Performance Analysis of Image Classification on Bangladeshi Datasets

TL;DR

The paper addresses whether a handcrafted CNN or pretrained architectures better suit image classification on Bangladeshi datasets. It implements a custom CNN and uses ResNet-50 and ConvNeXt-Tiny as fixed feature extractors with selective fine-tuning, under identical settings across binary and multiclass tasks. Key findings show pretrained models outperform the custom network in accuracy and convergence speed, with ConvNeXt-Tiny often achieving the best overall performance, albeit with higher compute demands; the custom CNN remains competitive for simpler tasks and resource-limited environments. The work provides practical guidance on architecture choice based on dataset complexity and available computational resources, highlighting the value of transfer learning for domain-specific, small- to medium-sized datasets.

Abstract

Convolutional Neural Networks (CNNs) have demonstrated remarkable success in image classification tasks; however, the choice between designing a custom CNN from scratch and employing established pre-trained architectures remains an important practical consideration. In this work, we present a comparative analysis of a custom-designed CNN and several widely used deep learning architectures, including VGG-16, ResNet-50, and MobileNet, for an image classification task. The custom CNN is developed and trained from scratch, while the popular architectures are employed using transfer learning under identical experimental settings. All models are evaluated using standard performance metrics such as accuracy, precision, recall, and F1-score. Experimental results show that pre-trained CNN architectures consistently outperform the custom CNN in terms of classification accuracy and convergence speed, particularly when training data is limited. However, the custom CNN demonstrates competitive performance with significantly fewer parameters and reduced computational complexity. This study highlights the trade-offs between model complexity, performance, and computational efficiency, and provides practical insights into selecting appropriate CNN architectures for image classification problems.

Figures (15)

• Figure 1: Training loss and Acuuracy curves of Custom CNN model across epochs on Auto-RickshawImageBD Dataset.
• Figure 2: Training loss and Acuuracy curves of RestNet-50 across epochs on Auto-RickshawImageBD Dataset.
• Figure 3: Training loss and Acuuracy curves of ConvNext-Tiny across epochs on Auto-RickshawImageBD Dataset.
• Figure 4: Training loss and Acuuracy curves of Custom CNN model across epochs on Auto-FootpathVisionBD Dataset.
• Figure 5: Training loss and Acuuracy curves of RestNet-50 across epochs on Auto-FootpathVisionBD Dataset.