The salmon industry is one of the important productive sectors of the country and is constantly growing expansion . In 2023, salmon production in Chile reached 1,089,924 tons, this leads to countless problems, including diseases and the subsequent death of fish. The skin of fish is responsible for important functions such as protection against physical damage, maintenance of homeostasis, and constitutes the first immunological barrier. Diseases in fish in aquaculture have bacterial, viral, and parasitic origins, which spread rapidly in high-density populations in artificial systems. Convolutional neural networks (CNN ) have been used to identify species, perform biometric studies, and for effective diagnosis of sick fish using high-quality fish images. The present work aims to diagnose salmon with skin lesions or without skin lesions (Yes/No, respectively) by image classification with a CNN technique and a pre-trained VGG16 architecture using freeze and not unfreeze layers.
One hundred images of salmon were classified into healthy fish (without skin lesions) and diseased fish (with skin lesions). The images were sized to 224x224 pixels. The image and label lists were then converted to numerical values, and the images were then normalized. The dataset was split into training (80%) and test (20%) sets.
Subsequently, random transformations were applied to the training images. These included: rotation (15 degrees), cropping (0.2), zooming (0.2), horizontal flipping, and pixel filling with the ’ nearest ’ mode. The technique used was Transfer Learning using a pre-trained architecture: VGG16. To compile the model, an Adam optimizer was used with a learning rate of 1 × 10−3, a loss function: The appropriate one for binary classification with softmax output, and Accuracy was used as metrics to monitor performance. 8 Batch and 90 epochs were used as training hyperparameters .
Results (Table 1) show that the pre-trained architecture VGG16 with freeze layers (method 1) has a higher accuracy than the VGG16 method with unfreeze layers (method 2) (0.81 vs 0.52). Similarly, method 1 is able to classify fish without skin lesions (No), but not method 2, where the precision, Recall and F1-score for the category “No” have a value of 0.
These results show that it is possible diagnostic or sort disease salmons with skin lesions using Convolutional Neural Network (CNN) and transfer learning. Its necessary improve classification accuracy for a implement it in the salmon industry.