Image Retrieval Based on Deep Learning

Abstract

The digital image has an important role in many fields such as biomedical, robotics, weather forecasting and object recognition.Due to the widespread use of social media sites, cloud services, and cellphones, large image databases are easily accessible.Searching by text is a common and simple method, however if the algorithm is running properly, searching by visual content will be much more sensitive.The goal of this research is to create a content-based image retrieval method that is more accurate in image retrieval.In this approach, intelligent systems can assist and work successfully.This study analyses three deep learning-based proposal methodologies: CNN, convolutional layers fused with LSTM, and Convolutional layers fused with GRU.The models were tested on four distinct databases of varying sizes, including Corel1K, Cifar-10, Cifar-100, and Mnist 70K.In comparison to state-of-the-art models, the three presented algorithms have significantly reduced computation time and provided very high picture retrieval levels of accuracy.For Corel1K, Cifar-10, Cifar-100, and Mnist 70k, CNN's proposed model scored 93.3, 94%, 85.5 %, and 99.9 %, respectively.the second proposed model scored 94.5%, 95%, 86.5%, and 99.9 for Corel1K, Cifar-10, Cifar-100, and Mnist 70k, respectively.Finally, for Corel1K, Cifar-10, Cifar-100, and Mnist 70k, the third proposed model reached 95.5%, 96 % , 87.5 %, and 99.9%, respectively.