Handwriting Recognition via Ink Segmentation and Cutting Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current handwriting recognition methods, particularly those using the CTC approach, face significant challenges in processing complex logographic writing systems like Chinese, Japanese, and Korean due to the large number of characters and denser ink paths, leading to substantial processing time and resource requirements, making real-time recognition difficult on both powerful servers and handheld devices.
Innovation Solution
A method that involves identifying consecutive ink segments and cutting points in user handwritten input, generating ink point feature vectors, and applying them to a recurrent neural network (RNN) classifier, with filtered outputs processed through a connectionist temporal classification (CTC) output layer to determine character probabilities, optimizing processing efficiency by reducing the number of inputs to the CTC layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the CTC approach is used to recognize logographic characters, then the system can handle complex writing systems, but the processing time and resource requirements become substantial
Solution Approach 1:
The patent segments the dense ink path of logographic characters into multiple strokes by detecting stroke endpoints and junction points. Each stroke is processed separately through the neural network, reducing the computational burden compared to processing the entire dense ink sequence at once. This segmentation enables the CTC approach to handle logographic characters efficiently.
Solution Approach 2:
The patent extracts key features from the ink data, specifically identifying stroke endpoints and junction points as critical features. By focusing on these extracted features rather than processing all ink points, the system reduces processing time while maintaining recognition accuracy for logographic characters.
2Measurement precision
If the CTC approach processes all ink points in the sequence, then comprehensive character recognition is achieved, but the processing becomes computationally expensive
Solution Approach 1:
The patent extracts only the essential features from the ink sequence - specifically stroke endpoints and junction points - rather than processing all ink points. This extraction maintains recognition accuracy by focusing on structurally important points while dramatically reducing the computational resources required.
Solution Approach 2:
The patent creates a simplified representation or 'copy' of the original ink data by identifying and processing only key stroke features. This copied representation contains the essential information needed for accurate character recognition but requires significantly fewer computational resources to process.
3Loss of information
If the system processes dense ink paths of logographic characters, then complete character information is captured, but the processing time increases substantially
Solution Approach 1:
The patent extracts the essential structural information from dense ink paths by identifying stroke endpoints and junction points. These extracted features capture the complete character information needed for recognition while reducing processing time by eliminating redundant ink point data.
Solution Approach 2:
The patent segments dense ink paths into discrete strokes based on endpoint and junction point detection. This segmentation preserves complete character information by maintaining the structural relationships between strokes while reducing processing time by handling each segment separately rather than as one continuous dense sequence.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and method of recognizing a sequence of characters of a user handwritten input. The method comprises: retrieving data representing the user handwritten input comprising a sequence of ink points (Pt) of the characters; identifying consecutive ink segments (Sn) comprising respective subsets of the sequence of ink points and at least one cutting point (CPt) at an edge of each ink segment; generating a sequence of ink point feature vectors (Vt) based on the sequence of ink points; applying the sequence of ink point feature vectors (Vt) as inputs to a recurrent neural network-based, RNN, classifier; filtering outputs (Ht) of the RNN classifier and transferring only the outputs of the RNN classifier corresponding to the at least one cutting point to a connectionist temporal classification, CTC, output layer; and determining (S170) probabilities of each character of the pre-defined writing system occurring at each consecutive cutting point.