Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Comparative Analysis of Custom CNN Architectures versus Pre-trained Models and Transfer Learning: A Study on Five Bangladesh Datasets

Ibrahim Tanvir, Alif Ruslan, Sartaj Solaiman

TL;DR

This paper systematically compares custom CNNs, pre-trained models used as feature extractors, and transfer learning with fine-tuning across five Bangladesh image classification datasets. It finds that fine-tuning pre-trained networks consistently yields the highest accuracy, with notable gains on challenging tasks such as Paddy Variety BD, and even perfect accuracy on Road Damage BD. Feature extraction offers a practical middle ground with substantial speed and parameter reductions, while custom CNNs lag on complex tasks despite smaller size. The results provide concrete, domain-relevant guidance for practitioners balancing accuracy, data availability, and resource constraints in real-world Bangladeshi applications.

Abstract

This study presents a comprehensive comparative analysis of custom-built Convolutional Neural Networks (CNNs) against popular pre-trained architectures (ResNet-18 and VGG-16) using both feature extraction and transfer learning approaches. We evaluated these models across five diverse image classification datasets from Bangladesh: Footpath Vision, Auto Rickshaw Detection, Mango Image Classification, Paddy Variety Recognition, and Road Damage Detection. Our experimental results demonstrate that transfer learning with fine-tuning consistently outperforms both custom CNNs built from scratch and feature extraction methods, achieving accuracy improvements ranging from 3% to 76% across different datasets. Notably, ResNet-18 with fine-tuning achieved perfect 100% accuracy on the Road Damage BD dataset. While custom CNNs offer advantages in model size (3.4M parameters vs. 11-134M for pre-trained models) and training efficiency on simpler tasks, pre-trained models with transfer learning provide superior performance, particularly on complex classification tasks with limited training data. This research provides practical insights for practitioners in selecting appropriate deep learning approaches based on dataset characteristics, computational resources, and performance requirements.

Comparative Analysis of Custom CNN Architectures versus Pre-trained Models and Transfer Learning: A Study on Five Bangladesh Datasets

TL;DR

This paper systematically compares custom CNNs, pre-trained models used as feature extractors, and transfer learning with fine-tuning across five Bangladesh image classification datasets. It finds that fine-tuning pre-trained networks consistently yields the highest accuracy, with notable gains on challenging tasks such as Paddy Variety BD, and even perfect accuracy on Road Damage BD. Feature extraction offers a practical middle ground with substantial speed and parameter reductions, while custom CNNs lag on complex tasks despite smaller size. The results provide concrete, domain-relevant guidance for practitioners balancing accuracy, data availability, and resource constraints in real-world Bangladeshi applications.

Abstract

This study presents a comprehensive comparative analysis of custom-built Convolutional Neural Networks (CNNs) against popular pre-trained architectures (ResNet-18 and VGG-16) using both feature extraction and transfer learning approaches. We evaluated these models across five diverse image classification datasets from Bangladesh: Footpath Vision, Auto Rickshaw Detection, Mango Image Classification, Paddy Variety Recognition, and Road Damage Detection. Our experimental results demonstrate that transfer learning with fine-tuning consistently outperforms both custom CNNs built from scratch and feature extraction methods, achieving accuracy improvements ranging from 3% to 76% across different datasets. Notably, ResNet-18 with fine-tuning achieved perfect 100% accuracy on the Road Damage BD dataset. While custom CNNs offer advantages in model size (3.4M parameters vs. 11-134M for pre-trained models) and training efficiency on simpler tasks, pre-trained models with transfer learning provide superior performance, particularly on complex classification tasks with limited training data. This research provides practical insights for practitioners in selecting appropriate deep learning approaches based on dataset characteristics, computational resources, and performance requirements.
Paper Structure (24 sections, 1 figure, 2 tables)

This paper contains 24 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Research Contributions
  3. Related Work
  4. Methodology
  5. Datasets
  6. Model Architectures
  7. Custom CNN Architecture
  8. Pre-trained Models
  9. Training Approaches
  10. Evaluation Metrics
  11. Experimental Results
  12. Overall Performance Comparison
  13. Dataset-Specific Analysis
  14. Footpath Vision Dataset
  15. Auto Rickshaw Dataset
  16. ...and 9 more sections

Figures (1)

  • Figure 1: Comprehensive performance visualizations showing (a) test accuracy across datasets, (b) model size comparison, (c) training time, (d) custom vs pretrained vs transfer learning, (e) F1-scores, and (f) performance improvement over custom CNN.