A Comparative Evaluation of Hybrid Fusion Strategies for Turtle Identification

Mohd Erman Safawie Che Ibrahim, Wan Nural Jawahir Hj Wan Yussof, Muhammad Suzuri Hitam, Ezmahamrul Afreen Awalludin, Mohamad Fathullah Ruslan, Siti NurFarahim Shaharudin
dx.doi.org/10.17576/apjitm-2026-1501-21

Abstract

This paper presents a comparative study of Coiflet1 wavelet- and Gabor-based fusion strategies combined with deep learning for individual sea turtle identification. The visual identification of marine turtles continues to be challenging because of the small amount of annotated data, high inter class similarity and class imbalance. To achieve this, a feature to image transformation framework is presented in which statistical descriptors obtained from Coiflet1 wavelet transforms and Gabor filters are restructured into surrogate image representations and then fed into a pre-trained ResNet-18 convolutional neural network. Four fusion schemes are applied which are Coiflet1 only, Gabor only, sequential Coiflet1 and Gabor, and parallel Coiflet1 and Gabor. All models are tested on a custom dataset that consists of 1,426 labeled images of turtles in 20 different identities. We evaluate the performance using several evaluation metrics, including accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, Cohen kappa, and training time. Experimental results have shown that all the models show impressive performance (accuracies more than 93%). The Coiflet1 only configuration outperforms the others with 97.57% accuracy and a Cohen?s kappa of 0.97. It also achieves the best precision (99.12%), recall (94.83%), and F1 score (96.37%), while the Gabor-only (96.53% accuracy) and hybrid fusion configurations both trail. These results show that if structured as surrogate images that independently reformulate handcrafted wavelet features, it does better than those in more complex fusion schemes in resource-constrained deep learning. The proposed framework sets a new quantitative level for hybrid wavelet-CNN architectures and provides insight into the trade-offs among fusion complexity, model generalization and computational efficiency.

keyword

Deep Learning; Wavelet Transform; Gabor Filters; Feature Fusion; Ecological Monitoring

Area

Pattern Recognition