Article Preview
Top1. Introduction
The present era has witnessed technology etching upon every aspect of our life. Be it, online transactions, reservations and what not, technology has paved its path into our lives. In the current pandemic situation, online education has undergone a massive transition. This is where our audiobooks will be tremendously helpful, especially for school, college students and professors (Noman et. al., 2016). There are a lot of existing systems, including Audible by Amazon, AppleBooks by Apple. Though this sounds good, the existing audiobooks available in the market are not free and feasible for everyone. Not only that, these audiobooks are available only for some prescribed books such as comics or novels. If you take a look at the market, there are a number of services offered by different companies starting from Amazon Audible to iTunes and so on. But in all these services, a narrator reads the story which the user listens to. This involves a lot of human work, and cost.
Hence, we have decided to come up with an audiobook that can be used for any printed material including books, school textbooks as well as storybooks, research papers, newspapers, documents, a single sheet of text or multiple pages wherein users have to just upload the printed information and enjoy listening to the content. Moreover, this is free of cost, and therefore can be used by anyone. Further, we use an auto-generated voice that translates the text to voice which saves a lot of labour.
The proposed framework has 3 important modules. Preparing clean and clear images for the recognition engines is often taken for granted as a trivial task that requires little attention. However, this step undoubtedly influences the overall performance of the system. The first module involves denoising the images that are fed by the user. For this purpose, Mayank et. al., (2020) have used CNN architecture for separating the foreground and background. This work acknowledges the problems of restoration of the useful content of handwritten documents and reconstruction of the ‘most likely’ appearance of the original documents. Jose et. al., (2005), Lavanya et. al., (2020) and Chunwei et. al., (2015) have discussed neural networks approach for cleaning the dirty documents.
The second step is text extraction. Text Extraction is useful in information retrieving, searching, editing, documenting, archiving or reporting of image text. Jian et. al., (2013) have tried to find a new way which can comprehensively utilize existing methods to detect and extract text from born- digital image. Cartic et. al., (2012) uses edge detection methods. Similarly, we plot bounding boxes and identify the Region of Interest, ROI.
Following this, we have designed post processing steps to identify the errors and suggest suitable alternatives. However, the basic outline for building this work emerged from the extensive work undertaken by Sasirekha et. al., (2013). A different approach is used to extract text from newspapers, wherein the user has the choice to manually select the columns he wishes to. The underlying analysis will provide a detailed review of each of the modules with the outcome of each step.