Right from the very beginning, text has vital importance in the human life. As compared to the vision based applications preference is always given to the precise and rich information embodied in text. Considering the importance of text, scene text recognition and detection is also equally important in human life. In document analysis and studies of computer scene text detection and recognition are the prime topic of research studies. Particularly, in present times a number of recent research studies are conducted on this topic. A substantial progress made in this field is dependent upon these research studies. The main purpose of this research study is to conduct a survey on this topic and study different directions. Research will focus on identification of state-of-the art algorithms and bring into light the up-to-date research work. Moreover, research will also predict future potential directions of scene text detection and recognition. Somehow, research will ensure link of resources (only publically available), online demos, and source code. In short, present work will help out the future researchers to use this research study and findings as basis to conduct future advanced research studies.

The Vision-based applications are the great source of information. Somehow, information becomes more beneficial to the human beings when embodied in the text. Textual information presented in videos and images can be studied by detecting and recognizing text. Somehow, reading and localizing text in the natural scenes are not quite easy tasks. It is true that text is beneficial in so many ways, but when it comes to detect as well as recognize this text from images, the process is not an easy one, as it has to use variety of mechanism and algorithms. The process is not as simple as it may look, but good thing about this process is that when it is successful it comes with great results. Regardless of its success, some common challenges associated with the scene text detection and recognition process is presented below [1]:

·         Complexity of Scene Text Detection and Recognition

Complex backgrounds are difficult to detect and recognize. Sometimes a clearly written text in regular font on complex backgrounds becomes unreadable. Some examples are natural scenes including bricks and grass in the background.

·         Interference of Scene Text Detection and Recognition

Interference factors including non-uniform illumination, noise, low resolution images, blur backgrounds or text, as well as distortion create challenges for scene text detection and recognition.  

·         Diversity of Scene Text Detection and Recognition     

Scenes text detection and recognition is not difficult if font styles and sizes are same in the document. However, in the presence of different font’s sizes, fonts, scales, and colors detection and recognition process takes time and sometimes comes up with wrong outcomes.

·         Noise & Distortion of Scene Text and Recognition

It has been observed that sometimes images are blurred and there is too much a noise as well as distortion in the image, which means that it is hard for the detection technology to detect and recognize correct text from the images. If process has to be more advanced in this regard that it may also detect blur words with noise in the image, then various explorations of the process can be made, but still, it would remain to be a big challenge.

             Recently conducted research studies have introduced several algorithms that can be useful in the field of scene text detection and recognition [2]. Detection of multi-oriented and diverse text is possible and evaluation of algorithm regarding this can be selected through MSRA_TD500. Another database that contains images with mutely oriented text and backgrounds is known as NEOCR. Basically, in-ability or poor capability of scene detection and recognition in the multi-oriented texts and complex background are key limitations that need to be addressed in this research work. Present work will use survey methodology to study algorithms and limitations of scene text detection and recognition.

Methodology on Scene Text Detection and Recognition

            It is important to describe that what methodology is going to be used for this survey paper about scene text detection and recognition. The research methodology selected for this paper is based on qualitative method, where secondary research data and resources will be analyzed to collect the relevant data. It means that existing research literature will be reviewed in the discussion section to see that what kind of research has been done on the topic, and what topics and areas have been covered in the previous research studies. It is important to get view of the previous research to know that what limitations have been there, and what can be done in future in this regard. The research methodology is the most important part of the paper, as it determines that how research will be conducted. In this paper, there will be no primary research method used to collect data and paper will only be dependent on exiting secondary resources such as journal and peer-reviewed research articles written by various researchers in this field. The reviewed data will be used to make conclusion as well as future recommendations that what else can be reviewed in future research work regarding scene text detection and recognition [3].

Discussion on Scene Text Detection and Recognition

            The process varies in different ways to recognize & detect text, and major purpose for all processes is to ensure that text in the images is correctly read. If method is not good enough, then it will detect wrong words, and whole process of text detection & recognition will be showing incorrect results. So, it is important to review literature that how things have been developed in the past and what kind of methods came with which kind of results to show that things are going towards the right direction or not. So, next part of this brief discussion will review various secondary resources [4].

Literature Review of Scene Text Detection and Recognition

            It is important to see that what kind of recent advances has been made in scene text detection and recognition process. It is also vital to look at these advances so that estimation is made about the future trends in this regard. One of the research studies tried to explore that how things have been developing with the passage of time. The research has found out that various kinds of challenges are faced in the process such as variation, occlusion, distortion, complexity, blur as well as noise. The study reviewed publically available literature from various resources to conduct a detailed survey about recent advances in the field. The point of survey was based on three things; first introduction of up to date work, second identification of algorithms, which are state of the art, and third making predictions about future research directions. The research came up with various approaches and concluded that considerable and significant progress has been made in this area of research, and things are looking bright for the future [3]. Here are two images for recognition of text is given:

           It was explained in a research study that scene text can come up with rich kind of semantic information, and this can be used in different ways in applications, which are vision based. There have been various conventional methods for scene text detection and recognition, and each had its own advantages and disadvantages. The research found out after analyzing these methods that key issues faced by these methods included multi-orientation, loss function as well as sequence labeling & language model. There are various benchmark evaluation protocols & datasets, which can be used to analyze the performance of the overall detection and recognition process [5]

             A research study was conducted recently in 2019, which tried to analyze Curved scene text detection by using longitudinal & transverse sequence connection to see that how process goes with its accuracy. It was described that it is always hard to detect a text, which is curved. So, study developed the dataset of curved text named CTW1500, and there were more than 10,000 actual text annotations in given 1500 images. To detect the text, polygon-based text detector was used in the process. This method used by the researchers came up with good results, and even some of advanced methods were actually outperformed by this particular method. So, a new method was developed in the research to help out the process of detection, where it becomes way easier to detect curved text, which is generally hard to detect and recognize by other methods in this field of technology [6]

            A study that demonstrated an end-to-end educable sensory network, which resolves the  responsibilities in a novel integrated framework. By the allocation of the conventional characteristic maps, we can instantaneously educate the two prototypes in a solitary combined pipeline. Furthermore, we calculated a completely  recognized, which has been verified to be active and effectual. Within this method, assumed the calculation of ,  is able to attain around  calculation tradable. In this outline, we mostly center on the horizontal or near-horizontal texts, as for the upcoming task, there is a requirement to give more responsiveness in managing the texts of several angles [7].

Resource: https://arxiv.org/pdf/1811.08611.pdf

There is still a gap amid the practical prominence and the essential performance point out that  endure to be unexplained problems. Even though countless development has been completed, there are still many study occasions. If compared with the performance of  on ,  is quiet distant overdue. The development will derive not just from the resilient feature of recognition models, but likewise, commencing from well-made information allocation, opinion also the optimization approaches as well. The newly advanced  has significantly developed the feature arrangement performance by studying the hierarchical multi-scale representations. The combination of deep learning along with the improved division,  and also extraordinary command language prototypes could extra increase the performance [8].