Enhancing Content Based Image Retrieval Through Deep Convolutional Feature Extraction Combined with Ensemble Learning Classifiers

Abstract

Searching a large image database by keywords alone rarely gives satisfactory results, mainly because a keyword can never fully describe what is actually inside a picture. This gap between what a machine reads as metadata and what a human sees is usually called the semantic gap, and Content Based Image Retrieval, or CBIR, exists to close it. This paper builds on an earlier VGG16 and XGBoost based CBIR pipeline and extends it by placing the ensemble classification stage at the centre of the study. VGG16 is used purely as a feature extractor, and the resulting feature vectors are passed to a gradient boosted ensemble for prediction and similarity ranking. The system was implemented and tested on 1000 images drawn from the public Kaggle Flowers Recognition dataset, across five categories, daisy, dandelion, rose, sunflower and tulip. The trained model reached a training accuracy of 1.00 but a validation accuracy of only 0.72, a gap pointing to overfitting on the high dimensional 25088 length feature vectors relative to the dataset size, discussed openly here as one of the more useful findings of the study. The paper reviews thirty related works and closes with a comparison table and directions for future work.