[WEEK 5] Image Colorization with Reconstruction
Last weeks, we try to improve colorization with Gaussian filter, sharpening and Bilateral filter. Each filter had good portions. For example, Gaussian filter lost detail but for some inputs, its colorization can be better than model output without any filter. Sharpening increase edge information but for colorization it did not give good results. We decided to combine our results.
For combining output of Gaussian filter and sharpening, we need color layer, detail and large scale of image. For color layer, we use output of Gaussian filter. Because results of Gaussian filter, colorization operation can be successful for nature images. We get details from output of sharpening. We applied the bilateral filter to the image. Reason of using bilateral filter, it decreases noise of image and it does not lose edge information. After getting image from bilateral filter, we divided it and intensity layer. We multiply detail and large scale of output of Gaussian filter and we get intensity values that include edge information. After that multiplication, we multiply this intensity layer and color layer. After this calculation, we get result image.
If we compare output of model and output of model after reconstruction, we have found that we have achieved better results. Reason of this, Gaussian filter gives successful colorization results. Sharpening does not make very good colorization but highlighted the details become better. After merge outputs of Gaussian and sharpening our results have become better.
