How to reliably detect edges (under shadows or poor lighting) in the images?

Python
1

How to reliably detect edges (under shadows or poor lighting) in the images ?

I’m trying to identify object boundaries, with edge detection but run into a problem when images have shadows or poor lighting or lower res.

Here is a sample photo.

familyroom_craft-room-preview
familyroom_craft-room-preview728×496 84.6 KB

I use edge detection and sharpening with this code:

def sharpen_edges(binary_data):
    image = Image.open(io.BytesIO(binary_data))
    image_np = np.array(image)

    # Convert to grayscale for edge detection
    gray_image = cv2.cvtColor(image_np, cv2.COLOR_RGB2GRAY)

    # Apply Canny edge detection
    edges = cv2.Canny(gray_image, threshold1=50, threshold2=200)

    # Convert edges to RGB and overlay on the original image
    edges_rgb = cv2.cvtColor(edges, cv2.COLOR_GRAY2RGB)

    # Increase the contrast of edges by blending them with the original image
    sharpened_np = cv2.addWeighted(image_np, 1.0, edges_rgb, 1.5, 0)

    # Optional: Apply a slight Gaussian blur to soften the edges a bit
    sharpened_np = cv2.GaussianBlur(sharpened_np, (3, 3), 0)

    # Convert back to PIL image and save to buffer
    sharpened_image = Image.fromarray(sharpened_np)
    buffer = io.BytesIO()
    sharpened_image.save(buffer, "PNG")
    sharpened_image_data = buffer.getvalue()

    return sharpened_image_data

Appreciate any advice on this issue. I’m open-cv newbie so please bear with me as I try to understand what’s happening.

2

I couldn’t upload the result of edge detection. Here it is:

FURNITURE-sharpened-input
FURNITURE-sharpened-input728×496 96.8 KB

As you can see the area under the sofa - it’s not able to identify the wooden frame under this sofa as its in the shadow of sofa itself.

I tried plenty of techniques like different edge detection (like Laplacian, Sobel) or shadow removal, but its not working as expected.

3

wrong approach. that will never do the job.

you need AI for this. “semantic segmentation”.

1 Like