Res2NeXt Backbone with Mixed Pooling for Scene Text Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional scene text detection methods face challenges due to variations in text characteristics and external conditions like uneven illumination, leading to incomplete and inaccurate detection.
Innovation Solution
A method using a rich feature structure network (Res2NeXt) with mixed pooling, which acquires and fuses single-layer feature maps, and performs progressive scale expansion to enhance text detection precision by capturing context information and distance dependencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional network structures are used for scene text detection, then the detection process is simpler, but the detection precision and completeness are insufficient
Solution Approach 1:
The patent segments the feature extraction process into multiple parallel convolutional branches with different kernel sizes (3x3, 5x5, 7x7, 9x9), each capturing text features at different scales. This segmentation allows the network to handle diverse text sizes and shapes independently, improving detection precision without requiring a single complex monolithic structure
Solution Approach 2:
The patent creates a composite network structure by combining multiple convolutional branches with different kernel sizes, ReLU activation functions, and Batch Normalization layers. This composite architecture integrates the strengths of different component structures, achieving superior text detection performance compared to conventional single-structure networks
2Adaptability or versatility
If the network structure is simplified, then the device complexity is reduced, but the ability to handle text variations and external conditions deteriorates
Solution Approach 1:
The patent applies local quality by using convolutional kernels of different sizes (3x3 for small text, 5x5 for medium text, 7x7 for large text, 9x9 for very large text) in different branches. Each branch is optimized for specific text scale characteristics, allowing the network to adapt to local variations in text size and shape while maintaining an organized modular structure
Solution Approach 2:
The patent creates a universal network structure that can detect text across multiple scales and orientations simultaneously. The multi-branch architecture with different kernel sizes enables the same network to handle various text types (small signs, medium labels, large banners, oversized displays) and external conditions (illumination variations, occlusions) through a single unified system
3Measurement precision
If mixed pooling and feature fusion are added to capture context information, then text detection precision improves, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary action by applying max pooling and average pooling operations at multiple intermediate stages of the network, before the final detection layer. This early and intermediate pooling captures contextual information and reduces feature map dimensions in advance, decreasing the computational energy required for subsequent processing while improving detection precision
Data Source
AI summary
A method and a system for detecting a scene text are provided. The method includes: acquiring a scene text picture in a preset manner, pre-processing the acquired scene text picture, detecting the pre-processed scene text picture with a training model for scene text detection, and acquiring a detection result. Accordingly, the method and the system have an original PSENet (Progressive Scale Expansion Network) backbone network ResNet (Deep Residual Network) is replaced with a rich feature structure network (i.e. Res2NeXt (Combination of Res2Net and ResNeXt)) to improve a network feature extraction capability, thereby increasing a text detection precision of the network; mixed pooling is added at an appropriate location in the backbone network to acquire useful context information by performing pooling operations of different kernel shapes and capture long and short distance dependency relationships between different locations, thereby further increasing the text detection precision of the network.