Publication: Optimization of convolutional neural network architecture using particle swarm optimization for signature verification
Loading...
Date
2023-08
Authors
Tam, Yi Yang
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Fine-tuning the build convolutional neural networks (CNN)’s hyperparameters manually one-by-one to achieve the desired accuracy after determining which
hyperparameters should be fine-tuned is an arduous and inefficient task. Optimization Algorithms such as Particle Swarm Organization (PSO) are used to automate the design selection of hyperparameters and select the best performing hyperparameters from its population search. The aim of this project is to build and use Particle Swarm Optimization to optimize the build convolutional neural networks (CNN)’s architecture through fine-tuning the hyperparameters. The datasets selected to test the performance of this PSO-CNN is based on Signature Verification datasets such as CEDAR, Tam (2021)’s digital signature datasets and Mustaqim (2022)’s and (Auni’s,2022)’s digital signature datasets with Goodojoq and Baseus Pen. The selected dataset would then undergo image pre-processing to standardize the input to [64 64 1] and image augmentation of [-5 5] rotation and XY translation and split into the specified number of training, validation and testing. The PSO would then undergo 10 population searches to determine the required values of hyperparameters such as number of convolutional blocks, number of convolutional layers, filter number, filter size and batch size to build and train the CNN. The training of CNN would include
validation to perform early stopping should no further improvement in both validation accuracy and validation loss occurred for 5 validation frequency. When the 10 population searched had concluded, the best performing CNN results would be recorded and undergo K-fold validation to determine the final performance of the CNN. PSO-CNN has succeeded in automating the design selection of hyperparameters of CNN for signature verification application with each of the datasets having their own CNN’s hyperparameters selection. The PSO-CNN had achieved accuracy of 93.00% for Goodojoq Writer Dependent dataset, 92.00% for Baseus Writer Dependent dataset, 93.47% for Mixed Writer Dependent dataset, 84.22% for Goodojoq writer independent dataset, 82.94% for Baseus writer independent dataset ,85.11% for mixed writer independent dataset, 99.77% for CEDAR writer dependent dataset, 99.85% for CEDAR writer independent dataset and 87.17% for Tam writer dependent dataset.