SiamMBFAN: Siamese tracker with multi-branch feature aggregation network

An empirical hybridized Siamese network using hypercube natural aggregation algorithm for handling imbalance data learning

“…A Siamese network (Bedi et al, 2020; Javaid & Jan, 2021; Zhang et al, 2022; Zhao et al, 2022) is a type of neural network architecture that is commonly used for tasks such as similarity matching, face recognition, signature verification, and others. It consists of two identical sub‐networks that share the same weights and are trained on pairs of input examples.…”

“…The main contributions of this work can be summarized as; (a) ten different imbalanced datasets ranging from 1.81 to 8.78 imbalanced ratio (IR) are collected from knowledge extraction based on evolutionary learning (KEEL) data repository (https://sci2s.ugr.es/keel/keel‐dataset/datasets/imbalanced/all//imb_IRlowerThan9.zip (n.d.); (b) first, the support vector machine (SVM) (Liu et al, 2011; Valero‐Carreras et al, 2023) is used to generate correctly predicted candidate solutions to build solution pairs; (c) then, the key advantages of Siamese network (Bedi et al, 2020; Javaid & Jan, 2021; Zhang et al, 2022; Zhao et al, 2022) such as; use of identical networks with same configuration, same parameters and same weights are utilized; (d) then, correctly predicted candidate solutions are fed to this Siamese network which is more robust to class imbalance with one‐shot learning approach to get better predictions with less number of candidate solutions; (e) as per empirical observations, a very less number of samples belonging to minority class are being correctly predicted by SVM, therefore, this Siamese network has been improvised to generate more synthetic samples by building optimized number of candidate solution pairs using HYNAA for improving the classification accuracy by maintaining a balance between minority and majority class samples; and (f) finally, the performance of proposed strategy has been measured by comparing with basic SMOTE and our previous work SMOTE‐PSOEV (Rout et al, 2022) based on ROC‐AUC learning curves and the various performance of the strategies and the computational effectiveness of this work has been evaluated.…”

“…The outputs of the two networks are then compared to produce a similarity score or distance metric. Considering the key advantages of Siamese networks few authors have used this network to deal with this data imbalance problem by hybridizing this network with some other methodologies (Bedi et al, 2020; Javaid & Jan, 2021; Zhang et al, 2022; Zhao et al, 2022). Zhao et al (2022) proposed a hybridized network of two CNNs in the Siamese network with same features and same weights named as SCNN and online reweighted example (ORE) to handle data imbalance learning.…”