A Study of the Performance and Efficiency of Transfer Learning for Image Classification.

Computer vision (CV) is a field with different tasks, including image
classification. Applications of image classification range from agriculture,
healthcare, and remote sensing to intelligent systems. Although a remarkable
performance has been achieved by deep learning (DL) models, their achievement
is usually determined by availability of large-scale annotated datasets and
extensive computational resources. These challenges are addressed by the
emergence of transfer learning (TL) as an effective solution that leverages
knowledge from pre-trained models. A comprehensive study is presented in this
paper on TL for classification of images with evaluation of its performance and
efficiency across multiple Convolutional Neural Network (CNN) architectures.
Feature extraction and fine-tuning were analyzed as two main TL strategies
using pre-trained models. Their performance was compared in terms of
computational efficiency, convergence speed, and accuracy. As demonstrated by
experimental results, TL does not only significantly improve classification
performance but also reduces training time and data requirements,
demonstrating suitability for resource-constrained and real-world applications.