Сделай высоко контрастный портрет крупным планом м...

python
from PIL import Image, ImageEnhance, ImageFilter, ImageOps
img = Image.open('/mnt/data/IMG_1988.jpeg')
img.size, img.mode


((960, 1280), 'RGB')

python
import cv2, numpy as np, os, math, random, pathlib, subprocess, textwrap, tempfile, imageio.v2 as imageio
import numpy as np
img_bgr = cv2.imread('/mnt/data/IMG_1988.jpeg')
h,w = img_bgr.shape[:2]
h,w


(1280, 960)

python
# create composite on black
fg = cv2.bitwise_and(img_bgr, img_bgr, mask=mask2)
black = np.zeros_like(img_bgr)
comp = np.where(mask2[:,:,None]==255, img_bgr, black)

preview_path = '/mnt/data/preview_comp.png'
cv2.imwrite(preview_path, comp)
preview_path


'/mnt/data/preview_comp.png'

python
# create composite on black
fg = cv2.bitwise_and(img_bgr, img_bgr, mask=mask2)
black = np.zeros_like(img_bgr)
comp = np.where(mask2[:,:,None]==255, img_bgr, black)

preview_path = '/mnt/data/preview_comp.png'
cv2.imwrite(preview_path, comp)
preview_path


'/mnt/data/preview_comp.png'



<<ImageDisplayed>>

python
def grabcut_rect(rect, iter=7):
    mask = np.zeros((h,w), np.uint8)
    bgdModel = np.zeros((1,65), np.float64)
    fgdModel = np.zeros((1,65), np.float64)
    cv2.grabCut(img_bgr, mask, rect, bgdModel, fgdModel, iter, cv2.GC_INIT_WITH_RECT)
    mask2 = np.where((mask==cv2.GC_FGD) | (mask==cv2.GC_PR_FGD), 255, 0).astype('uint8')
    return mask2

rect2 = (250, 730, 690, 540)
mask2b = grabcut_rect(rect2, iter=7)
comp2 = np.where(mask2b[:,:,None]==255, img_bgr, np.zeros_like(img_bgr))
cv2.imwrite('/mnt/data/preview_comp2.png', comp2)


True

python
def grabcut_rect(rect, iter=7):
    mask = np.zeros((h,w), np.uint8)
    bgdModel = np.zeros((1,65), np.float64)
    fgdModel = np.zeros((1,65), np.float64)
    cv2.grabCut(img_bgr, mask, rect, bgdModel, fgdModel, iter, cv2.GC_INIT_WITH_RECT)
    mask2 = np.where((mask==cv2.GC_FGD) | (mask==cv2.GC_PR_FGD), 255, 0).astype('uint8')
    return mask2

rect2 = (250, 730, 690, 540)
mask2b = grabcut_rect(rect2, iter=7)
comp2 = np.where(mask2b[:,:,None]==255, img_bgr, np.zeros_like(img_bgr))
cv2.imwrite('/mnt/data/preview_comp2.png', comp2)


True



<<ImageDisplayed>>

python
rect_front = (420, 690, 520, 580)
mask_front = grabcut_rect(rect_front, iter=7)
comp_front = np.where(mask_front[:,:,None]==255, img_bgr, np.zeros_like(img_bgr))
cv2.imwrite('/mnt/data/preview_front.png', comp_front)


True

python
rect_front = (420, 690, 520, 580)
mask_front = grabcut_rect(rect_front, iter=7)
comp_front = np.where(mask_front[:,:,None]==255, img_bgr, np.zeros_like(img_bgr))
cv2.imwrite('/mnt/data/preview_front.png', comp_front)


True



<<ImageDisplayed>>

python
import importlib, pkgutil, sys, subprocess, os, textwrap, json, math, statistics
"rembg" in [m.name for m in pkgutil.iter_modules()]


False

python
import cv2, numpy as np

mask = mask2b.copy()
gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5,5), 0)
edges = cv2.Canny(blur, 50, 150)
# distance transform needs binary image where non-zero are foreground (for distance to zero)
inv_edges = (edges==0).astype(np.uint8)  # 1 where no edge
dist = cv2.distanceTransform(inv_edges, cv2.DIST_L2, 5)
# normalize
dist_norm = dist / (dist.max() + 1e-6)
gray_norm = gray.astype(np.float32)/255.0
score = 0.6*(1.0 - dist_norm) + 0.4*(1.0 - gray_norm)  # edge proximity + darkness
# apply only within initial mask
score_masked = score * (mask/255.0)
#

True

python
import cv2, numpy as np

mask = mask2b.copy()
gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5,5), 0)
edges = cv2.Canny(blur, 50, 150)
# distance transform needs binary image where non-zero are foreground (for distance to zero)
inv_edges = (edges==0).astype(np.uint8)  # 1 where no edge
dist = cv2.distanceTransform(inv_edges, cv2.DIST_L2, 5)
# normalize
dist_norm = dist / (dist.max() + 1e-6)
gray_norm = gray.astype(np.float32)/255.0
score = 0.6*(1.0 - dist_norm) + 0.4*(1.0 - gray_norm)  # edge proximity + darkness
# apply only within initial mask
score_masked = score * (mask/255.0)
#

True



<<ImageDisplayed>>

python
for thr in [0.26,0.30,0.34]:
    refined = np.where((mask==255) & (score > thr), 255, 0).astype('uint8')
    refined2 = cv2.morphologyEx(refined, cv2.MORPH_CLOSE, kernel, iterations=2)
    refined2 = cv2.morphologyEx(refined2, cv2.MORPH_OPEN, kernel, iterations=1)
    comp_ref = np.where(refined2[:,:,None]==255, img_bgr, np.zeros_like(img_bgr))
    cv2.imwrite(f'/mnt/data/preview_thr_{int(thr*100)}.png', comp_ref)


<<ImageDisplayed>>

python
img_mod = img_bgr.copy()
img_mod[soil] = (0,0,0)  # black
comp_soil = np.where(mask_obj[:,:,None]==255, img_mod, np.zeros_like(img_mod))
cv2.imwrite('/mnt/data/preview_soil_black.png', comp_soil)


True