Publication

Abstract : The accurate identification of fish species is an important endeavour in the study of marine biodiversity and underwater ecosystems. This paper presents a hybrid classification framework that integrates handcrafted phenotypic and deep learning-based features. The resultant classification framework utilizes images obtained from the Fish4Knowledge dataset and uses masks to ensure the extraction of features only from the fish areas within the images. The masks were used to segment the photographs into fish and non-fish sections, enabling the extraction of morphologic and spatial characteristics (shape, texture, colour) along with six pretrained CNNs (VGG16, ResNet50, MobileNetV3, Xception, EfficientNetB0, InceptionV3) from deep neural networks. Subsequently, all the extracted features were subjected to dimensionality reduction using PCA and were evaluated by the following classifiers: SVM, Random Forest and XGBoost. The highest accuracy was achieved with the combination of handcrafted features and the Xception embeddings, which provided an overall classification accuracy of 97.42% when combined with an SVM classifier. When compared to other studies, the hybrid approach described in this paper has shown to be competitive in providing a practical, effective methodology for classifying underwater fish species.