meniki

[boxosh6162-ship-it]

import cv2
import numpy as np
import os

Load the image

image_path = 'olamiz 03.jpg'
img = cv2.imread(image_path)

Since we don't have a deep learning segmentation model loaded directly,

let's use a standard GrabCut algorithm or simple thresholding/masking if appropriate,

but GrabCut is much better for foreground extraction like a person and a bicycle.

Let's define an initial bounding box around the person and the bicycle.

Image dimensions:

h, w, c = img.shape

The person and bicycle span almost the entire height and width, but let's give a safe box:

top around 10%, bottom 100%, left 10%, right 95%

mask = np.zeros(img.shape[:2], np.uint8)
bgdModel = np.zeros((1, 65), np.float64)
fgdModel = np.zeros((1, 65), np.float64)

rect = (int(w * 0.05), int(h * 0.05), int(w * 0.92), int(h * 0.95))

Apply GrabCut

cv2.grabCut(img, mask, rect, bgdModel, fgdModel, 5, cv2.GC_INIT_WITH_RECT)

Modify the mask so that all definite background and probable background pixels are 0,

and all definite foreground and probable foreground pixels are 1.

mask2 = np.where((mask == 2) | (mask == 0), 0, 1).astype('uint8')

Let's refine the mask a bit with morphological operations to smooth edges

kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
mask2 = cv2.morphologyEx(mask2, cv2.MORPH_CLOSE, kernel)
mask2 = cv2.morphologyEx(mask2, cv2.MORPH_OPEN, kernel)

Create 4-channel PNG

rgba = cv2.cvtColor(img, cv2.COLOR_BGR2BGRA)
rgba[:, :, 3] = mask2 * 255

Save the result

output_path = 'personaj_va_velosiped.png'
cv2.imwrite(output_path, rgba)
print(f"File saved to {output_path}")