Mobile Text Recognition with Coordinate-Based Single-Pass NPU Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text recognition solutions in mobile terminals suffer from resource waste, excessive calculation overheads, and long response times due to the use of sliding windows, which require multiple invocations of neural networks and inefficient parallel processing of text lines.
Innovation Solution
A method that scales the input image to detect text line areas, determines corresponding coordinates, and processes these areas concurrently using a recognition model without sliding windows, optimizing NPU usage and reducing redundant calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sliding window technology is used to partition the raw image into sub-images, then text detection coverage is improved, but calculation overhead and resource consumption increase significantly
Solution Approach 1:
The patent extracts only the necessary text line regions from the original image using coordinate information, eliminating the need for sliding window partitioning. By directly obtaining text line images through coordinate mapping, the system avoids redundant calculations while maintaining complete text detection coverage.
Solution Approach 2:
The patent performs preliminary coordinate transformation and text line extraction before neural network processing. By pre-processing the image to identify and isolate text regions using coordinate information, the system prepares data in advance, reducing the need for repeated neural network invocations and sliding window operations.
2Reliability
If sliding window is used with step 4 or 8 for text line selection, then edge context is preserved, but repeated calculation occurs 4 or 8 times causing resource waste
Solution Approach 1:
The patent extracts text line regions directly using coordinate information without sliding window partitioning. By mapping coordinates from the scaled image back to the original image and extracting only the necessary text regions, the system eliminates repeated calculations while preserving edge context through precise coordinate-based selection.
3Reliability
If sliding window partitioning is applied to the raw image, then text line detection is achieved, but extra merge operation is required and resource waste occurs
Solution Approach 1:
The patent extracts text line images directly through coordinate mapping without sliding window partitioning. By obtaining text regions through coordinate information and directly feeding them to the neural network, the system eliminates the need for subsequent merge operations while maintaining detection accuracy.
4Reliability
If NPU is invoked multiple times for processing sub-images, then complete text recognition is achieved, but response time becomes excessively long
Solution Approach 1:
The patent performs preliminary coordinate transformation and text line extraction before neural network processing. By pre-processing the image to identify and isolate text regions using coordinate information, the system prepares data in advance, reducing the need for repeated neural network invocations and sliding window operations.
Solution Approach 2:
The patent extracts only the necessary text line regions from the original image using coordinate information, eliminating the need for sliding window partitioning. By directly obtaining text line images through coordinate mapping, the system avoids redundant calculations while maintaining complete text detection coverage.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of this application provide a text recognition method. The method includes: scaling a to-be-recognized image based on a first scale ratio; determining first coordinate information corresponding to a text line area in the scaled to-be-recognized image; determining, based on the first scale ratio, second coordinate information corresponding to the first coordinate information, where the second coordinate information is coordinate information of the text line area in the to-be-recognized image; and performing character recognition on text line images corresponding to the second coordinate information by using a recognition model, and determining text line content corresponding to the text line images, where the to-be-recognized image includes the text line images. The embodiments of this application avoids resource waste caused by invoking an NPU for a plurality of times in sub-images obtained through partitioning by using a sliding window. Based on actual requirements for low response delay and low power consumption of a mobile terminal application and NPU technical specifications, an advantage of a device-end AI dedicated chip can be fully utilized to concurrently process a plurality of text lines at a time. Therefore, NPU usage and user experience of an OCR product are improved.