Convolutional neural network modelling and image analysis techniques for the detection of fish diseases

Abstract

Fish rearing or pisciculture holds the key to food security and economic well-being for several countries across the globe. Diseases in fish are the biggest threat due to their rapid spread rate and high calamity, leading to a sharp decrease in fish yield. Deep learning techniques such as Convoluted Neural Networks (CNNs) hold an extremely promising impact on disease detection and raising the predictability of production amount. In this paper, the sequential CNN model indicates rigour and high reliability. The experimental setup consists of a model based on TensorFlow, Keras and 8-core TPU to accelerate computational Machine Learning (ML) tasks. The dataset obtained from Kaggle consists of 457 files depicting seven distinct classes (one healthy and six diseased classes), including all major fish diseases. Image data preprocessing is done by resizing and rescaling to train the optimised model. Image augmentation is done to expand the available data set and resolve overfitting issues within the CNN model. Modelling involves multiclass classification with a convolutional layer to extract features, keeping non-linearity in the model by an activation function. By transfer learning, inadequacies of the dataset are minimised. The proposed CNN model architecture efficaciously classifies various fish diseases. Identification and categorisation are done using Python, and the algorithm’s learning efficiency is predicted to be quite high. The model makes reasonably accurate predictions with an accuracy close to 91 %. A good pattern of learning during the training of the model is observed. These observations indicate the model's remarkable capacity to correctly identify the diseases.