Scholarlink

As artificial intelligence (AI) advances, many opportunities come into play for creating innovative and valuable applications. However, there are restrictions or high costs that limit the process of gathering data, particularly when it comes to rare bird images. This study explores using DenseNet121 for multi-label image classification in bird detection, with special emphasis given to cases when we have minimal images of endangered bird species. The International Union for Conservation of Nature’s (IUCN) Red List of Threatened Species has identified 44,000 species at risk of becoming extinct, focusing on bird populations facing an unprecedented decrease. Our objective is to develop a system that can detect and identify 10 classes of birds in a condition of limited image availability with a high level of accuracy and efficiency. This study utilizes a sample of 10 Indigenous Australian bird species to evaluate the performance of a classification algorithm on authentic images, computer-generated images, and a combination of the two. The work employs a reliable diffusion technique to generate artificial images by providing prompts encompassing each bird species’ vernacular and taxonomic names. The evaluation of the trained models shows that the real image-trained equivalent achieves an accuracy of 92.13%. Hence, we present a method that yielded good results and guided the way for using pre-trained models and fine-tuning with real and synthetic datasets generated with Diffusion models for bird identification (BI). This will help save a lot of time for scientists and researchers interested in processing, classifying, and identifying rare birds, especially those with less raw data or natural images.

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

Bird Identification and Classification Model with DenseNet and StableDiffusion on Limited Hybrid, Natural and Synthetic Datasets

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kushal Pokhrel,
  • Cesar Sanin,
  • Md. Kowsar Hossain Sakib,
  • Md Rafiqul Islam,
  • Edward Szczerbicki

摘要

As artificial intelligence (AI) advances, many opportunities come into play for creating innovative and valuable applications. However, there are restrictions or high costs that limit the process of gathering data, particularly when it comes to rare bird images. This study explores using DenseNet121 for multi-label image classification in bird detection, with special emphasis given to cases when we have minimal images of endangered bird species. The International Union for Conservation of Nature’s (IUCN) Red List of Threatened Species has identified 44,000 species at risk of becoming extinct, focusing on bird populations facing an unprecedented decrease. Our objective is to develop a system that can detect and identify 10 classes of birds in a condition of limited image availability with a high level of accuracy and efficiency. This study utilizes a sample of 10 Indigenous Australian bird species to evaluate the performance of a classification algorithm on authentic images, computer-generated images, and a combination of the two. The work employs a reliable diffusion technique to generate artificial images by providing prompts encompassing each bird species’ vernacular and taxonomic names. The evaluation of the trained models shows that the real image-trained equivalent achieves an accuracy of 92.13%. Hence, we present a method that yielded good results and guided the way for using pre-trained models and fine-tuning with real and synthetic datasets generated with Diffusion models for bird identification (BI). This will help save a lot of time for scientists and researchers interested in processing, classifying, and identifying rare birds, especially those with less raw data or natural images.