Elasmobranchs, a group of cartilaginous fishes that includes sharks and rays, are among the most threatened marine taxa. Southeastern Spain has recently been identified as a key conservation area, despite the fact that more species have been documented than are officially recognized [7]. This underscores the urgent need for tools to support species identification in fish markets and fisheries-environments in interaction with the marine ecosystem. However, the lack of annotated datasets and the common practice of grouping species under generic labels hinder precise classification and effective conservation strategies. To address this, we propose an informed classification framework based on a zero-shot multimodal strategy using the CLIP model. Our approach integrates knowledge in the form of expert visual descriptions and schematic illustrations that highlight distinctive traits. This enables inference grounded in expert sources and field guides, promoting both interpretability and adaptability in emerging monitoring and management contexts. The code is available at: https://github.com/Tech4DLab/e-Lasmobranc-project .

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Informed Zero-Shot Elasmobranch Classification Using Expert Knowledge and Illustrations

  • Ismael Beviá-Ballesteros,
  • Mario Jerez-Tallón,
  • Bernabé Sánchez-Sos,
  • Nieves Aranda-Garrido,
  • Isabel Abel-Abellán,
  • Irene Elvira Antón-Linares,
  • Marcelo Saval-Calvo,
  • Francisca Giménez-Casalduero,
  • Jorge Azorín-López,
  • Andrés Fuster-Guilló

摘要

Elasmobranchs, a group of cartilaginous fishes that includes sharks and rays, are among the most threatened marine taxa. Southeastern Spain has recently been identified as a key conservation area, despite the fact that more species have been documented than are officially recognized [7]. This underscores the urgent need for tools to support species identification in fish markets and fisheries-environments in interaction with the marine ecosystem. However, the lack of annotated datasets and the common practice of grouping species under generic labels hinder precise classification and effective conservation strategies. To address this, we propose an informed classification framework based on a zero-shot multimodal strategy using the CLIP model. Our approach integrates knowledge in the form of expert visual descriptions and schematic illustrations that highlight distinctive traits. This enables inference grounded in expert sources and field guides, promoting both interpretability and adaptability in emerging monitoring and management contexts. The code is available at: https://github.com/Tech4DLab/e-Lasmobranc-project .