Explainable multi stream deep learning for fine grained camel breed classification using a Novel Arabian and Non...

Area of Science:

• Computer Science
• Artificial Intelligence
• Zoology

Background:

• Camels are vital to arid ecosystems and desert communities.
• Accurate identification of camel breeds, especially Arabian camels, is challenging due to visual similarities.

Purpose of the Study:

• To introduce a novel dataset of Arabian and Non-Arabian camel images.
• To propose an explainable multi-stream deep learning architecture for fine-grained camel breed classification.
• To establish a foundation for automated camel breed identification and livestock management.

Main Methods:

• A dataset of 1,620 camel images was collected and annotated.
• A two-stage hierarchical adaptive framework was developed: binary classification (Arabian vs. Non-Arabian) and multi-class classification (five Arabian breeds).
• A multi-stream deep learning architecture utilizing DenseNet121, online data augmentation, class-balanced focal loss, Adam optimizer, and Grad-CAM for interpretability was employed.

Main Results:

• The DenseNet121 model achieved 98% accuracy in binary classification and 76% in multi-class classification.
• The multi-stream design enhanced feature extraction and classification accuracy.
• Grad-CAM visualization provided transparency into the model's decision-making process.

Conclusions:

• The proposed explainable AI model demonstrates high accuracy in identifying camel breeds.
• The developed dataset and methodology provide a strong basis for future research in automated livestock identification.
• This technology has the potential to significantly improve livestock management practices.