Picture document blind denoising system, method and device

Abstract

The invention belongs to the technical field of computer vision, and particularly relates to a picture document blind denoising system, method and device. The system comprises: a preprocessing unit which is used for performing Gaussian blur processing, graying processing and binarization processing on a to-be-processed picture to generate a first processing result; a straight line detection unit which is used for carrying out straight line detection on the first processing result to generate a first detection result; a table detection unit which is used for carrying out table detection on thebasis of the first detection result and generating a second detection result; a text direction processing unit which is used for carrying out text direction detection on the basis of the second detection result to generate a third detection result, and processing the first processing result according to the third detection result to generate a second processing result; a stretch removing unit which is used for carrying out stretch removing processing on the second processing result to generate a third processing result; and a distortion removing unit which is used for carrying out stretch removing or distortion removing processing on the third processing result to generate a final processing result. And the method can be perfectly adapted to a deep network model for OCR.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a system, method and device for blind denoising of picture documents. Background technique [0002] OCR has been a hot spot in the CV field for a long time. With the rise of deep learning, it has achieved leapfrog progress. In terms of recognition accuracy, it has stagnated at around 75% and 85% at the sentence and character levels from traditional machine learning for a long time. , until now it is approaching 95% and 100%; in terms of generalization ability, a deep network model with sufficient complexity is enough to learn the characteristics of image documents of different qualities in different scenarios, and achieve true generalization, while in contrast In contrast, traditional machine learning has strict requirements on the document quality and scenarios to be recognized. Regardless of traditional machine learning or deep learning, OCR still follows t...

AI Technical Summary

Problems solved by technology

With the high complexity of its network model, deep learning can directly process some simple noise pictures, but preprocessing work is still required

Images