Illumination correction

soutobias commented 3 months ago

What:

Illumination correction addresses uneven lighting in images to ensure uniform light distribution. This is crucial in underwater imaging, where uneven lighting is common due to lens vignetting and light source positioning. Lens vignetting reduces brightness towards the edges of an image, caused by lens design flaws, focal length, or aperture settings. Uneven light source positioning in underwater vehicles, like AUVs or ROVs, creates high-brightness spots near the light source and darker areas elsewhere, affecting color accuracy and overall image quality.

Why:

Proper illumination correction is essential for accurate image analysis and interpretation. Inconsistent lighting can distort colors, obscure details, and hinder the evaluation of underwater scenes. Correcting illumination helps standardize images, making them more suitable for tasks such as biodiversity assessment, species identification, and visual documentation.

How:

Illumination correction can be achieved through various techniques:

Background Subtraction:
- Median Filtering: Removes noise and enhances uniformity by replacing each pixel with the median value of neighboring pixels.
- Gaussian Blur: Smoothens the image and reduces lighting variations by applying a Gaussian filter to estimate and subtract background illumination.
Histogram Matching:
- This technique normalizes lighting by adjusting the image’s histogram to match a reference histogram. It ensures consistent light distribution across the image.
Retinex Algorithms:
- Single-Scale Retinex (SSR): Enhances image details by improving local contrast and brightness. It simulates human visual perception to correct lighting non-uniformities.
- Multi-Scale Retinex (MSR): Extends SSR by analyzing multiple scales, providing a more comprehensive correction for varying lighting conditions.

Python Code Examples:

Median Filtering for Background Subtraction:

import cv2
import numpy as np

def apply_median_filter(image_path, ksize=5):
    # Load image
    img = cv2.imread(image_path, cv2.IMREAD_COLOR)

    # Apply median filter
    filtered_img = cv2.medianBlur(img, ksize)

    # Save and display result
    cv2.imwrite('median_filtered.jpg', filtered_img)
    cv2.imshow('Median Filtered Image', filtered_img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

# Example usage
apply_median_filter('image.jpg')

Histogram Matching:

import cv2
import numpy as np

def match_histograms(source_path, reference_path):
    # Load images
    source = cv2.imread(source_path, cv2.IMREAD_COLOR)
    reference = cv2.imread(reference_path, cv2.IMREAD_COLOR)

    # Convert to YUV color space
    source_yuv = cv2.cvtColor(source, cv2.COLOR_BGR2YUV)
    reference_yuv = cv2.cvtColor(reference, cv2.COLOR_BGR2YUV)

    # Equalize histograms of Y channel
    source_yuv[:,:,0] = cv2.equalizeHist(source_yuv[:,:,0])
    reference_yuv[:,:,0] = cv2.equalizeHist(reference_yuv[:,:,0])

    # Convert back to BGR color space
    result = cv2.cvtColor(source_yuv, cv2.COLOR_YUV2BGR)

    # Save and display result
    cv2.imwrite('histogram_matched.jpg', result)
    cv2.imshow('Histogram Matched Image', result)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

# Example usage
match_histograms('source_image.jpg', 'reference_image.jpg')

Retinex Algorithm (Simple Implementation):

import cv2
import numpy as np

def retinex_enhancement(image_path, sigma=30):
    # Load image
    img = cv2.imread(image_path, cv2.IMREAD_COLOR)

    # Convert to float32 for better precision
    img_float = np.float32(img)

    # Apply Gaussian blur
    blur = cv2.GaussianBlur(img_float, (0, 0), sigma)

    # Retinex enhancement
    img_retinex = np.log10(img_float + 1) - np.log10(blur + 1)
    img_retinex = np.exp(img_retinex) - 1
    img_retinex = np.uint8(np.clip(img_retinex, 0, 255))

    # Save and display result
    cv2.imwrite('retinex_enhanced.jpg', img_retinex)
    cv2.imshow('Retinex Enhanced Image', img_retinex)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

# Example usage
retinex_enhancement('image.jpg')

What to expect:

The result will be an image with improved and more uniform illumination. Corrected images will show a more consistent light distribution with reduced brightness gradients and better visual clarity, allowing for clearer and more accurate analysis of image details.

What makes it difficult:

Dynamic Lighting: Varying lighting conditions across different images or within a single image can complicate correction efforts, making it challenging to achieve uniform results.
Information Loss: Over-aggressive correction may lead to loss of important image details, potentially flattening the image and reducing its informational value.
Algorithm Tuning: Different algorithms and parameters may be needed based on specific lighting conditions and image characteristics, requiring careful tuning for optimal results.

Success Metrics:

Uniformity of Light Distribution: Successful illumination correction will result in an evenly lit image with minimal overexposed or underexposed areas.
Detail Preservation: Enhanced images should maintain important details while providing a clear and consistent view of all features, ensuring that the visual information remains intact and useful for analysis.

LoicVA commented 2 months ago

For histogram matching, what would be the reference image?

Mojtabamsd commented 2 months ago

for high resolution images, rolling ball is slow, need to optimize the code or resize image before processing

soutobias commented 2 months ago

@Mojtabamsd , I apply some changes on your code on the develop branch. Please take a look and see if everything is fine

paidiver / paidiverpy