Ensemble CNN Text Detection for High-Entropy Product Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Extracting text from product images in e-commerce settings is challenging due to varied text shapes, sizes, and the presence of multiple visual objects, which affects the accuracy and coverage of brand and product information extraction.
Innovation Solution
An end-to-end text detection strategy combining segmentation algorithms and ensemble models, including convolutional neural networks (CNNs) for text detection and recognition, with selective non-maximal suppression to enhance accuracy and coverage in high entropy images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text extraction methods are used on product images, then the process is simpler, but the accuracy and coverage of text detection deteriorates due to varied text shapes, sizes, and multiple visual objects
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into multiple specialized models: a first CNN model for detecting text in certain orientations and a second CNN model for detecting text in other orientations. Each model is trained to handle specific text configurations, and their outputs are combined to achieve comprehensive text detection. This segmentation approach improves detection accuracy for varied text shapes and sizes while managing system complexity through modular model design.
2Adaptability or versatility
If multiple types of text in various shapes and sizes are detected, then text coverage improves, but detection difficulty increases due to the presence of many visual objects
Solution Approach 1:
The patent implements local quality by training different CNN models with specialized architectures and parameters optimized for specific text characteristics. The first model is configured for certain text orientations and shapes, while the second model is configured for different orientations and shapes. Each model develops specialized feature extraction capabilities tailored to its specific text type, enabling high adaptability across diverse text formats while managing detection complexity through specialized local processing.
3Loss of information
If text is extracted from high entropy images with compacted text and objects, then information coverage improves, but detection precision requirements increase
Solution Approach 1:
The patent merges the outputs of multiple specialized CNN models to achieve comprehensive text detection in high entropy images. By combining the detection results from the first model (handling certain text types) and the second model (handling other text types), the system achieves complete information coverage without requiring any single model to achieve impossible precision levels. The merging process integrates detections from multiple sources, reducing the precision burden on individual models while maintaining overall high detection accuracy.
Data Source
AI summary
Segmentation first breaks the images into segments or regions, with the segments of the region having text or symbols. The segmented image is separately applied to two different CNN-based models. Each model produces text boxes where potential text might exist. Then, a selective NMS algorithm is applied to the output of each model to produce a final group of text regions. These text regions are analyzed and actions are taken.