Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluation of Convolutional Neural Network For Image Classification with Agricultural and Urban Datasets

Shamik Shafkat Avro, Nazira Jesmin Lina, Shahanaz Sharmin

TL;DR

The study evaluates a CustomCNN across five diverse datasets spanning urban and agricultural imagery to understand how architectural choices impact multi-domain image classification. The CustomCNN combines residual learning, SE attention, and progressive channel expansion, and is benchmarked against ResNet50 and VGG16 with and without ImageNet pretraining. Results show pretrained models generally achieve higher accuracy and stability, but the lightweight CustomCNN remains competitive in many tasks and offers substantial efficiency benefits; in some domain-specific datasets (e.g., PaddyVarietyBD) transfer learning can underperform due to domain mismatch. The work provides guidance on when to favor compact architectures versus deep pretrained models for real-world smart city and agricultural imaging deployments.

Abstract

This paper presents the development and evaluation of a custom Convolutional Neural Network (CustomCNN) created to study how architectural design choices affect multi-domain image classification tasks. The network uses residual connections, Squeeze-and-Excitation attention mechanisms, progressive channel scaling, and Kaiming initialization to improve its ability to represent data and speed up training. The model is trained and tested on five publicly available datasets: unauthorized vehicle detection, footpath encroachment detection, polygon-annotated road damage and manhole detection, MangoImageBD and PaddyVarietyBD. A comparison with popular CNN architectures shows that the CustomCNN delivers competitive performance while remaining efficient in computation. The results underscore the importance of thoughtful architectural design for real-world Smart City and agricultural imaging applications.

Evaluation of Convolutional Neural Network For Image Classification with Agricultural and Urban Datasets

TL;DR

The study evaluates a CustomCNN across five diverse datasets spanning urban and agricultural imagery to understand how architectural choices impact multi-domain image classification. The CustomCNN combines residual learning, SE attention, and progressive channel expansion, and is benchmarked against ResNet50 and VGG16 with and without ImageNet pretraining. Results show pretrained models generally achieve higher accuracy and stability, but the lightweight CustomCNN remains competitive in many tasks and offers substantial efficiency benefits; in some domain-specific datasets (e.g., PaddyVarietyBD) transfer learning can underperform due to domain mismatch. The work provides guidance on when to favor compact architectures versus deep pretrained models for real-world smart city and agricultural imaging deployments.

Abstract

This paper presents the development and evaluation of a custom Convolutional Neural Network (CustomCNN) created to study how architectural design choices affect multi-domain image classification tasks. The network uses residual connections, Squeeze-and-Excitation attention mechanisms, progressive channel scaling, and Kaiming initialization to improve its ability to represent data and speed up training. The model is trained and tested on five publicly available datasets: unauthorized vehicle detection, footpath encroachment detection, polygon-annotated road damage and manhole detection, MangoImageBD and PaddyVarietyBD. A comparison with popular CNN architectures shows that the CustomCNN delivers competitive performance while remaining efficient in computation. The results underscore the importance of thoughtful architectural design for real-world Smart City and agricultural imaging applications.
Paper Structure (55 sections, 1 equation, 23 figures, 5 tables)

This paper contains 55 sections, 1 equation, 23 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Overview of CNN
  4. ResNet
  5. VGG16
  6. Transfer Learning in CNN
  7. Methodology
  8. Custom CNN Architecture
  9. Overall Architecture Design
  10. Design Rationale and Architectural Choices
  11. Residual Connections for Stable Deep Learning
  12. Squeeze-and-Excitation Attention Mechanism
  13. Choice of Convolutional Kernel Size
  14. Progressive Channel Expansion
  15. Padding Strategy
  16. ...and 40 more sections

Figures (23)

  • Figure 1: Training vs Validation Accuracy for Custom CNN showing (a) Learning Rate: 0.0001, Dropout: 0.0 and (b) Learning Rate: 0.001, Dropout: 0.5.
  • Figure 2: Training vs Validation Accuracy showing (a) Custom CNN (Learning Rate: 0.0001, Dropout: 0.5) and (b) ResNet50 from scratch.
  • Figure 3: Training vs Validation Accuracy showing (a) ResNet50 with Transfer Learning and (b) VGG16 from scratch.
  • Figure 4: Training vs Validation Accuracy for VGG16 with Transfer Learning.
  • Figure 5: Precision--recall curve (left) and ROC curve (right) for Dataset 1.
  • ...and 18 more figures