Character String Recognition With Additional Region Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing character recognition technologies, such as CRAFT and Star-Net, face challenges in accurately identifying characters due to varying probabilities of character existence, leading to recognition failures, especially when characters are partially obscured or not adjacent.
Innovation Solution
The proposed method involves determining additional character regions by analyzing pixel columns perpendicular to the character arrangement direction, using a machine learning model to enhance character region estimation, and clustering to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If character recognition is performed based on probability images from models like CRAFT, then the processing can be automated and scaled, but recognition accuracy deteriorates when characters are partially obscured or not adjacent
Solution Approach 1:
The patent segments the character recognition task into two independent stages: first, detecting character regions using probability images from models like CRAFT; second, recognizing characters within those segmented regions using optical character recognition. This segmentation allows each stage to optimize independently, maintaining automation while improving accuracy for challenging cases like obscured or non-adjacent characters.
Solution Approach 2:
The patent performs preliminary character region detection and segmentation before the actual character recognition. By pre-identifying the regions where characters are likely to exist using probability images, the system prepares the data in advance, allowing the subsequent recognition step to focus only on relevant areas, thereby improving accuracy without sacrificing automation.
2Productivity
If character region detection is performed using probability images, then processing efficiency is improved, but recognition accuracy deteriorates when characters have low existence probability in certain portions
Solution Approach 1:
The patent divides the processing into efficient region detection using probability images and subsequent precise character recognition within segmented regions. This segmentation maintains the efficiency benefits of probability-based detection while enabling accurate recognition even when characters have low existence probability in certain portions, as the system focuses recognition efforts on the segmented character regions rather than the entire image.
Solution Approach 2:
The patent applies local quality by treating different regions of the image differently: using probability images for general region detection (efficient processing) and then applying detailed character recognition only within the segmented character regions (high accuracy). This allows the system to optimize for efficiency globally while optimizing for accuracy locally where characters are detected.
3Measurement precision
If the probability of character existence is calculated for each pixel, then detailed character detection is achieved, but processing complexity increases
Solution Approach 1:
The patent segments the processing into a first stage that uses pre-computed probability images (from models like CRAFT) to detect character regions, and a second stage that performs character recognition within those regions. This segmentation reduces the complexity of the overall system by leveraging existing probability image generation techniques while adding only the necessary region-based recognition component, achieving detailed detection without proportionally increasing processing complexity.
Solution Approach 2:
The patent introduces probability images as an intermediary representation between the raw image and the final character recognition. Instead of directly analyzing each pixel for character existence (which would be complex), the system uses probability images as a mediator that encapsulates character presence information in a compressed, pre-processed format, reducing processing complexity while maintaining detection precision.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The information processing device obtains a character string image which includes a plurality of characters, and which includes the characters arranged in an arrangement direction, obtains a probability image representing a probability of an existence of a character in each of pixels included in the character string image, obtains a plurality of character regions in which the characters are estimated to respectively exist in the character string image based on the probability image, obtains an additional character region which is located in the character string image, and which does not overlap the plurality of character regions based on a determination result on whether or not a pixel of a non-background color exists in a direction perpendicular to the arrangement direction at every position in the arrangement direction in the character string image, and recognizes the plurality of characters from the character regions and the additional character region.