This is a term project investigating efficient and transferable multilingual scene text recognition tasks through deep learning. The datasets experimented are CVSI-2015 and MLT-2019, both are publicated online. These datasets pose specific challanges, such as varying aspect ratios of input images and large data size which are not feasible through conventional training approaches. The techniques include depth-wise convolutional neural networks, deep averaging network, vision transformers with 2d positional embeddings and more, which provide a feasible way to deal with some of the challenges in a small scale setting.
The report and .ipynb files are attached here. More descriptions will be added in the future.