Classification of electric gutar types using MobileNetV2 and transfer learning

Goal

The aim of the study is to create and evaluate a transfer learning model for multiclass classification of electric guitar types.

Methods

The work uses an open dataset consisting of 7 classes, 605 images. MobileNetV2 was used as the architecture, followed by the addition of global averaging, dropout, and fully connected output layers. The training is performed in two stages: first, only the classification heads, then the fine-tuning of the last 40 layers. Data augmentation has been applied to increase generalizing ability.

Results

The model achieved an accuracy of 73.77% on an independent test sample. The best metrics were shown by the Telecaster and Les Paul classes. The main errors occurred between visually similar pairs: ES and SG.

Conclusions

The results confirm the effectiveness of MobileNetV2 in small data tasks. The solution has practical implications for mobile applications, cataloging, and learning. In the future, it is planned to expand the dataset, use Grad-CAM and optimize the model for peripheral computing devices.