Convolutional Feature Vectors and Support Vector Machine for Animal Sound Classification.

Area of Science:

• Machine Learning
• Bioacoustics
• Pattern Recognition

Background:

• Deep learning networks excel at pattern classification but struggle when training data lacks class discernibility.
• This limitation is evident in tasks like animal sound classification, where similar sounds lead to sharp accuracy drops.

Purpose of the Study:

• To propose a novel deep learning approach to overcome the limitations of conventional methods in classifying animal sounds with low inter-class discernibility.
• To improve the accuracy and robustness of pattern classification in challenging datasets.

Main Methods:

• A novel approach combining multiple Convolutional Neural Networks (CNNs), each pre-trained for specific classes to generate mid-level features.
• Features from individual CNNs are merged into a combined CNN unit.
• Support Vector Machine (SVM) is employed for the final overall classification.

Main Results:

• An animal sound database comprising 3 classes and 102 species was established for experimental validation.
• The proposed method demonstrated superior performance compared to prominent conventional methods on the established dataset.
• The approach effectively addresses the learning failures caused by poor class separation in training data.

Conclusions:

• The proposed hybrid deep learning model effectively classifies animal sounds, outperforming existing methods.
• This approach offers a robust solution for pattern classification tasks with limited inter-class feature distinctiveness.
• Further research can explore its application in other domains requiring fine-grained classification.