Classification of Infected Salmon Using CNN Deep Features and Optuna-Optimized SVM

Abstract

Fish diseases are a major challenge in the aquaculture industry, impacting productivity and the economy, particularly in salmon farming. This study aims to develop an image classification system for infected salmon using Convolution Neural Network (CNN) deep features approach and Support Vector Machine (SVM) classifier optimized with Optuna. The dataset consists of 1,208 images that were balanced through augmentation before being divided into 70% training data and 30% test data. Features were extracted from the middle layer of three pretrained CNN architectures: EfficientNetB1 (block6d_add), ResNet50 (conv4_block6_out), and VGG16 (block4_pool), then selected using the Least Absolute Shrinkage and Selection Operator (LASSO) method to address high-dimensionality issues. An SVM classification model was trained using stratified 5-fold cross-validation, both with default parameters and hyperparameter optimization results from Optuna. The results show that the model with features from EfficientNetB1 tuned by Optuna achieved the highest accuracy of 99.34%, a significant improvement over the default model 98.23%. Meanwhile, ResNet50 and VGG16 achieved optimal accuracies of 98.23% and 98.89%, respectively, after tuning. This study contributes to the development of an adaptive and accurate early detection system for infected fish.