IEEE ICCECE 2025
Understanding the Roles of Geometric Forms and Proportions in CNN-Based Image Classification
This research investigates how geometric variation, dataset size, CNN architecture, and proportional image resizing affect image-classification performance. The work compares natural and human-made objects to understand how differences in visual structure influence model learning and generalization.
Research question
How do object geometry, dataset composition, image proportions, and CNN design choices affect classification accuracy across natural and human-made objects?
Approach
Two custom image datasets were developed. The first contained cars, trucks, and roses, while the second contained lionesses and kittens. CNN performance was evaluated across different dataset sizes, convolutional layers, training epochs, batch sizes, and learning rates. The study also tested proportional resizing based on the approximate real-world size relationship between a lioness and a kitten.
Key results
Training accuracy achieved on the cars, trucks, and roses dataset.
Improvement in lioness and kitten testing accuracy after proportional resizing.
Natural forms such as roses were more difficult to classify than visually structured human-made objects.
Technical focus
The experiments examined the relationship between object geometry and CNN behavior, including how scale, proportional preprocessing, network depth, and dataset composition influence classification accuracy.
DOI: 10.1109/ICCECE61355.2025.10941645
.jpeg)
.jpeg)
.jpeg)