Undergraduate thesis · Deep learning
Corn Leaf Disease Detection Using Deep Learning
- Institution
- Daffodil International University
- Period
- 2026
- Models
- DenseNet121 · EfficientNet-B0
A controlled comparison of two convolutional backbones, DenseNet121 and EfficientNet-B0, for classifying corn leaf disease from images. The study was built around a question the accuracy headline usually hides: does an architecture comparison still hold once the split, the augmentation and the training budget are held fixed for both models?
What the comparison isolates
01
DenseNet121
A densely connected backbone, where each layer receives the feature maps of every earlier layer. Included for its strong performance on small, fine-grained image datasets.
02
EfficientNet-B0
A compound-scaled backbone that balances depth, width and resolution. Included as the more parameter-efficient alternative to test whether fewer parameters cost anything in accuracy.
03
Transfer learning from ImageNet
Both models start from ImageNet weights rather than random initialisation, which is the realistic setting when labelled agricultural imagery is scarce.
Results
Documentation in progress
The dataset card, per-class metrics and the DenseNet121 / EfficientNet-B0 comparison table are published together with the thesis document, so the numbers here stay traceable to the report rather than being quoted from memory. The LaTeX source and figures live in the public repository.
Limitations
Corn leaf disease is a narrow slice of agricultural vision: results here should not be read as evidence about field conditions, occlusion or other crops.
A two-model comparison is a starting point, not a survey — the honest conclusion is about these two backbones under one fixed protocol.
Reported accuracy is only meaningful alongside the split it was measured on, which is why the protocol is documented before the number.