[WEEK 6] Image Colorization with Reconstruction
The results of colorization we obtained from our chosen model were not satisfactory. If you look at the results we have shown before, you can see that we cannot get a satisfactory result from the colorization process. We cannot be able to change the model as we realize this situation too late, we did not have time. In this section, we decided to show the results of our application with a different model. Richard Zhang has created a demo of the his model on the internet. We have achieved better colorization results using this model. Of course, there were some minor erroneous results in colorization.We have given the images after using Gaussian filter and sharpening as input in this demo. Then we combined the results as in the reconstruction part. The results we get are more accurate than the results of the model. The reason for this is that we get the color layers from the image we obtained as a result of Gaussian, and the details from the image we obtained as a result of sharpening. In this way, we were able to obtain more accurate results.
If we look at the results from different model comparison, as we have already guessed, the training of the models is one of the most impressive factors for colorization. As we have seen, the model we are using is more likely to be trained with nature images. Because when we look at the result, we can see that the weighted green colorization is done. If we look at the other model , we see that it cannot make an accurate coloring due to its own training dataset. At this point, we can say that our method is an effective method to correct this situation. The results we achieved after implementing our own implementation were successful. This shows that the problems and solutions we think are correct.
