in range(0, heigth, SLICE_SIZE):
for x1 in range(0, width, SLICE_SIZE):
y2 = y1 + SLICE_SIZE
x2 = x1 + SLICE_SIZE
new_img = img.crop((x1, y1, x2, y2))
color = get_avg_color(new_img)
close_img_name = find_closiest(color, color_list)
close_img_name = OUT_DIR + str(close_img_name) + '.jpg'
paste_img = Image.open(close_img_name)
total_images += 1
print("%s images \r" % total_images),
background.paste(paste_img, (x1, y1))
return background
def get_image_paths():
paths = []
for file_ in os.listdir(IN_DIR):
paths.append(IN_DIR + file_)
print("一共找到了%s" % len(paths) + "张图片")
return paths
def resize_pic(in_name,size):
img = Image.open(in_name)
img = ImageOps.fit(img, (size, size), Image.ANTIALIAS)
return img
def convert_image(path):
img = resize_pic(path,SLICE_SIZE)
color = get_avg_color(img)
img.save(str(OUT_DIR) + str(color) + ".jpg")
def convert_all_images():
paths = get_image_paths()
pool = Pool()
pool.map(convert_image, paths)
pool.close()
pool.join()
def read_img_db():
img_db = []
for file_ in os.listdir(OUT_DIR):
print(file_)
if file_ == 'None.jpg':
pass
else:
file_ = file_.split('.jpg')[0]
file_ = tuple(map(float, file_[1:-1].split(',')))
img_db.append(file_)
return img_db
if __name__ == '__main__':
parse = argparse.ArgumentParser()
parse.add_argument("-i",'--input',required=True,help='input image')
parse.add_argument("-d", "--db", type=str, required=True,help="source database")
parse.add_argument("-o", "--output", type=str, required=True,help="out directory")
parse.add_argument("-is",'--inputSize',type=str, required=False,help="inputSize")
parse.add_argument("-os",'--outputSize',type=str, required=False,help="outputSize")
args = parse.parse_args()
start_time = time.time()
args = parse.parse_args()
image = args.input
if args.db:
IN_DIR = args.db
if args.output:
OUT_DIR= args.output
if args.inputSize:
SLICE_SIZE = args.inputSize
if args.outputSize:
OUT_SIZE = args.outputSize
img = resize_pic(image,OUT_SIZE)
convert_all_images()
list_of_imgs = read_img_db()
# 为了添加图片信息量，合并两张图片
out = make_puzzle(img, list_of_imgs)
img = Image.blend(out, img, 0.5)
img.save('out.jpg')
print("耗时: %s" % (time.time() - start_time))
print("已完成")
命令行输入
python main.py -i test.jpg（你的图片） -d D:/acg/img/（你的图片集合） -o output/（你的马赛克图片输出位置）
当然，你还可以给你的 background 滤色
上传一张结果图（根据喜欢，精度可以更高），知乎限制5m了
非洲小鸟，黑白对比度处理任然是败笔，建议还是 hsv
这个模型更适合颜色适配。
后续将会发布优化版本。
作者：今晚的风儿很喧嚣
