Abstract
This paper presents a process for classifying predatory birds by
family and species. The motivation for this study arises from the
high variability observed among birds of different species and the
importance of performing classification efficiently and in a timely
manner. Additionally, this work aims to analyze the impact of using
RGB channels in comparison to grayscale images on classification
performance, as well as the effect of applying data augmentation
techniques during training. The dataset contains 42,475 images,
distributed across 6 families and 41 species. The process employs
fine-tuning, using the ResNet-50 model. Early stopping was applied
to control overfitting and obtain the best model. The test results
highlight the effectiveness of the proposed process in classification
tasks, with performance varying across different input configurations.
For species classification, the model trained with grayscale
channels achieved an F1-Score of 0.80. Using RGB channels improved
the performance significantly, resulting in an F1-Score of
0,86. Further applying data augmentation techniques to the RGB
slightly improved the metrics, achieving an F1-Score of 0.87. These
results demonstrate the benefits of incorporating color information
and data augmentation in enhancing classification accuracy.