Abstract:The increasing demand for effective retrieval from image datasets has been driven by the rapid growth of digital images. Image retrieval is a method for creating a structured database based on searching for similar images to a user's query image. Effective image representation and extraction of distinctive features are among the challenges faced by image retrieval systems. The proposed approach utilizes a triplet loss function based on binary cross-entropy to train a Siamese network, allowing for deep metric learning and creating a discriminative feature space with maximum discrimination between classes and minimum intra-class distance. In this approach, a pre-trained neural network is serialized with Siamese network. Initially, image features are extracted using a pre-trained convolutional neural network. Then, a Siamese network is trained to create a discriminative feature space using deep metric learning. Learning in this method is controlled by a triplet loss function based on binary cross-entropy with anchor, positive, and negative samples. The transformation from the initial feature space to the discriminative feature space is achieved through deep metric learning in the Siamese network. The proposed approach proves to effective in discriminating features of diverse classes in the dataset. Visualization of the feature space is also demonstrated using the t-SNE statistical technique. Furthermore, the explainability of the proposed approach is presented by examining the Shapley value. The proposed approach was examined on the Corel10K, Caltech101, and Caltech256 datasets. The best reported results with the precision metric were 0.988 for the Corel10K dataset and 0.951 and 0.902 for the Caltech101 and Caltech256 datasets, respectively. Experimental results show that the proposed approach effectively improves the retrieval results of similar samples to the query image creating discrimination in the feature vectors, even with dimensionality reduction.

A case weighted similarity deep measurement method based on a self-attention Siamese neural network

Self-attention driven adversarial similarity learning network

Hybridization between Neural Computing and Nature-Inspired Algorithms for a Sentence Similarity Model Based on the Attention Mechanism

Chinese Sentences Similarity via Cross-Attention Based Siamese Network

Modality-Specific Cross-Modal Similarity Measurement With Recurrent Attention Network

Sentence-Embedding and Similarity via Hybrid Bidirectional-LSTM and CNN Utilizing Weighted-Pooling Attention

A Siamese neural network for learning the similarity metrics of linear features

Enhancing image retrieval through entropy-based deep metric learning

Hybrid Attention Based Neural Architecture for Text Semantics Similarity Measurement

Hierarchical Deep Multi-task Learning with Attention Mechanism for Similarity Learning

Compressed Self-Attention for Deep Metric Learning with Low-Rank Approximation

Learning a Deep Similarity Network for Hyperspectral Image Classification

Metric networks for enhanced perception of non-local semantic information

Traditional Chinese Medicine Text Similarity Calculation Model Based on the Bidirectional Temporal Siamese Network

Similarity-based Text Recognition by Deeply Supervised Siamese Network

Unsupervised Deep Metric Learning with Transformed Attention Consistency and Contrastive Clustering Loss

A Comparative Study of Neural-Network Feature Weighting

Re-LSTM: A long short-term memory network text similarity algorithm based on weighted word embedding

Signal-To-Noise Ratio: A Robust Distance Metric For Deep Metric Learning

DeepSIM: a novel deep learning method for graph similarity computation

Modeling Time Series Similarity with Siamese Recurrent Networks