Reading or Recognizing Text from Images is a challenging Task in the field of Computer Vision. This is mainly because Text in Images exhibit diversity and variability. Backgrounds of Images are virtually unpredictable as some images may have dark background, some may have light background and there are even cases where image fonts and background have same color with a small outline differentiating Text from the Background. There might be patterns of Text appearing in the image with different fonts, orientations, and varying lengths. The Text in Images can also be blurred or distorted.
Given an input Image we need to predict the Text in the Image with a reasonable accuracy >80% (Exact match with the actual Text Labels) and should have a good letter match accuracy.
Well that is fine you say, are there any constraints? Yes, there are.
There is a Low Latency requirement meaning, given a picture the model should be able to predict the Text quickly Predict Text from Images with Variable length Labels/Words.
Accuracy is one of the main metrics to evaluate the Model. Additionally the model can be evaluated on other metrics as well like given below:
1.Letter Accuracy 2.CTC Loss