Performance Analysis of Multiple Deep Learning Models for Image Retrieval Problems

In modern times, the exponential growth of digital images threatens conventional image retrieval frameworks. Shallow machine learning algorithms degrade image retrieval performance owing to the semantic gap between low-level and high-level features. Hence, deep learning (DL) is a possible way to escape from this. Previous studies have yet to determine the superior DL model or combination for similarity-based image retrieval. In this paper, we apply deep features instead of traditional features for image retrieval. We first execute eight pretrained DL models individually and then all pairs combining two, three, and four models. Experimental results show that model combinations outperform single models, with increased accuracy as the number of models increases. We combine a maximum of four models. Our empirical findings from various model combinations indicate that the most effective combination includes ResNet101V2, InceptionV3, InceptionResNetV2, and DenseNet201, yielding a mean average precision (mAP) of 97.55% and mean average recall (mAR) of 19.51%.