Image Recognition Using CNN Feature Clustering and Probability Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition methods are prone to noise and inaccuracy due to reliance on web page text data and large image databases, often mislabeling images with similar content but different semantics, resulting in imprecise and wide-ranging content labels.
Innovation Solution
The method employs convolutional neural network (CNN) features for image recognition, involving two rounds of similarity matching and calculation to select target image classifiers and determine class and entity labels, improving precision by converting CNN features into feature vectors and calculating probability values through vector inner products and exponential mapping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If web page text data and large image databases are used for image recognition, then the coverage of recognized content is wide, but the precision of recognition deteriorates due to noise and mislabeling
Solution Approach 1:
The patent segments the image database into multiple class clusters, each corresponding to a specific class label (e.g., animal, plant, building). This segmentation allows the system to narrow down the search scope from the entire large database to relevant clusters only, improving precision while maintaining wide coverage through multiple labeled categories.
Solution Approach 2:
The patent extracts CNN features from images and extracts relevant class clusters based on probability values, separating the recognition process from noise-filled web page text data. This extraction of essential feature data eliminates harmful noise while preserving the core recognition capability.
2Ease of operation
If image databases are searched for similar data to determine content labels, then the recognition process is simplified, but the accuracy deteriorates because pictures with similar content but different semantics are likely to be found
Solution Approach 1:
The patent performs preliminary action by pre-building class clusters from the image database before the actual recognition process. Each class cluster is pre-labeled with class labels and entity labels based on CNN features. During recognition, the system only needs to search for similar images within relevant pre-clustered groups, maintaining simplicity while improving accuracy through pre-organized semantic categories.
Solution Approach 2:
The patent introduces class clusters as an intermediary between the image database and the recognition result. Instead of directly searching the entire database for similar images, the system first identifies relevant class clusters based on CNN features, then searches within those clusters. This intermediary structure prevents finding images with similar visual content but different semantics.
3Measurement precision
If CNN features are converted into feature vectors and probability values are calculated through vector inner products, then the precision of image recognition is improved, but the computational complexity increases
Solution Approach 1:
The patent applies partial action by calculating probability values only for relevant class clusters identified through CNN feature matching, rather than computing probabilities for all possible classes in the database. This selective calculation maintains high precision through vector inner products while reducing overall computational complexity by limiting the scope to relevant clusters only.
Data Source
AI summary
The present disclosure provides a method performed at a computing device for recognizing an image's content. The method includes: extracting one or more features from an image to be recognized; comparing the features of the image with a set of image classifiers and obtaining a probability value for each image classifier; selecting, from the set of image classifiers, at least one image classifier as a target image classifier according to the probability value of the image classifier; determining a degree of similarity between each image and the image to be recognized, and selecting, from the target image classifier, multiple images as target images when their respective degrees of similarity with the image to be recognized exceed a predefined threshold; and labeling, the image to be recognized according to a class label corresponding to the target image classifier and entity labels corresponding to the target images in the target image classifier.