A hybrid deep learning-based model for enhanced feature representation in image retrieval systems

Abstract

The exponential growth of image data volume has made the necessity of accurate and efficient retrieval systems more and more evident; in this regard, extracting comprehensive and meaningful features and their optimal selection has become a vital issue in image processing and machine learning research. The present study, with the aim of improving the representation of features in image retrieval systems, presents a novel hybrid model based on deep learning. In the proposed method, first, a rich set of image features is created by combining contextual features (Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) on image fragments), basic features (statistical features and Gray Level Cooccurrence Matrix (GLCM)), and deep features (extracted through a convolutional neural network). Then, in order to reduce the dimensions and select the most expressive features, a new feature selection algorithm based on reinforcement learning using learning automata is applied. Finally, the Fuzzy C-Means (FCM) clustering model is used to build a retrieval model and recall related images based on the selected features. The originality of this research lies in providing an integrated model that, by making intelligent integration of various methods of feature extraction and a reinforcement learning-based feature selection algorithm, attempts to break the bottleneck of current models in terms of diversity of image feature description and precision of retrieval. Evaluation of the proposed method on two datasets, Corel-1000 and ALOI, showed that this method provides significant performance compared to other methods, achieving average Precisions of 0.9711 and 0.8906 in these datasets, respectively. This indicates the efficiency of the proposed approach in extracting and selecting relevant and effective features for accurate image retrieval.