invisible-watermark

invisible-watermark is a python library and command line tool for creating invisible watermark over image (a.k.a. blink image watermark, digital image watermark). The algorithm doesn't rely on the original image.

Note that this library is still experimental and it doesn't support GPU acceleration, carefully deploy it on the production environment. The default method dwtDCT(one variant of frequency methods) is ready for on-the-fly embedding, the other methods are too slow on a CPU only environment.

supported algorithms

• Discrete wavelet transform + Discrete cosine transform frequency embedding algorithm variants.
• RivaGAN, a deep-learning model trained from Hollywood2 movie clips dataset.

speed

• default embedding method dwtDct is fast and suitable for on-the-fly embedding
• dwtDctSvd is 3x slower and rivaGan is 10x slower, for large image they are not suitable for on-the-fly embedding

accuracy

• The algorithm cannot guarantee to decode the original watermarks 100% accurately even though we don't apply any attack.
• Known defects: Test shows all algorithms do not perform well for web page screenshots or posters with homogenous background color

Supported Algorithms

• frequency methods

• dwtDct: DWT + DCT transform, embed watermark bit into max non-trivial coefficient of block dct coefficents

• dwtDctSvd: DWT + DCT transform, SVD decomposition of each block, embed watermark bit into singular value decomposition

• rivaGan: encoder/decoder model with Attention mechanism + embed watermark bits into vector.

background:

• Discrete wavelet transform
• Discrete cosine transform.
• RivaGAN, a deep-learning model trained from Hollywood2 movie clips dataset.

How to install

pip install invisible-watermark

Library API

Embed watermark

• example embed 4 characters (32 bits) watermark

import cv2
from imwatermark import WatermarkEncoder

bgr = cv2.imread('test.png')
wm = 'test'

encoder = WatermarkEncoder()
encoder.set_watermark('bytes', wm.encode('utf-8'))
bgr_encoded = encoder.encode(bgr, 'dwtDct')

cv2.imwrite('test_wm.png', bgr_encoded)

Decode watermark

• example decode 4 characters (32 bits) watermark

import cv2
from imwatermark import WatermarkDecoder

bgr = cv2.imread('test_wm.png')

decoder = WatermarkDecoder('bytes', 32)
watermark = decoder.decode(bgr, 'dwtDct')
print(watermark.decode('utf-8'))

CLI Usage

embed watermark:  ./invisible-watermark -v -a encode -t bytes -m dwtDct -w 'hello' -o ./test_vectors/wm.png ./test_vectors/original.jpg

decode watermark: ./invisible-watermark -v -a decode -t bytes -m dwtDct -l 40 ./test_vectors/wm.png

positional arguments:
  input                 The path of input

optional arguments:
  -h, --help            show this help message and exit
  -a ACTION, --action ACTION
                        encode|decode (default: None)
  -t TYPE, --type TYPE  bytes|b16|bits|uuid|ipv4 (default: bits)
  -m METHOD, --method METHOD
                        dwtDct|dwtDctSvd|rivaGan (default: maxDct)
  -w WATERMARK, --watermark WATERMARK
                        embedded string (default: )
  -l LENGTH, --length LENGTH
                        watermark bits length, required for bytes|b16|bits
                        watermark (default: 0)
  -o OUTPUT, --output OUTPUT
                        The path of output (default: None)
  -v, --verbose         print info (default: False)

Test Result

For better doc reading, we compress all images in this page, but the test is taken on 1920x1080 original image.

Methods are not robust to resize or aspect ratio changed crop but robust to noise, color filter, brightness and jpg compress.

rivaGan outperforms the default method on crop attack.

only default method is ready for on-the-fly embedding.

Input

• Input Image: 1960x1080 Image
• Watermark:
  • For freq method, we use 64bits, string expression "qingquan"
  • For RivaGan method, we use 32bits, string expression "qing"
• Parameters: only take U frame to keep image quality, scale=36

Attack Performance

Watermarked Image

Attacks	Image	Freq Method	RivaGan
JPG Compress		Pass	Pass
Noise		Pass	Pass
Brightness		Pass	Pass
Overlay		Pass	Pass
Mask		Pass	Pass
crop 7x5		Fail	Pass
Resize 50%		Fail	Fail
Rotate 30 degress		Fail	Fail

Running Speed (CPU Only)

Image	Method	Encoding	Decoding
1920x1080	dwtDct	300-350ms	150ms-200ms
1920x1080	dwtDctSvd	1500ms-2s	~1s
1920x1080	rivaGan	~5s	4-5s
600x600	dwtDct	70ms	60ms
600x600	dwtDctSvd	185ms	320ms
600x600	rivaGan	1s	600ms

RivaGAN Experimental

Further, We will deliver the 64bit rivaGan model and test the performance on GPU environment.

Detail: https://github.com/DAI-Lab/RivaGAN

Zhang, Kevin Alex and Xu, Lei and Cuesta-Infante, Alfredo and Veeramachaneni, Kalyan. Robust Invisible Video Watermarking with Attention. MIT EECS, September 2019.[PDF]