Add enhance_image4 function for advanced image enhancement using MSRCR

This commit is contained in:
2025-02-02 06:16:02 +01:00
parent 30ddd72321
commit 8d83435be5

70
main.py
View File

@ -308,6 +308,71 @@ def enhance_image3(image: bytes) -> bytes:
return enhanced_webp.tobytes()
def enhance_image4(image: bytes) -> bytes: # noqa: PLR0914
"""Enhance an image using a simplified Multi-Scale Retinex with Color Restoration (MSRCR) algorithm to better reveal details in dark areas.
This approach first denoises the image, then applies multi-scale retinex processing to
boost local contrast. A color restoration step helps maintain a natural look.
Args:
image (bytes): The input image to enhance.
Returns:
bytes: The enhanced image encoded in WebP format.
""" # noqa: E501
# Decode the image from bytes
nparr = np.frombuffer(image, np.uint8)
img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
# Denoise the image with conservative settings
img = cv2.fastNlMeansDenoisingColored(img, None, 5, 5, 7, 21)
# Convert to float64 and add 1 to avoid log(0)
img = img.astype(np.float64) + 1.0
# Define scales for the multi-scale retinex (you can experiment with these)
scales = [15, 80, 250]
retinex = np.zeros_like(img)
# Compute the retinex output over different scales
for sigma in scales:
# Gaussian blur with standard deviation sigma; kernel size is computed automatically
blur = cv2.GaussianBlur(img, (0, 0), sigma)
retinex += np.log(img) - np.log(blur)
# Average the retinex result over all scales
retinex /= len(scales)
# --- Color Restoration Step ---
# Compute the sum across color channels (with a small epsilon to avoid division by zero)
eps = 1e-6
sum_channels = np.sum(img, axis=2, keepdims=True) + eps
# The color restoration factor; alpha is chosen empirically (here 125 works well in many cases)
color_restoration = np.log(125 * img / sum_channels + 1)
# Combine the retinex output with the color restoration factor
msrcr = retinex * color_restoration
# Apply gain and offset adjustments to fine-tune brightness and contrast
gain = 1.5
offset = 20
msrcr = msrcr * gain + offset
# Normalize each channel to span the full 0-255 range
for channel in range(3):
ch_data = msrcr[:, :, channel]
ch_min, ch_max = ch_data.min(), ch_data.max()
msrcr[:, :, channel] = ((ch_data - ch_min) / (ch_max - ch_min + eps)) * 255
# Clip the values to valid 8-bit range and convert back to uint8
enhanced_img = np.clip(msrcr, 0, 255).astype(np.uint8)
# Encode the enhanced image to WebP format
_, enhanced_webp = cv2.imencode(".webp", enhanced_img)
return enhanced_webp.tobytes()
@client.tree.context_menu(name="Enhance Image")
@app_commands.allowed_installs(guilds=True, users=True)
@app_commands.allowed_contexts(guilds=True, dms=True, private_channels=True)
@ -350,7 +415,10 @@ async def enhance_image_command(interaction: discord.Interaction, message: disco
enhanced_image3: bytes = enhance_image3(image_bytes)
file3 = discord.File(fp=io.BytesIO(enhanced_image3), filename=f"enhanced3-{timestamp}.webp")
await interaction.followup.send("Enhanced version:", files=[file1, file2, file3])
enhanced_image4: bytes = enhance_image4(image_bytes)
file4 = discord.File(fp=io.BytesIO(enhanced_image4), filename=f"enhanced4-{timestamp}.webp")
await interaction.followup.send("Enhanced version:", files=[file1, file2, file3, file4])
except (httpx.HTTPError, openai.OpenAIError) as e:
logger.exception("Failed to enhance image")