PATCHMATCH算法解析

代码复杂度O(M*m*t*logM),M为图像的像素个数，m为patch size ,t为迭代次数

import numpy as np
from PIL import Image
import time

def cal_distance(a, b, A_padding, B, p_size):
    p = p_size // 2
    patch_a = A_padding[a[0]:a[0]+p_size, a[1]:a[1]+p_size, :]
    patch_b = B[b[0]-p:b[0]+p+1, b[1]-p:b[1]+p+1, :]
    temp = patch_b - patch_a
    num = np.sum(1 - np.int32(np.isnan(temp)))
    dist = np.sum(np.square(np.nan_to_num(temp))) / num
    return dist

def reconstruction(f, A, B):
    A_h = np.size(A, 0)
    A_w = np.size(A, 1)
    temp = np.zeros_like(A)
    for i in range(A_h):
        for j in range(A_w):
            temp[i, j, :] = B[f[i, j][0], f[i, j][1], :]
    Image.fromarray(temp).show()


def initialization(A, B, p_size):
    A_h = np.size(A, 0)
    A_w = np.size(A, 1)
    B_h = np.size(B, 0)
    B_w = np.size(B, 1)
    p = p_size // 2
    random_B_r = np.random.randint(p, B_h-p, [A_h, A_w])
    random_B_c = np.random.randint(p, B_w-p, [A_h, A_w])
    A_padding = np.ones([A_h+p*2, A_w+p*2, 3]) * np.nan
    A_padding[p:A_h+p, p:A_w+p, :] = A
    f = np.zeros([A_h, A_w], dtype=object)
    dist = np.zeros([A_h, A_w])
    for i in range(A_h):
        for j in range(A_w):
            a = np.array([i, j])
            b = np.array([random_B_r[i, j], random_B_c[i, j]], dtype=np.int32)
            f[i, j] = b
            dist[i, j] = cal_distance(a, b, A_padding, B, p_size)
    return f, dist, A_padding

def propagation(f, a, dist, A_padding, B, p_size, is_odd):
    A_h = np.size(A_padding, 0) - p_size + 1
    A_w = np.size(A_padding, 1) - p_size + 1
    x = a[0]
    y = a[1]
    if is_odd:#向左上方传播
        d_left = dist[max(x-1, 0), y]
        d_up = dist[x, max(y-1, 0)]
        d_current = dist[x, y]
        idx = np.argmin(np.array([d_current, d_left, d_up]))
        if idx == 1:
            f[x, y] = f[max(x - 1, 0), y]
            dist[x, y] = cal_distance(a, f[x, y], A_padding, B, p_size)
        if idx == 2:
            f[x, y] = f[x, max(y - 1, 0)]
            dist[x, y] = cal_distance(a, f[x, y], A_padding, B, p_size)
    else:#向右下方传播
        d_right = dist[min(x + 1, A_h-1), y]
        d_down = dist[x, min(y + 1, A_w-1)]
        d_current = dist[x, y]
        idx = np.argmin(np.array([d_current, d_right, d_down]))
        if idx == 1:
            f[x, y] = f[min(x + 1, A_h-1), y]
            dist[x, y] = cal_distance(a, f[x, y], A_padding, B, p_size)
        if idx == 2:
            f[x, y] = f[x, min(y + 1, A_w-1)]
            dist[x, y] = cal_distance(a, f[x, y], A_padding, B, p_size)

def random_search(f, a, dist, A_padding, B, p_size, alpha=0.5):
    x = a[0]
    y = a[1]
    B_h = np.size(B, 0)
    B_w = np.size(B, 1)
    p = p_size // 2
    i = 4
    search_h = B_h * alpha ** i
    search_w = B_w * alpha ** i
    b_x = f[x, y][0]
    b_y = f[x, y][1]
    while search_h > 1 and search_w > 1:
        search_min_r = max(b_x - search_h, p)
        search_max_r = min(b_x + search_h, B_h-p)#不能越界
        random_b_x = np.random.randint(search_min_r, search_max_r)#row
        search_min_c = max(b_y - search_w, p)
        search_max_c = min(b_y + search_w, B_w - p)
        random_b_y = np.random.randint(search_min_c, search_max_c)#col
        search_h = B_h * alpha ** i#用类似于二分法的方式不断缩小搜索范围
        search_w = B_w * alpha ** i
        b = np.array([random_b_x, random_b_y])
        d = cal_distance(a, b, A_padding, B, p_size)
        if d < dist[x, y]:#如果新的随机距离小于已有的距离，那么就更新
            dist[x, y] = d
            f[x, y] = b
        i += 1

def NNS(img, ref, p_size, itr):
    A_h = np.size(img, 0)
    A_w = np.size(img, 1)
    f, dist, img_padding = initialization(img, ref, p_size)#随机初始化A到B的匹配
    for itr in range(1, itr+1):
        if itr % 2 == 0:
            for i in range(A_h - 1, -1, -1):
                for j in range(A_w - 1, -1, -1):
                    a = np.array([i, j])
                    propagation(f, a, dist, img_padding, ref, p_size, False)#先根据周围像素的匹配情况来进行本像素的匹配
                    random_search(f, a, dist, img_padding, ref, p_size)#继续进行大范围的搜索
        else:
            for i in range(A_h):
                for j in range(A_w):
                    a = np.array([i, j])
                    propagation(f, a, dist, img_padding, ref, p_size, True)
                    random_search(f, a, dist, img_padding, ref, p_size)
        print("iteration: %d"%(itr))
    return f

if __name__ == "__main__":
    img = np.array(Image.open("001_0.png"))
    ref = np.array(Image.open("001_5.png"))
    p_size = 3#patch的长宽
    itr = 5#迭代次数
    start = time.time()
    f = NNS(img, ref, p_size, itr)
    end = time.time()
    print(end - start)
    reconstruction(f, img, ref)