名字

没想好

马赛克图片拼接

| Comments

以前就想做一个马赛克合成的图片了，最近不务正业，就写了一个 方法很简单:

  • 抓取一组图片的并缩放到统一大小
  • 计算每张图片的rgb值的中值
  • 所有的图片中值生成kd-tree
  • 加载目标图片，并缩放到想要的结果图片的大小，然后创建结果图片(np.zeros)
  • 每一小格结果图片的大小和抓取的小图大小一样,计算目标图片每小格的rgb 中值
  • 查找最想进的小图，并用小图的rgb值填充到结果图片里
  • 生成结果图片

马赛克图片生成代码:

import Image
import os
import numpy as np
from scipy import spatial

#calculate a mean rgb of an image
def meanRGB(img):
    rgb = np.array(img.getdata())
        mean_r = np.mean(rgb[:,0])
        mean_g = np.mean(rgb[:,1]) 
        mean_b = np.mean(rgb[:,2])
        return [mean_r, mean_g, mean_b]

if __name__ == "__main__":
    #grid size
    mosaic_img_size = (20, 20)
    #size of target image to generated
    target_img_size = (1000, 1000)
    target_img = Image.open("./VanGogh-starry_night_ballance1.jpg")
    target_img = target_img.resize(target_img_size)
    #grid images directory
    directory = "./icons/"
    listing = os.listdir(directory)
    img_list = []
    color_list = []
    #load images and calculate mean rgb
    for infile in listing:
        if infile.endswith("jpg"):
            img = Image.open(directory+infile)
            img = img.resize(mosaic_img_size, Image.ANTIALIAS)
            img_list.append(img)
            color_list.append(meanRGB(img))

    #create kdt for future pnn query
    data = np.array(color_list)
    tree = spatial.KDTree(data, leafsize=5)

    count_x = 0
    count_y = 0
    new_image = np.zeros((target_img_size[0],target_img_size[1],3), dtype=np.uint8)
    target_img_data = np.reshape(np.array(target_img.getdata()), (target_img_size[0],target_img_size[1],3))
    length = mosaic_img_size[0]*mosaic_img_size[1]

    #calculate mean rgb of each grid in the target image, search a nearest neighbor and fill the new image with
    #the rgb values of the grid image
    while count_x+mosaic_img_size[0] < target_img_size[0]:
        count_y = 0
        while count_y+mosaic_img_size[1] < target_img_size[1]:
            mean_r = np.mean(target_img_data[count_x:count_x+mosaic_img_size[0],count_y:count_y+mosaic_img_size[1],0])
            mean_g = np.mean(target_img_data[count_x:count_x+mosaic_img_size[0],count_y:count_y+mosaic_img_size[1],1])
            mean_b = np.mean(target_img_data[count_x:count_x+mosaic_img_size[0],count_y:count_y+mosaic_img_size[1],2])
            img_index = tree.query(np.array([mean_r,mean_g,mean_b]))[1]
            new_image[count_x:count_x+mosaic_img_size[0],count_y:count_y+mosaic_img_size[1],:] = np.reshape(np.array(img_list[img_index].getdata()), (mosaic_img_size[0],mosaic_img_size[1],3))

            count_y += mosaic_img_size[1]
        count_x += mosaic_img_size[0]
    
    #image created
    im = Image.fromarray(new_image)
    im.show()
    im.save("./1VanGogh-starry_night_ballance1.jpg")

里面有些小bug，但不影响使用，譬如最右最下有一圈黑色的，是代码中某个逻辑没写好，没有填充进去，但是不管它了- -(貌似把小于号改成小于等于就可以了)

原图: 效果: 768的向量空间长度，图片量一定要大一些，要不然的结果会是，颜色很单一，拼出来的马赛克效果不好。自己的友邻太少了，可能结果会不理想，所以这里我抓了豆瓣活动上的豆友，570多位。。

Comments