Vision-only Deep Learning Baseline for Meme-based Humor Classification

📄 Published Research - Namal Business Conference 2026 (NBC 2026) 🏷️ Track: Technology and Innovation | Abstract ID: T4.02 👤 Author: Hamza Maqsood - University of Management and Technology, Lahore

---

📌 Overview

This paper presents a deep learning framework for visual humor classification using a fine-tuned ResNet18 convolutional neural network.

Prior research on humor classification has largely focused on text-based or multimodal approaches. This work addresses the gap by proposing a vision-only baseline that operates exclusively on meme images.

---

🔬 Methodology

• Model: ResNet18 - fine-tuned with ImageNet pretrained weights
• Task: Multi-class meme humor classification
• Data: Labelled meme dataset with stratified train/val/test splits
• Techniques: Transfer learning · Data augmentation · Early stopping
• Evaluation: Confusion matrix · Precision · Recall · F1-score

---

📊 Key Findings

• ResNet18 effectively learns discriminative visual features for humor recognition
• Vision-only approach achieves stable generalization on unseen data
• Demonstrates viability of image-only humor classification without text modality

---

🛠️ Tech Stack

Python · PyTorch · ResNet18 · Transfer Learning · OpenCV

---

📎 Reference

Abstract Book — NBC 2026, Namal University Mianwali View Publication