Background: The main purpose of this thesis isto study the impact of data augmentation on theMobileNet architecture, specifically inhandwritten signature identification systems. Themodel selected for this investigation is theMobileNetV3-Small.This study examines howaffine data augmentation techniques, such asrotation, scaling, and translation, enhance theperformance and generalisation capabilities ofthe model.

Objectives: This thesis aims to evaluate theeffectiveness of data augmentation in enhancingthe performance of the MobileNet architecturefor identifying hand-written signatures. Thestudy involves implementing the MobileNetV3-Small CNN architecture, integrating a dataaugmentation pipeline, and evaluating its impacton model performance. This evaluation includestraining the model on both original andaugmented datasets to study accuracy andgeneralisation capabilities in signatureidentification tasks.

Methods: This research uses a formalexperimental research methodology to study theimpact of data augmentation on MobileNetV3-Small for handwritten signature identification.The dataset used for this study is gathered fromKaggle and the dataset is preprocessed to makeit suitable for training the model. Themethodology consists of the implementation of MobileNetV3-Small architecture and theintegration of a data augmentation pipeline thatapplies techniques such as rotation, scaling andtranslation to the preprocessed dataset. Theperformance of the model is evaluated bytraining on both original and augmenteddatasets, the evaluation is performed by usingmetrics such as model accuracy, precision, recall,F1 score, and loss.

Results: The findings indicate that dataaugmentation enhances the model’s accuracyand generalisation ability. The MobileNetV3-Smallmodel trained with the original dataset hasachieved an accuracy of 75.26% and 85.94% ofaccuracy when trained with augmented data,which demonstrates better performance insignature identification, showing improvedaccuracy and a lower loss percentage than themodel trained without data augmentation.

Conclusions: In this thesis, the increasedvalidation accuracy, F1 score, and decreased losspercentage show that the traditional affinetransformations positively impacted theperformance of the MobileNetV3-Small modeland improved its generalisation. This studyconfirms that even simple, label-preservingtransformations can improve the capabilities oflightweight models like MobileNetV3-Small incomplex recognition tasks.