Computer Vision News - November 2019

3 Summary Tesseract OCR with Python 27 Limitations Tesseract is using a pre-trained model trained to identify regular characters. Hence, it has two main limitations: the first is that Tesseract is not a magic pill. In order to recognize correctly, it needs to get a 'comfortable image', meaning the text needs to be the main object in the image, apparent and aligned. To solve this issue, we need to integrate our code into a detection mechanism that first extracts the ROI and only later invokes the recognition. Yolo, for example, can make good work to solve this issue. The second vulnerability of such pre-trainedmodel is the resistance to noise. We can take our image, add some noise and our systemwill fail to recognize the characters. For example, the image below will produce the following output: Solving this issue depend on the application. In some applications, we will need to train our own model and build task specific model. In most applications, we can use several preprocessing manipulations to increase the robustness of our OCR system. This is what we will see next.