On-device Document Auto-rotation via Segmented Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in efficiently correcting the orientation of document images on mobile devices due to memory and performance limitations, which hinders effective optical character recognition (OCR) and requires slow testing of each possible orientation.
Innovation Solution
An AI-augmented method using a small, fast, single-layer neural network on a user device to detect document orientation by converting images to monochrome, segmenting into bounding boxes, filtering non-textual snippets, and processing textual snippets to determine the correct orientation and automatically rotate the image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large, high-quality neural network is used for document orientation detection, then the accuracy of orientation detection is improved, but the memory and performance requirements exceed mobile device capabilities
Solution Approach 1:
The patent divides the document image into multiple smaller patches or regions. Instead of processing the entire document image at once with a large neural network, the system processes smaller patches individually or in groups. This segmentation approach reduces the computational burden and memory requirements while maintaining orientation detection accuracy through aggregation of results from multiple patches.
2Manufacturing precision
If server-based OCR rotation correction is used, then the orientation correction quality is improved, but the processing time increases significantly
Solution Approach 1:
The patent performs preliminary orientation detection and correction on mobile devices before transmitting documents to the server. By detecting document orientation using the segmented image approach and pre-correcting images on-device, the system reduces the need for time-consuming server-based rotation correction, thereby decreasing overall processing time while maintaining correction quality.
3Manufacturing precision
If OCR-based rotation correction is used, then the orientation correction capability is improved, but the processing speed decreases due to testing each possible orientation
Solution Approach 1:
The patent extracts the orientation detection task from the full OCR processing pipeline. By using a dedicated, lightweight neural network specifically for orientation detection on mobile devices, the system separates this function from the main OCR processing that occurs on the server. This extraction enables faster orientation detection without compromising the overall OCR correction capability.
Data Source
AI summary
An auto-rotation module having a single-layer neural network on a user device can convert a document image to a monochrome image having black and white pixels and segment the monochrome image into bounding boxes, each bounding box defining a connected segment of black pixels in the monochrome image. The auto-rotation module can determine textual snippets from the bounding boxes and prepare them into input images for the single-layer neural network. The single-layer neural network is trained to process each input image, recognize a correct orientation, and output a set of results for each input image. Each result indicates a probability associated with a particular orientation. The auto-rotation module can examine the results, determine what degree of rotation is needed to achieve a correct orientation of the document image, and automatically rotate the document image by the degree of rotation needed to achieve the correct orientation of the document image.