English Word Image Recognition via 1D Convolution and Bidirectional LSTM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current English word recognition methods require expensive hardware and computing systems, especially when using deep networks with LSTM, making real-time recognition challenging on platforms without GPUs, and leading to recognition errors, slow operation, or crashes.
Innovation Solution
An English word image recognition method that uses a series of one-dimensional convolutional neural network operations and bidirectional LSTM networks to generate feature maps, allowing for probabilistic string recognition and reducing the need for high-end hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep network and LSTM method are used to recognize English words, then recognition accuracy is improved, but hardware cost and system complexity increase
Solution Approach 1:
The patent segments the recognition process into distinct layers: feature extraction layer (generating feature maps from image), sequence modeling layer (using bidirectional LSTM to process character sequences), and recognition layer (outputting final word recognition results). This segmentation allows each layer to be optimized independently, reducing overall hardware requirements while maintaining accuracy
Solution Approach 2:
The patent introduces feature maps as an intermediary representation between the input image and the LSTM processing. These feature maps condense visual information into a compact format that is more efficient for subsequent processing, reducing the computational burden on hardware while preserving recognition accuracy
2Productivity
If GPU is used to perform graphics operation for deep network recognition, then real-time recognition is achieved, but hardware cost increases
Solution Approach 1:
The patent replaces the need for GPU-based parallel processing with a streamlined neural network architecture that can be efficiently executed on standard CPUs. The simplified feature extraction and processing pipeline eliminates the requirement for specialized graphics hardware while maintaining real-time performance
Solution Approach 2:
The patent changes the computational parameters by using one-dimensional convolution operations instead of traditional two-dimensional convolutions, and by optimizing the feature map dimensions. These parameter changes reduce the computational complexity from O(n²) to O(n), enabling real-time recognition on standard hardware without GPUs
3Device complexity
If conventional recognition method recognizes each character one by one, then hardware requirements are reduced, but recognition errors increase
Solution Approach 1:
The patent introduces feature maps as an intermediary that captures spatial relationships between characters before sequence processing. This intermediary representation preserves contextual information that would be lost in character-by-character processing, reducing recognition errors while maintaining hardware efficiency
Solution Approach 2:
The bidirectional LSTM network implements feedback mechanisms by processing sequences in both forward and backward directions, allowing the model to incorporate contextual information from both past and future characters when predicting each character. This feedback loop significantly reduces recognition errors compared to unidirectional or independent character processing
Data Source
AI summary
The invention provides an English word image recognition method, mainly loading a to-be-recognized image and performing a one-dimensional convolutional neural network operation and a fully connected operation processing to generate a feature map, outputting the feature map by a bidirectional long short-term memory (LSTM) network and performing a fully connected operation to generate a feature map, then performing a probability recognition and outputting a probabilistic string, and then recognizing the probabilistic string and outputting a word recognition result to solve the problem of producing a large amount of operation in the conventional two-dimensional recognition operation, thereby achieving efficacies of reducing costs of recognition equipment and enabling fast and accurate recognition.