Fusing Local Shallow Features and Global Deep Features to Identify Beaks.

Area of Science:

• Marine Biology
• Computational Biology
• Ecology

Background:

• Cephalopods are vital to marine ecosystems, influencing resource development, ecological balance, and human food security.
• Accurate classification of cephalopods is crucial for sustainable resource management, with beak analysis being a common method.
• Existing classification methods require enhancement for improved accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a novel feature fusion method for accurate cephalopod beak identification.
• To compare the performance of different deep learning and traditional machine learning models for beak classification.
• To enhance the identification of cephalopod species through advanced computational techniques.

Main Methods:

• A feature fusion approach was developed, utilizing a convolutional neural network (CNN) architecture.
• Local shallow features, Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP), were extracted and classified using Support Vector Machines (SVM).
• Global deep features were extracted using the Resnet50 model, fused with HOG and LBP features, and then classified with SVM.

Main Results:

• The feature fusion model demonstrated effective integration of multiple features, leading to improved beak recognition accuracy.
• The HOG+Resnet50 combined feature method achieved the highest classification accuracies: 91.88% for upper beaks and 93.63% for lower beaks.
• Comparative analysis showed superior performance of the feature fusion approach over individual models.

Conclusions:

• The proposed feature fusion method significantly enhances the accuracy of cephalopod beak identification.
• The HOG+Resnet50 model offers a robust and accurate solution for classifying cephalopod beaks.
• This approach facilitates more effective identification studies, supporting cephalopod resource management and conservation efforts.