Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comparative Study of Custom CNNs, Pre-trained Models, and Transfer Learning Across Multiple Visual Datasets

Annoor Sharara Akhand

TL;DR

This study addresses the practical question of how three CNN training paradigms—training from scratch, using a frozen pre-trained network as a feature extractor, and fine-tuning a pre-trained backbone—perform across five real-world image datasets. It uses identical splits, preprocessing, and training hyperparameters to compare accuracy, macro F1-score, training time, and parameter counts. The main finding is that transfer learning consistently delivers the best predictive performance, while a carefully designed compact CNN offers a favorable efficiency-accuracy trade-off; frozen feature extractors provide solid baselines but usually underperform fine-tuning. These results supply actionable guidance for practitioners to select a CNN strategy based on domain similarity, data availability, and resource constraints.

Abstract

Convolutional Neural Networks (CNNs) are a standard approach for visual recognition due to their capacity to learn hierarchical representations from raw pixels. In practice, practitioners often choose among (i) training a compact custom CNN from scratch, (ii) using a large pre-trained CNN as a fixed feature extractor, and (iii) performing transfer learning via partial or full fine-tuning of a pre-trained backbone. This report presents a controlled comparison of these three paradigms across five real-world image classification datasets spanning road-surface defect recognition, agricultural variety identification, fruit/leaf disease recognition, pedestrian walkway encroachment recognition, and unauthorized vehicle recognition. Models are evaluated using accuracy and macro F1-score, complemented by efficiency metrics including training time per epoch and parameter counts. The results show that transfer learning consistently yields the strongest predictive performance, while the custom CNN provides an attractive efficiency--accuracy trade-off, especially when compute and memory budgets are constrained.

A Comparative Study of Custom CNNs, Pre-trained Models, and Transfer Learning Across Multiple Visual Datasets

TL;DR

This study addresses the practical question of how three CNN training paradigms—training from scratch, using a frozen pre-trained network as a feature extractor, and fine-tuning a pre-trained backbone—perform across five real-world image datasets. It uses identical splits, preprocessing, and training hyperparameters to compare accuracy, macro F1-score, training time, and parameter counts. The main finding is that transfer learning consistently delivers the best predictive performance, while a carefully designed compact CNN offers a favorable efficiency-accuracy trade-off; frozen feature extractors provide solid baselines but usually underperform fine-tuning. These results supply actionable guidance for practitioners to select a CNN strategy based on domain similarity, data availability, and resource constraints.

Abstract

Convolutional Neural Networks (CNNs) are a standard approach for visual recognition due to their capacity to learn hierarchical representations from raw pixels. In practice, practitioners often choose among (i) training a compact custom CNN from scratch, (ii) using a large pre-trained CNN as a fixed feature extractor, and (iii) performing transfer learning via partial or full fine-tuning of a pre-trained backbone. This report presents a controlled comparison of these three paradigms across five real-world image classification datasets spanning road-surface defect recognition, agricultural variety identification, fruit/leaf disease recognition, pedestrian walkway encroachment recognition, and unauthorized vehicle recognition. Models are evaluated using accuracy and macro F1-score, complemented by efficiency metrics including training time per epoch and parameter counts. The results show that transfer learning consistently yields the strongest predictive performance, while the custom CNN provides an attractive efficiency--accuracy trade-off, especially when compute and memory budgets are constrained.
Paper Structure (38 sections, 5 equations, 2 figures, 5 tables)

This paper contains 38 sections, 5 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep CNN Architectures
  4. Regularization and Architectural Design Choices
  5. Transfer Learning in Visual Recognition
  6. CNNs in Road and Infrastructure Monitoring
  7. CNNs in Agriculture and Plant Phenotyping
  8. Efficiency--Accuracy Trade-offs
  9. Positioning of This Work
  10. Datasets
  11. Dataset Overview
  12. Preprocessing
  13. Data Augmentation
  14. Class Imbalance and Sampling
  15. Methodology
  16. ...and 23 more sections

Figures (2)

  • Figure 1: Test accuracy and macro F1-score comparison across datasets.
  • Figure 2: Training time comparison across datasets and model types.