 |
The dataset consists of 1k¡¿1k pixels image of mandrill images (512¡¿ 512), as shown in Figure, which are combined in different orientations. This is because Mandrill images have a high spatial frequency and are difficult to identify; therefore, images that are as difficult to discriminate as possible are more effective for training. Of the total 5000 labelled patterns, 80% were used for training and 20% for testing. The back-propagation method was used for error propagation, and simulation was performed using python.
As a result of the machine learning training, the correctness rate exceeded 99% in the early stage of training and finally reached 99.875% on the training data and 99.5% on the test data.¡¡Because we plan to include ECC in the watermark extractor, which corrects about 10% of burst errors, this correct rate is sufficient for implementation.
|
IEEE/ICACT20220093 Slide.17
[Big Slide]