OpenCV透视变换后Logo出界及空白区域问题求解

Nice catch! Let's dig into why this is happening and walk through three solid solutions to fix both the cropping and unwanted white space issues with your OpenCV perspective warp.

The Root Cause #

Your current code applies the perspective transform directly to the original image dimensions. When you warp the logo, parts of it get pushed outside the original canvas and get cropped out. Meanwhile, the empty spots left behind by the warped logo get filled with the white borderValue you set—those are the useless white areas you're seeing.

Solution 1: Dynamically Resize Canvas to Fit the Warped Logo #

This approach calculates the exact bounds of your warped logo first, then creates a canvas that perfectly fits it. No cropping, no extra white space.

Step-by-Step Code #

import numpy as np
import cv2

def warp_logo_without_crop(img):
    # Generate your original perspective matrix
    perspective = np.eye(3, dtype=np.float32) + np.random.uniform(-0.0015, 0.0015, (3,3))
    perspective[2][2] = 1.2

    # Get the four corner points of the original image
    h, w = img.shape[:2]
    src_corners = np.float32([[0,0], [w-1,0], [w-1,h-1], [0,h-1]])

    # Transform those corners using the perspective matrix
    warped_corners = cv2.perspectiveTransform(src_corners.reshape(-1,1,2), perspective)

    # Calculate the new canvas size needed to fit all warped corners
    min_x, max_x = np.min(warped_corners[:,:,0]), np.max(warped_corners[:,:,0])
    min_y, max_y = np.min(warped_corners[:,:,1]), np.max(warped_corners[:,:,1])
    new_width = int(np.ceil(max_x - min_x))
    new_height = int(np.ceil(max_y - min_y))

    # Create a translation matrix to shift the warped logo to the top-left of the new canvas
    translate_mat = np.array([
        [1, 0, -min_x],
        [0, 1, -min_y],
        [0, 0, 1]
    ], dtype=np.float32)

    # Combine perspective and translation matrices for the final transform
    final_transform = translate_mat @ perspective

    # Warp the image to the new canvas size
    warped_img = cv2.warpPerspective(img, final_transform, (new_width, new_height), borderValue=(255,255,255))
    return warped_img

# Usage example
src_logo = cv2.imread("src_logo.png")
dst_logo = warp_logo_without_crop(src_logo)
cv2.imwrite("dst_logo.png", dst_logo)

Solution 2: Constrain Distortion to Keep Logo Within Original Bounds #

If you need to keep the original image size, you can limit the random perspective distortion until the warped logo stays entirely inside the canvas.

Code Implementation #

import numpy as np
import cv2

def warp_logo_in_bounds(img, max_distortion=0.001):
    h, w = img.shape[:2]
    src_corners = np.float32([[0,0], [w-1,0], [w-1,h-1], [0,h-1]])
    valid_transform = False

    # Keep generating transforms until we get one that doesn't crop the logo
    while not valid_transform:
        perspective = np.eye(3, dtype=np.float32) + np.random.uniform(-max_distortion, max_distortion, (3,3))
        perspective[2][2] = 1.2
        
        # Check if all warped corners stay within the original canvas
        warped_corners = cv2.perspectiveTransform(src_corners.reshape(-1,1,2), perspective)
        if (np.min(warped_corners[:,:,0]) >= 0 and
            np.max(warped_corners[:,:,0]) <= w-1 and
            np.min(warped_corners[:,:,1]) >= 0 and
            np.max(warped_corners[:,:,1]) <= h-1):
            valid_transform = True

    # Apply the valid transform
    warped_img = cv2.warpPerspective(img, perspective, (w, h), borderValue=(255,255,255))
    return warped_img

Note: If you set max_distortion too high, this might loop a few times before finding a valid transform, but it guarantees no cropping or extra white space.

Solution 3: Warp Only the Logo's Valid Area (If Original Has Blank Space) #

If your source logo image already has white padding around it, first extract the actual logo region, warp that, then paste it back into the original-sized canvas. This cuts down on unnecessary white space.

Code Example #

import numpy as np
import cv2

def warp_only_logo_area(img):
    # Convert to grayscale and find the logo's contour
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    _, thresh = cv2.threshold(gray, 240, 255, cv2.THRESH_BINARY_INV)
    contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    
    if not contours:
        return img  # No logo found, return original
    
    # Get the bounding box of the largest contour (the logo)
    x, y, roi_w, roi_h = cv2.boundingRect(max(contours, key=cv2.contourArea))
    logo_roi = img[y:y+roi_h, x:x+roi_w]

    # Warp the ROI using Solution 1's function
    warped_roi = warp_logo_without_crop(logo_roi)

    # Create a blank original-sized canvas and paste the warped logo back
    final_img = np.full_like(img, 255)
    # Calculate paste position (keep original position, adjust if needed to fit)
    paste_x = x
    paste_y = y
    if paste_x + warped_roi.shape[1] > final_img.shape[1]:
        paste_x = final_img.shape[1] - warped_roi.shape[1]
    if paste_y + warped_roi.shape[0] > final_img.shape[0]:
        paste_y = final_img.shape[0] - warped_roi.shape[0]
    
    final_img[paste_y:paste_y+warped_roi.shape[0], paste_x:paste_x+warped_roi.shape[1]] = warped_roi
    return final_img

内容的提问来源于stack exchange，提问作者Sungwon Park