Deep Learning · Computer Vision · Team Project

Lizard Species Classification

A seven-class image classification project using EfficientNetV2S transfer learning, with data augmentation, fine-tuning and per-class evaluation. Achieved 77.0% validation accuracy and a macro F1 score of 0.7631 on 200 validation images.

Technology Stack

Python TensorFlow Keras EfficientNetV2S scikit-learn Matplotlib
Validation accuracy77.0%
Validation macro F10.7631
Validation images200
Classes7

Overview

Developed as a three-person academic project, this classifier distinguishes seven lizard species. The work covered data preparation, transfer learning, fine-tuning and evaluation, with particular attention to differences in performance between visually similar classes.

Approach

We used an ImageNet-pretrained EfficientNetV2S backbone and 224 × 224 input images. Training began with a frozen backbone, followed by fine-tuning of upper layers. Data augmentation included flipping, rotation, zoom, translation and contrast adjustments. The final workflow also generated submission predictions using original and horizontally flipped images.

Evaluation

The dataset contained 1,334 labelled images, split into 1,134 training images and 200 validation images. The final model achieved 77.0% validation accuracy and a macro F1 score of 0.7631. A confusion matrix and per-class precision, recall and F1 scores helped identify which species remained difficult to distinguish.

All figures are measured on our held-out validation images. They are not a Kaggle leaderboard score or a production accuracy. The validation set came from a random shuffled split of the labelled training data.

Confusion matrix on the 200 validation images
Confusion matrix for the seven lizard classes. Most images fall on the diagonal; Green anole is most often confused with Brown anole and Cuban knight anole. Open Full Size
Loss and accuracy over all epochs. The dashed line separates Phase 1 (frozen backbone) from Phase 2 (fine-tuning).
Training and validation loss and accuracy curves. A dashed vertical line marks the switch from Phase 1 to Phase 2. Training accuracy keeps rising in Phase 2 while validation accuracy stays between roughly 70% and 77%. Open Full Size

Per-class results

Precision, recall and F1 score per class on the validation set
SpeciesPrecisionRecallF1Images
Black spiny-tailed iguana0.8060.8330.82030
Brown anole0.7080.6800.69425
Cuban knight anole0.6670.8460.74626
Desert iguana0.9690.9690.96932
Green anole0.8000.5520.65329
Green iguana0.6360.7500.68928
Lesser Antillean iguana0.8150.7330.77230

What the results show

Performance varied by class: Desert iguana reached an F1 score of 0.969, while Green anole reached 0.653. This highlighted the importance of examining class-level errors alongside overall accuracy.

Limitations

Role & Scope

A team project with José Gutiérrez Montero, Mateo Donado González and me. I contributed to model development and evaluation as part of the team.

Back to Projects