Hi,
I got a little sad today when I learned that I have 2300 JAV movies with that nasty yellow Chinese watermark. But the good news is that I am now able to identify them using a little AI tool.
My JAV movie library is somewhat extensive. In my early years of gathering, I didn't really care about quality or if movies had watermarks. So many have that opaque yellow watermark for Chinese casino 22366.com stamped in the bottom left or right corner.
I really wanted to know how many I had.
Fortunately I also have a screenshot with 3x9 tiles of shots from each movie. So I thought that I might write a tool that used a Neural Network and teach it to recognize these movies.
The tool extracts each screenshot tile, then calculates an Image Histogram of Orient Gradients marker for each image and feeds it to a Feedforward Neural Network (AI algorithm). The algorithm gradually learns to tell the difference from watermark/no-watermark using a set of training images.
If you are a programmer, you might want to try it yourself. Links to project below.
The tool is written in JavaScript, rather than the usual Python+TensorFlow combo.
You'll need to install node.js (https://nodejs.org) programming framework.
Then in the extracted source folder, setup with...
npm install
If it complains, you might need to install node-gyp first (https://github.com/nodejs/node-gyp). It can be a little cumbersome to install.
Then try the following commands:
node index set-feature yellowRight
node index train
node index test ABP-318
node index scan ABP
node index scan-all
There are 2 watermark features defined. Each must be trained separately. The above commands will activate the "yellow watermark at bottom-right" detection. There is also a "yellowLeft" feature available in the sample.
Training images are located in ./images/training-images
When a match is found, a small sample is stored in ./images/found
For optimal detection you can add more training images.
Edit the JSON files in ./data to add more. Same to define a new watermark feature.
Screenshot images must have the exact layout to work (I use MPC-HC player).
The project:
https://filejoker.net/bvxt66zavips
https://mega.nz/#F!9J9CWIwA!m8Gws0cph3YD8SQydlc0Sg
Have fun.
I got a little sad today when I learned that I have 2300 JAV movies with that nasty yellow Chinese watermark. But the good news is that I am now able to identify them using a little AI tool.
My JAV movie library is somewhat extensive. In my early years of gathering, I didn't really care about quality or if movies had watermarks. So many have that opaque yellow watermark for Chinese casino 22366.com stamped in the bottom left or right corner.
I really wanted to know how many I had.
Fortunately I also have a screenshot with 3x9 tiles of shots from each movie. So I thought that I might write a tool that used a Neural Network and teach it to recognize these movies.
The tool extracts each screenshot tile, then calculates an Image Histogram of Orient Gradients marker for each image and feeds it to a Feedforward Neural Network (AI algorithm). The algorithm gradually learns to tell the difference from watermark/no-watermark using a set of training images.
If you are a programmer, you might want to try it yourself. Links to project below.
The tool is written in JavaScript, rather than the usual Python+TensorFlow combo.
You'll need to install node.js (https://nodejs.org) programming framework.
Then in the extracted source folder, setup with...
npm install
If it complains, you might need to install node-gyp first (https://github.com/nodejs/node-gyp). It can be a little cumbersome to install.
Then try the following commands:
node index set-feature yellowRight
node index train
node index test ABP-318
node index scan ABP
node index scan-all
There are 2 watermark features defined. Each must be trained separately. The above commands will activate the "yellow watermark at bottom-right" detection. There is also a "yellowLeft" feature available in the sample.
Training images are located in ./images/training-images
When a match is found, a small sample is stored in ./images/found
For optimal detection you can add more training images.
Edit the JSON files in ./data to add more. Same to define a new watermark feature.
Screenshot images must have the exact layout to work (I use MPC-HC player).
The project:
https://filejoker.net/bvxt66zavips
https://mega.nz/#F!9J9CWIwA!m8Gws0cph3YD8SQydlc0Sg
Have fun.
Last edited:
