Evolving CNN Architectures: From Custom Designs to Deep Residual Models for Diverse Image Classification and Detection Tasks

TL;DR

The study addresses the challenge of selecting CNN architectures aligned with task complexity across binary classification, fine-grained multiclass recognition, and object detection. It compares a purpose-built custom CNN—featuring depthwise separable residual blocks and bottleneck extensions—with pretrained backbones (MobileNetV2, EfficientNetB0) and detection frameworks (Faster R-CNN, YOLOv8) on five real-world datasets. Key findings show that deeper custom networks with bottleneck blocks excel in fine-grained tasks, while lightweight pretrained models often suffice for binary problems; robust object detection typically requires full backbone fine-tuning. These results offer practical guidance for designing resource-aware models and training strategies tailored to dataset characteristics and deployment constraints.

Abstract

This paper presents a comparative study of a custom convolutional neural network (CNN) architecture against widely used pretrained and transfer learning CNN models across five real-world image datasets. The datasets span binary classification, fine-grained multiclass recognition, and object detection scenarios. We analyze how architectural factors, such as network depth, residual connections, and feature extraction strategies, influence classification and localization performance. The results show that deeper CNN architectures provide substantial performance gains on fine-grained multiclass datasets, while lightweight pretrained and transfer learning models remain highly effective for simpler binary classification tasks. Additionally, we extend the proposed architecture to an object detection setting, demonstrating its adaptability in identifying unauthorized auto-rickshaws in real-world traffic scenes. Building upon a systematic analysis of custom CNN architectures alongside pretrained and transfer learning models, this study provides practical guidance for selecting suitable network designs based on task complexity and resource constraints.

Figures (11)

• Figure 1: Confusion matrices of four models on the Road Damage Dataset (Test Set).
• Figure 2: Confusion matrices of four models on the Footpath Dataset (Test Set).
• Figure 3: Confusion matrices of the two models on the MangoImageBD Dataset (Test Set).
• Figure 4: Confusion matrices of the two models on the PaddyVarietyBD Dataset (Test Set).
• Figure 5: Confusion matrix of the MiniYOLO model on the Auto-RickshawImageBD test set.