Neural Network Table Partition Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in identifying the location of partitions in tables within images, especially when the table structure is unknown or semi-structured, making it difficult to extract information accurately.
Innovation Solution
A neural network is trained using normalized images to detect partition information, representing it as two values per row and column portion, indicating the presence and position of partitions, allowing for successful recognition and extraction of information from tables without prior knowledge of the table structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing methods are used to detect table partitions, then the processing speed may be maintained, but the accuracy of partition detection deteriorates when table structures are unknown or semi-structured
Solution Approach 1:
The patent transforms the input image into a feature map through multiple convolutional layers, changing the parameter representation from raw pixel values to extracted features. This transformation enables the network to adapt to various table structures while maintaining high detection accuracy, as the feature extraction process learns to identify partition patterns regardless of the specific table format
Solution Approach 2:
The patent introduces depth dimension through multiple convolutional layers and channels, transforming the 2D image data into 3D feature volumes. This dimensional expansion allows the system to capture complex partition patterns and structural relationships that cannot be detected by traditional 2D processing methods, improving both accuracy and adaptability
2Adaptability or versatility
If a neural network is trained to detect partition information, then the ability to handle various table formats improves, but the training time and computational resources increase
Solution Approach 1:
The patent divides the image processing task into multiple stages through sequential convolutional layers, each responsible for detecting specific features at different levels of abstraction. This segmentation allows the network to learn hierarchical patterns efficiently during training, reducing the overall training time while maintaining high adaptability to various table formats
Solution Approach 2:
The patent designs a universal neural network architecture that can handle various table formats through shared convolutional layers and pooling operations. This multi-functional design allows the same network structure to adapt to different table types without requiring separate training for each format, significantly reducing training time while maintaining versatility
3Stability of the object's composition
If the image is normalized to a predetermined size for neural network training, then the training consistency improves, but the processing of images with various sizes becomes more complex
Solution Approach 1:
The patent performs image normalization and resizing as a preliminary action before feeding the data into the neural network. By standardizing all input images to a predetermined size during the preprocessing stage, the training consistency is improved, and the network can focus on learning partition patterns without being distracted by variations in image dimensions
4Difficulty of detecting and measuring
If explicit dividing lines are used to define table partitions, then the partition location is easier to detect, but many tables rely on implicit partitions based on content distribution without explicit lines
Solution Approach 1:
The patent applies local quality analysis by using convolutional kernels that scan local regions of the image to detect partition patterns. The network learns to identify both explicit dividing lines and implicit partition boundaries by analyzing local content distribution, alignment patterns, and spacing characteristics, enabling it to handle both explicit and implicit partition types effectively
Data Source
AI summary
Methods and apparatus for training neural networks to identify information table partitions are described. Also described are methods and apparatus of using a trained neural network to process an image and provide partition information in an easy to use format. The format of the partition information is one which is simple to interpret, easy to communicate and uses values which facilitate successful training and recognition of partitions in tables whether the partitions be implicitly defined by data arrangement or explicitly define using lines. An image is treated as including a predetermined number of row and column portions. The neural network generates for each predetermined portion a partition present indicator value and a partition location value. The partition present value in some embodiments is a value in the range of 0 to 1 and the partition location value in some embodiments is a value in the range of −1 to +1.