Sequential Deformation Module for Arbitrary-Shaped Text Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current scene text detection technologies using deep neural networks face challenges in modeling large geometric deformations and require complex postprocessing steps, making it difficult to detect arbitrary-shaped text areas, such as curved text, under uncontrolled conditions.
Innovation Solution
A method and system incorporating a Sequential Deformation Module (SDM) and an Auxiliary Character Counting (ACC) supervised learning mechanism, which includes a feature extraction module, a sequential deformation module, and an object detection baseline network, performing iterative sampling and feature aggregation to adapt to text area deformations without the need for complicated postprocessing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If component level based object detection approach is used to detect text components, then detection capability for arbitrary-shaped text areas is improved, but postprocessing complexity increases and end-to-end integral optimization training becomes inconvenient
Solution Approach 1:
The patent merges the text component detection and text instance recognition into a unified end-to-end training framework. The sequential deformation module processes text components through multiple iterations, and the auxiliary character counting network integrates character-level predictions with instance-level output, eliminating the need for separate postprocessing steps while maintaining capability for arbitrary-shaped text detection
Solution Approach 2:
The proposed method achieves multi-functionality by simultaneously performing text component detection, character sequence prediction, and text instance recognition within a single integrated model. The sequential deformation module serves both as a feature extraction mechanism and as a deformation adaptation mechanism, while the auxiliary character counting network provides both character-level and instance-level outputs
2Device complexity
If instance level based object detection approach is used, then end-to-end training is simplified, but modeling capability for large geometric deformation is insufficient
Solution Approach 1:
The patent introduces a sequential deformation module that dynamically adapts to text geometry through iterative sampling. Instead of using a fixed receptive field, the model performs multiple iterations of sampling and deformation prediction, allowing the feature extraction process to adapt to various text shapes and orientations while maintaining end-to-end training simplicity
Solution Approach 2:
The sequential deformation module performs preliminary deformation modeling through iterative sampling before the final text instance recognition. By pre-computing deformation patterns and feature transformations in the sequential manner, the model prepares adapted feature representations that accurately capture arbitrary-shaped text areas while maintaining a simple training framework
3Productivity
If fixed receptive field of convolutional neural network is used, then computational efficiency is maintained, but capability to model large geometric deformation is constrained
Solution Approach 1:
The patent segments the text detection task into multiple iterative sampling steps within the sequential deformation module. Instead of using a single large receptive field, the model divides the deformation modeling into sequential iterations, each handling a portion of the geometric transformation. This segmentation allows the model to maintain computational efficiency while achieving comprehensive deformation modeling capability
Solution Approach 2:
The sequential deformation module introduces a temporal/iterative dimension to the traditional spatial receptive field approach. By adding the iteration step as an additional dimension, the model transforms the fixed 2D receptive field problem into a multi-step sequential process, enabling large geometric deformation modeling without proportionally increasing computational cost
Data Source
AI summary
A method and a system for detecting a scene text may include extracting a first feature map for a scene image input based on a convolutional neural network, and delivering the first feature map to a sequential deformation module; obtaining sampled feature maps corresponding to sampling positions by performing iterative sampling for the first feature map, obtaining a second feature map by performing a concatenation operation in deep learning according to a channel dimension for the first feature map and the sampled feature maps; obtaining a third feature map by performing a feature aggregation operation for the second feature map in the channel dimension, and delivering the third feature map to the object detection baseline network; and performing text area candidate box extraction for the third feature map and obtaining a text area prediction result as a scene text detection result through regression fitting.