Editing Halftone (section)

===Optimization-based filtering===

Another possibility for inverse halftoning is the usage of [[machine learning]] algorithms based on [[artificial neural network]]s.<ref>{{Citation|last1=Li|first1=Yijun|title=Deep Joint Image Filtering|date=2016|work=Computer Vision – ECCV 2016|pages=154–169|publisher=Springer International Publishing|isbn=978-3-319-46492-3|last2=Huang|first2=Jia-Bin|last3=Ahuja|first3=Narendra|last4=Yang|first4=Ming-Hsuan|series=Lecture Notes in Computer Science |volume=9908 |doi=10.1007/978-3-319-46493-0_10}}</ref> These learning-based approaches can find the descreening technique that gets as close as possible to the perfect one. The idea is to use different strategies depending on the actual halftone image. Even for different content within the same image, the strategy should be varied. [[Convolutional neural network]]s are well-suited for tasks like [[object detection]] which allows a category based descreening. Additionally, they can do edge detection to enhance the details around edge areas. The results can be further improved by [[generative adversarial network]]s.<ref>{{cite journal|last1=Kim|first1=Tae-Hoon|last2=Park|first2=Sang Il|date=2018-07-30|title=Deep context-aware descreening and rescreening of halftone images|journal=ACM Transactions on Graphics|volume=37|issue=4|pages=1–12|doi=10.1145/3197517.3201377|s2cid=51881126|issn=0730-0301}}</ref> This type of network can artificially generate content and recover lost details. However, these methods are limited by the quality and completeness of the used training data. Unseen halftoning patterns which were not represented in the training data are rather hard to remove. Additionally, the learning process can take some time. By contrast, computing the inverse halftoning image is fast compared to other iterative methods because it requires only a single computational step.