Re-identification of fish individuals of undulate skate via deep learning within a few-shot context

Journal: Ecological Informatics (Q1) 2023

Abstract

This paper proposes a few-shot Siamese Neural Network to identify undulate skate individuls from photos.

Pipeline

Transfer Learning (due to little data)(a pre-trained CNN) -> Data Augmentation -> Siamese Neural Network -> Emsemble Evaluation

Data

• Most of the images were captured using a reflex camera (objective 15-55 mm) and an action camera GoPro Hero 7 (extracting frames from 4K videos)

• Not sensitive to superficial features (e.g., shadows) and use visual apperance instead

• 1,138 images(1 - 31, everage 7 images for each individual) of 108 individuals (After preprocessing, 1219 images for 108 individuals)

• Test set: 370 images (4 recaptured individuals)