Improving the Performance of CBIR Using XGBoost Classifier with Deep CNN-Based Feature Extraction

Abstract

The main challenge faced by CBIR systems is the semantic gap, namely the existence of semantic differences between low-level pixel images captured by machines and high-level semantics perceived by humans. This is caused by the CBIR system which is very dependent on the feature extraction used. The success of deep learning techniques, especially Convolutional Neural Networks (CNN) in solving the problem of computer vision applications has inspired researchers to overcome the problem of the semantic gap. Until now, what is known by researchers, Deep CNN research for CBIR uses softmax and SVM classifiers. However, the XGBoost classifier achieves extraordinary performance even though compared to SVM. By applying the advantages of Deep CNN techniques and XGBoost classifier, the Deep CNN model is proposed as a feature extractor and XGBoost to form a classification model replacing the softmax and SVM classifier. From the proposed Deep CNN model, two Fully Connected (FC) layers (FC1 and FC2) are taken as feature representations. In this study, Features Vector (FV) taken from the Deep CNN model that have been produced are used as extractor features (FV.FC1 and FV.FC2). The performance of the Deep CNN for CBIR tasks generated using softmax, SVM, and XGBoost classifier observed. The performance evaluated is accuracy, precision, recall, and f1-score. Based on experimental results on the Wang, GHIM-10k, and Fruit-360 dataset, XGBoost classifier can increase accuracy, precision, recall, and f1-score in all datasets. The best performance of the CBIR system is using XGBoost classifier with the best feature extractor taken from FV.FC2. Whereas the FV.FC1 feature extractor does not produce significant performance when compared to the resources needed.