App Icon Matching Using OCR and Text Categories for Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing application matching methods fail to accurately calculate similarity due to the neglect of semantic information in application icons' text, leading to errors in matching results.
Innovation Solution
An application matching method that recognizes and categorizes text in icons using OCR and TextCNN, determines similarity based on text categories, and adjusts calculations accordingly to include or exclude text information, using models like FCN, TextCNN, word2vec, ORB, and SiameseNet for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If text in icons is treated only as image information during image matching, then the matching process is simple, but the semantic information is not fully utilized leading to inaccurate matching results
Solution Approach 1:
The patent segments the icon information into two distinct parts: image information (visual features) and text information (semantic features). Image matching is performed on the visual features while text matching is performed on the extracted text features. This segmentation allows each type of information to be processed independently with appropriate algorithms, resolving the contradiction by maintaining process simplicity while improving accuracy through comprehensive information utilization.
Solution Approach 2:
The patent merges the results of image matching and text matching through a weighted combination formula. The final similarity score combines both visual similarity and textual similarity, giving full weight to both image information and semantic information. This merging approach resolves the contradiction by integrating multiple information sources to achieve higher accuracy while maintaining a systematic and manageable process.
2Measurement precision
If text information from icons is used for similarity calculation, then semantic information is utilized, but the text may not be included in official application names or descriptions leading to additional processing complexity
Solution Approach 1:
The patent introduces an intermediary text extraction and processing module that bridges the gap between icon images and application metadata. This intermediary component extracts text from icons using OCR technology, processes the extracted text through cleaning and normalization, and then integrates it with official application names and descriptions. This intermediary layer resolves the contradiction by systematically handling the additional processing requirements while enabling accurate utilization of semantic information from icons.
3Measurement precision
If different policies are applied for different text categories, then the matching accuracy is improved, but the system complexity increases due to multiple processing paths
Solution Approach 1:
The patent applies local quality by treating different text categories with different processing policies. Specifically, promotional text in icons is removed before matching to avoid false positives, while version information text is retained and used to enhance matching accuracy. This localized differentiation resolves the contradiction by applying complexity only where necessary (differentiating text types) while maintaining simplicity in the overall framework through clear categorization rules.
Data Source
AI summary
An application matching method that includes: acquiring a plurality of icons comprising a first icon corresponding to a first application and a second icon corresponding to a second application, recognizing a text in at least one of the plurality of icons, determining a category of the text, determining a similarity between the first icon and the second icon based on a category determination result of the text, and determining whether the first application matches the second application based on the similarity.