BscThesis

Témakiírás:

Angol témacím: Applied artifical intelligence on Java, Node.JS or on other platform

Témacím: Alkalmazott Mesterséges Intelligencia Java, Node.JS vagy egyéb platformon

Dolgozat:

English title of the thesis: Examination and application of image processing algorithms (with neural nets)

Szakdolgozat címe: Képfeldolgozási algoritmusok vizsgálata és alkalmazása (neurális hálókkal)

Konzulens: Dr. Ekler Péter

---

Link to organisational Drive

Link to Colab

Link to Colab-backing Drive

Some of these links maybe asks for special permission to access. Please contact me, if you meet some problems here.

Idea:

image processing: learn good exposure and some other retouch values on image (exposure, contrast, white balance, tint, vibrance values), and apply these to other images. So get an automatic retouch tool

---

Tasks, subtasks

• raw processing:
• process jpgs with CNN (get train, valid sets)

  • #ofHiddenLayers ?

  • input: raw converted to binary - uniform size

  • output: predicted values

  • error function:

    • multiple error funtion for the different features?
    • multiple parallel CNN to learn the different features?

    MSE based on predicted and actual exif fields - compare the output with the fields of the exifdata

  • optimizer: Adam is OK and effective enough, but let's try others

• train the NNs, fix the trained NNs
  • dropout
  • in the beginning -- small dataset, but many epochs
• NN returns output
• generate .xmp file from output lists for raw image (same name as the raw)
• rewrite fields of originally created .xmp from raw -- based on the lists
  • process .xmp (load & write)
• load automate generated .xmp-s and raws into Lightroom or other image processing software