Tricho-Vision

Computer vision for trichotaxonomy and wildlife conservation

Tricho-Vision applies computer vision to trichotaxonomy — the classification of mammalian hair microstructures — to enhance wildlife conservation of priority species. Published in Ecological Informatics (2025).

Motivation

Manual identification of species via hair (cuticle and medulla) analysis is time-consuming and requires expert knowledge. Automated vision pipelines can scale species identification from microscopy images, aiding biodiversity monitoring and wildlife-crime forensics.

Contributions

  • Curated the first benchmark dataset of 76 conservation-priority species, including critically endangered taxa
  • Classifies hair across four taxonomic levels — Order, Family, Genus, Species
  • Benchmarks CNNs, ViTs, and Swin Transformers; Swin Transformers perform best across all levels
  • Image cropping further improves accuracy by diversifying the training set

Dataset

Property Value
Species in dataset 76 (incl. endangered)
Taxonomic levels 4 (Order / Family / Genus / Species)
Architecture families 3 (CNN / ViT / Swin)
Best architecture Swin Transformer
Hair features analyzed Cuticle patterns + medulla

Results

Accuracy (%) across taxonomic levels — Swin Transformers lead, with Swin-V2-Base best at the fine-grained Species and Genus levels:

Method Species Genus Family Order Inf. (ms)
ViT-Base 70.60 74.90 87.48 95.84 4.02
EfficientNet-b3 69.49 73.71 86.73 95.11 12.17
DeiT-Base 70.03 74.96 87.45 95.24 12.29
BEiT-Base 70.07 74.92 87.72 95.98 15.70
Swin-Base 71.60 75.07 88.41 95.98 3.56
Swin-V2-Base 72.63 76.27 88.25 95.90 14.64
ResNet-50 66.90 72.88 86.35 94.99 15.96
MobileNetV2 40.38 68.66 80.90 92.92 17.51

Accuracy drops steadily from coarse (Order) to fine-grained (Species) labels, and the best Swin model stays consistently ahead of the CNN baseline at every level:

{
  "tooltip": { "trigger": "axis" },
  "legend": { "data": ["Swin-V2-Base (best)", "ResNet-50 (CNN baseline)"], "top": "2%" },
  "grid": { "left": "3%", "right": "4%", "bottom": "3%", "containLabel": true },
  "xAxis": { "type": "category", "data": ["Order", "Family", "Genus", "Species"] },
  "yAxis": { "type": "value", "name": "accuracy (%)", "min": 60, "max": 100 },
  "series": [
    {
      "name": "Swin-V2-Base (best)",
      "type": "bar",
      "data": [95.9, 88.25, 76.27, 72.63],
      "itemStyle": { "color": "#4f8ef7", "borderRadius": [4, 4, 0, 0] }
    },
    {
      "name": "ResNet-50 (CNN baseline)",
      "type": "bar",
      "data": [94.99, 86.35, 72.88, 66.9],
      "itemStyle": { "color": "#b9c4d0", "borderRadius": [4, 4, 0, 0] }
    }
  ]
}

Activation Maps

Grad-CAM activation maps from the best model across four orders (Carnivora, Primates, Rodentia, Pholidota): (a) input hair microscopy, (b–d) progressive layer activations highlighting the attended hair-shaft and scale regions.

Publication

Das, A., Banerjee, P., Biswas, S. et al. Tricho-Vision: The use of computer vision in trichotaxonomy for enhancing wildlife conservation of priority species. Ecological Informatics, 2025.

Work done at Habitat Lens Private Limited.