The proposal is inspired by the idea G on RedHen's GSoC 2016 idea page. The main purpose is to develop models and code helping domain experts to research on art history. It includes five main tasks:
- Training a classifier for recognizing a limited number of known visual symbols in images of religious artworks. VGG-net based fully convolutional network models will be trained for this task. Multi-scale detection, non-maximum suppression and bounding box regression will be implemented.
- Training a detector for localizing person heads and hands in images of art with various styles. CNN models will be trained based on a state-of-art head detection method.
- Building a module identifying a few certain hand gestures in images of artworks based on task 2. With the cropped detection result of task 2, a classifier based on VGG-net will be built.
- Developing a tool generating a frontal view of a face given a artistic portrait with a side view of the face based on existing face frontalization code.
- Developing initial research code for analysis of color style changes over time of Christian paintings by machine learning.