AAML

Trying to solve the anti-aliasing problem with machine learning models. In this project, a 3x3 filter was used to estimate the middle pixel values based on RGBA. As an alternative research, it is tried to increase the image quality with these models.
For this project PNG, JPG and JPEG image files can be used for the dataset.
At this point, it is important to enlarge the data set. After that, training will be done using the machine learning model and dataset then tests will be done on many pictures.
At 28/01/2023, there are nearly 200.000.000 rows of data. The data is splitted into 5 datasets. Each dataset has unique rows in themselves. Therefore each dataset can be used as single datasets. In training, the datasets are selected as random.
At 18/03/2023, there are nearly 380.000.000 rows of data. The data is splitted into 10 datasets. Each dataset has unique rows in themselves. Therefore each dataset can be used as single datasets. In training, the datasets are selected as random.
Currently "ADAM_RGB", "model" and "RMSPROP_RGBA" are trained. The training is ongoing.
Using Python 3.7.8. Used packages can be found in requirements.txt file.

Attributes

At the row by row version, model predicts one row and changes the pixel values and predict below row. This can be useful because when model is predicting it uses upper predicted row.
At the full image version, model predicts with all of the original image's pixel values then changes the pixel values.