Enhancing occluded and standard bird object recognition using fuzzy-based ensembled computer vision approach with...

Area of Science:

• Ecology and Biodiversity
• Computer Science and Artificial Intelligence
• Machine Learning

Background:

• Accurate bird species classification is crucial for ecological studies and biodiversity conservation.
• Traditional classification methods are often labor-intensive and error-prone.
• Convolutional Neural Networks (CNNs) offer a more robust approach to image feature extraction and classification.

Purpose of the Study:

• To enhance bird species classification accuracy, particularly for occluded subjects, by developing a novel ensemble learning framework.
• To improve the generalization capability of bird image classification models.
• To address the limitations of existing methods in handling obstructed bird images.

Main Methods:

• An ensemble learning framework was developed by integrating top-performing CNN architectures (DenseNet and ResNet families).
• Fuzzy logic was employed to adaptively allocate model weights based on feature contributions.
• A dataset of 11,352 bird images from Caltech-UCSD Birds-200-2011 and Birds525 Species-Image Classification datasets was utilized, incorporating advanced augmentation techniques.

Main Results:

• The proposed fuzzy-based ensemble approach achieved 98.73% accuracy and a 98.75% F1-score for standard images.
• For occluded bird images, the model attained 95.78% accuracy and a 95.1% F1-score.
• Performance improvements of 2% for standard and 9% for occluded images were observed compared to existing methods, with 2-7% gains over individual CNN candidates.

Conclusions:

• The fuzzy-based ensemble learning framework significantly enhances bird species classification accuracy, especially for occluded images.
• The adaptive weighting mechanism based on fuzzy logic effectively improves model performance and reliability.
• Statistical validation confirmed the significance of the performance improvements, highlighting the method's potential for ecological applications.