Replication of “Real-time Scene Text Localization and Recognition” and “Text Localization in Real-world Images using Efficiently Pruned Exhaustive Search”

I present a replication of two papers: “Real-time Scene Text Localization and Recognition” [10] and “Text Localization in Real-world Images using Efficiently Pruned Exhaustive Search” [9]. The combination of these two papers by Neumann et. al. present a text localization system for scene images. [10] presents a method for generating potential character regions in the image, and [9] presents a method for combining these regions into words. Overall, even with the combination of two papers, this was a difficult project to replicate. This was to be expected to a certain extent as they are both papers from conference proceedings. In short, I was able to implement the majority of both papers, with the exception of a few portions of their system. I did however, gain great insights into the underlying algorithms while studying the system’s vague details. In this paper I present potentially novel pseudocode for a key algorithm in [9], tips for which data structures to use for implementation of this system, and the identity of the key components of the papers which are required for full replication. Lastly, I present qualitative experimental results which I obtained without the aforementioned key details.