ZooTag Model Card

Museum specimen classifier: 7 broad animal groups, transfer learning, TensorFlow Lite.

Model details

• Developed by: Group 8, Uppsala University (ZooTag Deep Learning Challenge).
• Date: 16 October 2026. Version 1.0.
• Model type: image classifier, transfer learning from EfficientNetV2B0 (frozen ImageNet backbone with a new dense head; the top 20 backbone layers were then fine-tuned at a low learning rate).
• Input: one RGB image of shape (224, 224, 3), float pixel values in [0, 255]. Normalisation is built into the model.
• Output: softmax probabilities over 7 labels, in the order AMPHIBIAN, BIRD, FISH, INSECT, MAMMAL, MARINE_INVERTEBRATE, REPTILE.
• Formats: Keras (.keras, .h5) and TensorFlow Lite (.tflite, dynamic-range quantised).
• Framework: TensorFlow 2.21 / Keras 3.

Intended use

Primary use: the Evolutionsmuseet visitor app. A visitor photographs a specimen in an exhibit and the app shows its broad taxonomic group, to encourage engagement.

• Intended users: museum visitors, through the mobile app.
• Out of scope: species-level identification, live animals in the wild, scientific or conservation decisions, and images that do not contain an animal (the model always returns one of the 7 labels and has no 'unknown' class).

Training data

Photos taken by the team at Evolutionsmuseet Zoology Section, Uppsala, on 14 October 2026, with a smartphone. The photos are deliberately varied: different angles and distances, visible backgrounds, reflections on the display glass, and partial views. They were sorted by hand into TRAINING and TESTING folders for each of the 7 classes, in JPG, JPEG and PNG format. Before training, 191 photos that were stored with an EXIF rotation flag were turned upright, and every image was converted to RGB JPEG.

Training set: 280 images. 15% of these were held out as a validation set for early stopping. Test set: 56 images, 8 per class, which were never seen during training. 2 training images that were byte-identical copies of test images were removed so the test set stays unseen (INSECT/IMG_8370.jpeg, MARINE_INVERTEBRATE/IMG_4346.jpg).

Training procedure

• Base model: keras.applications.EfficientNetV2B0 with ImageNet weights, include_top=False.
• Head: GlobalAveragePooling2D -> Dropout(0.3) -> Dense(7, softmax).
• Augmentation (training only): horizontal flip, rotation ±10%, zoom ±20%, translation ±10%, contrast and brightness ±20%. These simulate visitor photos.
• Optimiser: Adam, learning rate 1e-3, then 1e-5 for fine-tuning. Loss: categorical cross-entropy with label smoothing 0.1. Batch size 16.
• Early stopping on validation loss (patience 5), 30 epochs in total. Trained on CPU in a few minutes.

Evaluation

Results on the held-out test set (56 images): accuracy 87.5%, macro precision 89.2%, macro recall 87.5%, macro F1 87.5%.

TFLite conversion and verification: the saved zootag.h5 model was reloaded and converted with the TensorFlow Lite converter, using dynamic-range quantisation (6.6 MB). The .tflite model was then run on every test image. Its top prediction agreed with the .h5 model on 98% of images. TFLite test accuracy was 85.7%, and the largest difference in any class probability was 0.060, caused by quantisation. The sample below has two test images per class. The full per-image results are in tflite_verification.csv.

Limitations and risks

• Very small dataset (about 40 images per class) from a single museum and a single phone. With about 8 test images per class, one mistake changes per-class recall by about 12.5 percentage points, so all the metrics are uncertain.
• The specimens are stuffed or preserved and shown behind glass. Performance on live animals, drawings or other museums is likely to be lower.
• Reptiles are the hardest class: 3 of 8 test reptiles were predicted as amphibians, which have similar body shapes and are shown in similar dioramas. The other errors are single mix-ups (fish predicted as amphibian or marine invertebrate, an amphibian as a bird, an insect as a reptile). See the confusion matrix.
• The model can learn exhibit backgrounds, for example a fish tank, instead of the animal itself.
• There is no rejection option. Photos without an animal still get one of the 7 labels, often with high confidence.

Recommendations

• Show the top 2-3 labels with their confidence, and treat low confidence (< 0.5) as 'not sure'.
• Collect more photos, especially of the weakest classes, before relying on per-class results.
• Add an 'other / no animal' class for production use.

How to use

import numpy as np, tensorflow as tf
from PIL import Image
labels = open('labels.txt').read().split()
it = tf.lite.Interpreter(model_path='zootag.tflite'); it.allocate_tensors()
img = np.asarray(Image.open('photo.jpg').convert('RGB').resize((224, 224)), np.float32)
it.set_tensor(it.get_input_details()[0]['index'], img[None])
it.invoke()
probs = it.get_tensor(it.get_output_details()[0]['index'])[0]
print(labels[probs.argmax()], probs.max())