Unsupervised Word Image Clustering Using Zero-Sum CNN Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Robust character segmentation for Indic scripts remains a challenge, leading to poor accuracy in Optical Character Recognition (OCR) systems, which requires large amounts of training data and computing resources, and is time-consuming, especially for unsupervised word image clustering.
Innovation Solution
An unsupervised word image clustering method using an untrained convolution neural network architecture with randomly initialized filters constrained to zero sum, applying ReLU non-linearity and pooling operations to increase position invariance, followed by graph clustering based on feature vectors extracted from word images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised feature based word clustering is used, then clustering accuracy is improved, but training time and computing resources increase significantly
Solution Approach 1:
The patent applies preliminary action by using unsupervised feature extraction with pre-trained convolutional neural network filters (e.g., from ImageNet) to extract meaningful word image features without requiring supervised training on word-specific data. This preliminary feature extraction enables subsequent clustering to be performed efficiently on already-processed features, avoiding the need for time-consuming supervised training while maintaining clustering accuracy.
2Measurement precision
If large amount of training data is used for supervised word clustering, then clustering performance is improved, but data requirements and storage increase
Solution Approach 1:
The patent applies self-service by using pre-trained convolutional neural network filters that have already learned general image features from large-scale datasets like ImageNet. These filters automatically adapt to word image features through unsupervised learning, allowing the system to extract meaningful features without requiring large quantities of labeled word training data. The pre-trained filters serve themselves by transferring their learned capabilities to the word clustering task.
3Adaptability or versatility
If randomly initialized deep networks are used for word recognition, then object recognition capability is improved, but shape feature extraction accuracy deteriorates
Solution Approach 1:
The patent applies parameter changes by modifying the initialization and training approach of convolutional neural network filters. Instead of using completely random initialization, the system uses pre-trained filters from general image recognition tasks and fine-tunes them through unsupervised learning on word images. This parameter adjustment allows the network to maintain its object recognition capability while improving shape feature extraction accuracy for word-specific characteristics.
Data Source
AI summary
The present application provides a method and system for unsupervised word image clustering, comprises capturing one or more image wherein the one or more image comprises at least one word images. Extracting at least one feature vector using an untrained convolution neural network architecture, wherein the convolution filters are initialized by random filter based deep learning techniques using Gaussian random variable with zero mean and unit standard deviation, and wherein the convolution filters are constrained to sum to zero. The extracted feature vectors are used for clustering, wherein clustering is performed in two stages. First stage includes clustering word images which are similar using a graph connected component. Second stage clustering includes clustering a remaining word images which are not clustered during the first stage by evaluating the remaining images against the clusters formed during the first stage and assigning them to clusters based on the evaluation.