A comparative study on image-based snake identification using machine learning.

Area of Science:

• Computer Science
• Zoology
• Ecology

Background:

• Automated snake identification is crucial for snake bite management and public safety.
• Accurate identification can help prevent venomous snake encounters and guide appropriate medical treatment.
• Current methods lack comprehensive comparison across various machine learning techniques.

Purpose of the Study:

• To compare the accuracy of state-of-the-art machine learning algorithms for snake image identification.
• To evaluate both holistic and deep learning approaches for classifying six snake species.
• To identify the most effective algorithm for practical applications, including mobile devices.

Main Methods:

• Employed holistic methods: k-nearest neighbors (kNN), support vector machine (SVM), and logistic regression (LR).
• Utilized dimension reduction techniques: principle component analysis (PCA) and linear discriminant analysis (LDA) as feature extractors.
• Compared holistic methods with deep learning models, specifically convolutional neural networks (CNNs) like MobileNetV2 and VGG.

Main Results:

• Holistic methods combined with PCA yielded less than 50% accuracy.
• Linear Discriminant Analysis (LDA) significantly improved holistic method performance, with LDA-SVM achieving 84% accuracy.
• Convolutional Neural Networks (CNNs) demonstrated superior performance, with MobileNetV2 reaching 93.16% accuracy.
• Analysis of VGG model activations indicated that deep layers use color and shape for species discrimination.

Conclusions:

• MobileNetV2 is a highly effective deep convolutional neural network for snake image classification.
• The study highlights the potential for developing mobile applications for real-time snake identification.
• Accurate automated identification can significantly enhance snake bite management and prevention strategies.