[WEEK 2] Image Colorization Results With Filters
This week, we have applied unsharpen filter to input image and we have observed how this filter affects result.
Filters can change the result of colorization operation. Unsharpen filter clarifies the edges of input images. This causes more successful object detection. But as a side effect, network may not be trained depending on clarified edges. For example, characteristic color of a tree is green. Network trained on this characteristic. If we apply unsharpen filter to this tree, leaves of this tree becomes more visible and if network does not trained on leaf, colorization operation becomes unsuccessful. At this point, we realised that colorization with sharpened input image depends on network. If network has large amount of data, sharpened input images can give more successful results but restricted data causes unsuccessful colorization operations.
At these examples, we can clearly see that sharpened input images directly affects the result positively.
And this example shows sometimes we are not as lucky as colorize the first image. Network percieved man’s hair as grass because of sharpened edges.
Different filters can give us different results. For example Gaussian Filter which we will refer at next post, is low-pass filter which means lost of edge informations. We realized that as sharpened edges can cause worse results, lose of edge informations can also causes worse results.
To sum up, we examined on effects of unsharpen filter this week. And depending on network we can got better or worse results compared to not filtered input image. At filtering input image, success of colorization operation depends on network.
