Document Segmentation Neural Network with Atrous Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document segmentation methods based on deep learning are limited by the amount of computational resources, leading to insufficient clarity in labeling objects within documents.
Innovation Solution
An electronic device and method for document segmentation using a neural network model that includes a first model configured to perform upsampling, concatenation, inverted residual blocks, atrous convolution operations, and convolution operations to generate a segmented document, while consuming reduced computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If a convolution neural network with fewer convolution layers is used, then the amount of computational resources is reduced, but the clarity of labeling objects in the document is insufficient
Solution Approach 1:
The neural network is divided into an encoding network and a decoding network with multiple independent modules (first model, second model, third model). Each module performs specific functions such as feature extraction, upsampling, and segmentation. This modular segmentation allows the system to achieve high labeling clarity without requiring an excessive number of convolution layers, thus reducing computational resource consumption while maintaining precision.
2Manufacturing precision
If a convolution neural network with more convolution layers is used, then the clarity of labeling objects in the document is improved, but the amount of computational resources increases
Solution Approach 1:
The patent extracts and emphasizes the most critical feature extraction and processing modules, removing redundant convolution layers. The encoding network extracts key features, and the decoding network with its specific modules (upsampling, concatenation, atrous convolution) processes these features efficiently. This extraction of essential functions maintains high labeling clarity while significantly reducing the computational burden compared to deeper networks.
Solution Approach 2:
Instead of applying excessive convolution operations throughout the network, the patent applies atrous convolution with specific dilation rates (1, 2, 4) only at critical stages in the decoding network. This partial application of complex operations where most needed maintains segmentation precision while avoiding the computational cost of applying such operations uniformly across all layers.
3Measurement precision
If multiple feature maps with different dilation rates are processed, then the segmentation precision is improved, but the computational complexity increases
Solution Approach 1:
The patent merges multiple feature maps with different dilation rates (1, 2, 4) through concatenation operations in the decoding network. By combining these feature maps that capture different receptive fields and contextual information, the system achieves high segmentation precision. The merging happens at strategic points rather than processing each feature map independently through full convolutional paths, thus reducing overall computational complexity while maintaining precision.
Data Source
AI summary
An electronic device and a method for document segmentation are provided. The method includes: obtaining a first feature map and a second feature map corresponding to an original document; performing a first upsampling on the second feature map to generate a third feature map; concatenating the first feature map and the third feature map to generate a fourth feature map; inputting the fourth feature map to a first inverted residual block (IRB) and performing a first atrous convolution operation based on a first dilation rate to generate a fifth feature map; inputting the fourth feature map to a second IRB and performing a second atrous convolution operation based on a second dilation rate to generate a sixth feature map; concatenating the fifth feature map and the sixth feature map to generate a seventh feature map; performing a convolution operation on the seventh feature map to generate a segmented document.