Deep Learning Stamp Localization and Text Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document processing systems face challenges in accurately localizing and removing stamps or seals from documents due to variations in color, shape, and placement, which can interfere with text recognition and system accuracy.
Innovation Solution
A deep learning-based model using a combination of color filter methods and grayscale models to predict stamp location and segment stamp pixels, employing a line mask for further segmentation of foreground text lines, thereby enabling efficient stamp removal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If color filter methods are used to detect stamps, then stamp localization accuracy is improved for color stamps, but the method fails when stamps are grayscale or same color as text
Solution Approach 1:
The patent segments the stamp detection task into two independent channels: a color filter channel for detecting color stamps and a grayscale channel for detecting grayscale or same-color stamps. This segmentation allows each channel to specialize in specific stamp types, resolving the contradiction between accuracy for color stamps and adaptability to grayscale stamps.
Solution Approach 2:
The patent creates a universal stamp detection system by combining two channels that can handle different stamp types. The first channel handles color stamps while the second channel handles grayscale and same-color stamps, making the overall system universally applicable to all stamp types regardless of color characteristics.
2Productivity
If stamp removal is performed without line mask segmentation, then processing speed is maintained, but text line recognition accuracy deteriorates due to overlapping text
Solution Approach 1:
The patent segments the text processing into two stages: first detecting text lines using the line mask channel, then performing stamp removal only in regions where stamps overlap with text. This segmentation allows the system to maintain processing speed by avoiding unnecessary text line detection while ensuring accurate text recognition where needed.
Solution Approach 2:
The patent applies local quality by differentiating the processing approach based on spatial location. The line mask channel identifies regions where text lines are present, and stamp removal is selectively applied only in those overlapping regions. This localized approach ensures text recognition accuracy is maintained where stamps interfere with text while avoiding unnecessary processing elsewhere.
3Measurement precision
If deep learning model with two channels is implemented, then stamp localization and text segmentation accuracy is improved, but device complexity increases
Solution Approach 1:
The patent segments the deep learning model into two independent channels that can be trained and processed separately. Each channel handles specific aspects of stamp detection, allowing for simplified individual channel architectures while maintaining overall system accuracy. This segmentation reduces the complexity burden on any single channel.
Solution Approach 2:
The patent merges the outputs of two separate channels into a unified stamp detection result. By combining the color filter channel and grayscale channel results, the system achieves high accuracy for all stamp types while distributing the computational load across two simpler channels rather than requiring one complex all-purpose channel.
Data Source
AI summary
In a form recognition system, a deep learning system may be trained to perform stamp localization for stamp removal to facilitate form recognition. In embodiments, a stamp mask identifies locations of stamps or seals on forms, and a line mask identifies pixels of the stamps. Where a stamp or seal overlaps with underlying text on a form, and a color or grayscale of the stamp or seal is sufficiently similar to that of the underlying text, a combination of the stamp mask and the line mask may enable removal of the stamp or seal without degrading the underlying text in the form, and facilitate form recognition.