Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Experimental Comparison of Light-Weight and Deep CNN Models Across Diverse Datasets

Md. Hefzul Hossain Papon, Shadman Rabby

TL;DR

The paper investigates whether a carefully designed lightweight CNN can match or approach the performance of large pretrained networks across diverse domain-specific vision tasks in resource-limited environments. It introduces a compact CustomCNN with three convolutional blocks, global average pooling, and a two-layer classifier, trained under a reproducible workflow with stratified data splits, class weighting, and early stopping. Through experiments on five Bangladeshi datasets (RoadDamageBD, PaddyVarietyBD, MangoImageBD, FootpathVision, Auto-RickshawImageBD), the study compares the CustomCNN against EfficientNetB0 and ResNet18, with and without transfer learning, revealing favorable efficiency-accuracy trade-offs for lightweight models in several domains while transfer learning often yields higher performance at greater computational cost. The findings provide practical guidance for deploying vision systems in low-resource settings, suggesting when to favor a compact architecture versus pretrained, transfer-learned models depending on deployment constraints and task difficulty.

Abstract

Our results reveal that a well-regularized shallow architecture can serve as a highly competitive baseline across heterogeneous domains - from smart-city surveillance to agricultural variety classification - without requiring large GPUs or specialized pre-trained models. This work establishes a unified, reproducible benchmark for multiple Bangladeshi vision datasets and highlights the practical value of lightweight CNNs for real-world deployment in low-resource settings.

Experimental Comparison of Light-Weight and Deep CNN Models Across Diverse Datasets

TL;DR

The paper investigates whether a carefully designed lightweight CNN can match or approach the performance of large pretrained networks across diverse domain-specific vision tasks in resource-limited environments. It introduces a compact CustomCNN with three convolutional blocks, global average pooling, and a two-layer classifier, trained under a reproducible workflow with stratified data splits, class weighting, and early stopping. Through experiments on five Bangladeshi datasets (RoadDamageBD, PaddyVarietyBD, MangoImageBD, FootpathVision, Auto-RickshawImageBD), the study compares the CustomCNN against EfficientNetB0 and ResNet18, with and without transfer learning, revealing favorable efficiency-accuracy trade-offs for lightweight models in several domains while transfer learning often yields higher performance at greater computational cost. The findings provide practical guidance for deploying vision systems in low-resource settings, suggesting when to favor a compact architecture versus pretrained, transfer-learned models depending on deployment constraints and task difficulty.

Abstract

Our results reveal that a well-regularized shallow architecture can serve as a highly competitive baseline across heterogeneous domains - from smart-city surveillance to agricultural variety classification - without requiring large GPUs or specialized pre-trained models. This work establishes a unified, reproducible benchmark for multiple Bangladeshi vision datasets and highlights the practical value of lightweight CNNs for real-world deployment in low-resource settings.
Paper Structure (19 sections, 10 equations, 11 figures, 6 tables, 1 algorithm)

This paper contains 19 sections, 10 equations, 11 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. Dataset Construction and Preprocessing
  4. Custom CNN Architecture
  5. Convolutional Feature Extractor
  6. Block Definitions.
  7. Global Average Pooling and Classifier
  8. Weight Initialization
  9. Training Strategy
  10. Early Stopping
  11. Overall Workflow
  12. Experimental Results
  13. RoadDamageBD
  14. PaddyVarietyBD.
  15. MangoImageBD.
  16. ...and 4 more sections

Figures (11)

  • Figure 1: Architecture diagram of the proposed Custom CNN model.
  • Figure 2: Performance comparison of CustomCNN, scratch-trained, and transfer learning models on the RoadDamageBD dataset.
  • Figure 3: Comaprison of training Curve for testing models on RoadDamageBD dataset
  • Figure 4: Performance comparison of CustomCNN, scratch-trained, and transfer learning models on the FootpathVision dataset.
  • Figure 5: Comaprison of training Curve for testing models on Footpath Vision dataset
  • ...and 6 more figures