Explainable Machine Vision Techniques for Fish Disease Detection with Deep Transfer Learning

Fish disease detection is crucial for maintaining the health and sustainability of aquaculture systems. Traditional methods of disease identification in fish often rely on manual observation, which can be inefficient and susceptible to humanly made mistakes. In recent years, machine learning techniques, particularly convolutional neural networks (CNNs), have shown promise in automating the detection process. However, the performance of CNNs can be limited by their reliance on fixed-size input images and their inability to capture long-range dependencies. In this work, five deep learning models such as the CNN, DenseNet121, ViT, VGG16 and VGG19 models are proposed to study their performance on our dataset using accuracy as the major evaluation metric. First, exploratory data analysis is carried out to understand the features and classes of our dataset, then preprocessing steps are carried out to resize them to a standard size, normalized the pixels as well as augmenting all classes to increase on variations of images for each class, to create a dataset balance for all classes. Then the dataset are splitted into training and validation sets and also extracted feature from images using VGG16 so as to capture the underlying dependencies in them. We trained all of our models on the training set then finally evaluated them to get their performance. The study implements deep transfer learning to all our models to improve the models' predictions and also solve the problem of poor generalization. Finally three different Explainable AI (XAI) techniques on each of the models and these include LIME, Grad-CAM and integrated gradient to gain an understanding on which exact features did the models base on to make predictions. Further more we investigated which XAI technique works better for each model. In our future prospects, more XAI techniques such SHAP and Layer-wise backward propagation are annalysed and also carry out our research together with engagements from domain experts to guide us on which evaluation metrics perfectly suit this sensitive study area.