Classification Efficiency of Pre-Trained Deep CNN Models on Camera Trap Images.

Area of Science:

• Computer Science
• Artificial Intelligence
• Ecology

Background:

• Convolutional Neural Network (CNN) models are widely used in image recognition.
• Pre-trained models offer a potential solution for specialized image classification tasks.
• Wildlife monitoring often faces challenges like variable image quality and imbalanced datasets.

Purpose of the Study:

• To evaluate the performance of 36 pre-trained CNN models on a real-world wildlife image classification task.
• To assess the effectiveness of these models in identifying the Eurasian lynx (Lynx lynx).
• To understand the impact of image quality and dataset imbalance on model performance.

Main Methods:

• Utilized 36 distinct pre-trained CNN models.
• Trained and evaluated models on a dataset of camera trap images from Croatia.
• Focused on binary classification: presence or absence of Eurasian lynx.
• Analyzed performance metrics considering image quality and dataset imbalance.

Main Results:

• Significant variation in performance was observed across the 36 evaluated CNN models.
• Model performance was affected by the highly imbalanced nature of the dataset.
• Image quality issues presented a notable challenge for accurate classification.
• Some models demonstrated potential for real-world application despite limitations.

Conclusions:

• Pre-trained CNN models show promise for wildlife image classification but require careful selection and potential fine-tuning.
• Dataset characteristics, such as imbalance and image quality, are critical factors influencing model success in ecological applications.
• Further research is needed to optimize CNN performance for challenging, real-world wildlife monitoring scenarios.