Topic-Modeling Guided Semantic Clustering for Enhancing CNN-Based Image Classification Using Scale-Invariant Feature...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are powerful for image classification.
• Traditional CNN training can benefit from incorporating semantic information.
• Image data often possesses inherent semantic structures that can be leveraged.

Purpose of the Study:

• To propose a topic-modeling guided framework to enhance CNN image classification.
• To introduce semantic clustering before CNN training.
• To evaluate the effectiveness of different feature extraction and topic modeling strategies.

Main Methods:

• Images processed using Scale-Invariant Feature Transform (SIFT) or Block Gabor Filtering (BGF) for feature descriptors.
• K-means clustering and Bag of Words histograms to build visual vocabularies and represent images.
• Latent Dirichlet Allocation (LDA) for topic modeling and semantic clustering.
• Training cluster-specific CNN models (AlexNet, GoogLeNet, ResNet variants).
• Topic-guided integration strategies: Maximum Proportion Topic (MPT) and Weight Proportion Topic (WPT).

Main Results:

• Both SIFT-based and BGF-based pipelines outperformed non-clustered CNNs and baseline methods.
• The SIFT pipeline achieved a peak accuracy of 95.24% using the MPT strategy.
• The BGF pipeline reached 93.76% accuracy with the WPT strategy.
• Semantic clustering via topic modeling significantly improved CNN classification performance.

Conclusions:

• Topic modeling guided semantic clustering is an effective method for enhancing CNN image classification.
• The proposed framework offers a substantial improvement over conventional approaches.
• Specific feature extraction and topic integration strategies yield varying but improved results.