Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reliable Deep Learning for Small-Scale Classifications: Experiments on Real-World Image Datasets from Bangladesh

Muhammad Ibrahim, Alfe Suny, MD Sakib Ul Islam, Md. Imran Hossain

TL;DR

This study evaluates a compact CNN on five real-world Bangladeshi datasets, comparing it with scratch-trained heavy models and transfer learning under identical conditions. The lightweight architecture achieves competitive accuracy on low-cardinality tasks (Datasets 1–3) with far fewer parameters and lower training time, while high-cardinality tasks (Dataset 5) favor pre-trained models despite a substantial resource burden. Data augmentation improves the lightweight model’s performance on fine-grained classifications, narrowing but not eliminating the gap to transfer learning. Overall, the work provides practical guidance for deploying efficient deep learning solutions in edge and resource-constrained environments, demonstrating that larger models are not always necessary for robust performance.

Abstract

Convolutional neural networks (CNNs) have achieved state-of-the-art performance in image recognition tasks but often involve complex architectures that may overfit on small datasets. In this study, we evaluate a compact CNN across five publicly available, real-world image datasets from Bangladesh, including urban encroachment, vehicle detection, road damage, and agricultural crops. The network demonstrates high classification accuracy, efficient convergence, and low computational overhead. Quantitative metrics and saliency analyses indicate that the model effectively captures discriminative features and generalizes robustly across diverse scenarios, highlighting the suitability of streamlined CNN architectures for small-class image classification tasks.

Reliable Deep Learning for Small-Scale Classifications: Experiments on Real-World Image Datasets from Bangladesh

TL;DR

This study evaluates a compact CNN on five real-world Bangladeshi datasets, comparing it with scratch-trained heavy models and transfer learning under identical conditions. The lightweight architecture achieves competitive accuracy on low-cardinality tasks (Datasets 1–3) with far fewer parameters and lower training time, while high-cardinality tasks (Dataset 5) favor pre-trained models despite a substantial resource burden. Data augmentation improves the lightweight model’s performance on fine-grained classifications, narrowing but not eliminating the gap to transfer learning. Overall, the work provides practical guidance for deploying efficient deep learning solutions in edge and resource-constrained environments, demonstrating that larger models are not always necessary for robust performance.

Abstract

Convolutional neural networks (CNNs) have achieved state-of-the-art performance in image recognition tasks but often involve complex architectures that may overfit on small datasets. In this study, we evaluate a compact CNN across five publicly available, real-world image datasets from Bangladesh, including urban encroachment, vehicle detection, road damage, and agricultural crops. The network demonstrates high classification accuracy, efficient convergence, and low computational overhead. Quantitative metrics and saliency analyses indicate that the model effectively captures discriminative features and generalizes robustly across diverse scenarios, highlighting the suitability of streamlined CNN architectures for small-class image classification tasks.
Paper Structure (18 sections, 12 figures, 7 tables)

This paper contains 18 sections, 12 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Network Design Rationale
  5. Datasets
  6. Results and Discussion
  7. Dataset 1: Auto-Rickshaw Image Classification
  8. Dataset 2: Footpath Encroachment Classification
  9. Training Performance
  10. Dataset 3: Road Damage Classification
  11. Performance of Custom CNN
  12. Comparative Analysis with Transfer Learning
  13. The Efficiency Trade-off
  14. Dataset 4 & 5: MangoImage & Paddy Variety Classification
  15. Performance of Custom CNN
  16. ...and 3 more sections

Figures (12)

  • Figure 1: Overview of the proposed convolutional network architecture. The network consists of convolutional blocks followed by fully connected layers, designed for efficient feature extraction and classification.
  • Figure 2: Training and validation loss curves for Dataset 2.
  • Figure 3: Training and validation accuracy curves for Dataset 2.
  • Figure 4: Training and validation curves for Dataset 2.
  • Figure 5: Confusion matrix for Dataset 2.
  • ...and 7 more figures