FINAL NOTES [WEEK 7]

If we apply a filter that highlights details we can get better details on output image but at color side we can become unsuccessful. Like this, if we apply a filter that makes image smoother we lose details but at color side we can get better results. Reason of that is models are likely to recognize whole object.

When we increased the kernel size and sigma value or shortly increased the Gaussian blur level, the results were poor. Because the increase in Gaussian blur level means getting more difficult to recognize objects and also means extremely detailed outputs. Combine of these, means details are more than it should be and colorizing is poor because of lack of recognizing objects. So after some tries we found optimum values and we get better results than program’s original results. If we need to summarize; filter selection is important but without optimum parameter value selection, filter selection loses its importance.

Model is fundamental factor of image colorization. For more successful colorization results, the model must be trained with as many data sets as possible. At this point we saw that our model is really weak. Our model is poor at images different than nature images. Therefore our results are green color weighted. Then we tried our method with different model. Different model is stronger than ours because it has ability that recognizes more different objects than nature images.

At reconstruction part, we use Bilateral filter for getting details. Reason of why we prefer Bilateral filter is Bilateral filter does not lose edge information on other words it does not make blurring makes smoothing only. With applying Bilateral filter to sharpened output, we get a large scale and smoothed image. By using this large-scale and smoothed image, we simply divide the sharpened image to large-scale image to obtain the details. Therefore, we have got details of sharped image. Now we must get color layer. We use Gaussian filtered output image. Simply we divide Gaussian filtered output image to its intensity layer. Now time to reconstruction with this features. First we multiply large scale image by details. This operation gives us intensity layer of the image that will be reconstructed. After that, the multiplication operation between intensity layer of the image that will be reconstructed and color layer which we obtained from Gaussian filtered output gives us final result.

So, we try to examine how filters affect the colorization operation.