Multi-Label Classification and Deep Image Analysis Methods on Marine Microalgae and Rotifer Identification Using Transfer Learning and Two-Dimensional Convolutional Neural Networks

Microalgae form the base of the food web in aquatic ecosystems and holds great promise as a source of alternative energy production. However, their consumption by pests, and rotifers with the potential to devastate the microalgal ‘crop’, is a commercial bottleneck to the developing industry. Early detection of invaders through deep learning methods is important, albeit challenging. This paper reports on research to design a deep learning image processing architecture that classifies and extracts the rotifer-infested features from microalgae images. A 2-dimensional convolutional neural network ( Conv2D ) was trained using transfer learning techniques along with fine-tuning and image augmentation methods. The model, deep learning ‘ TA-Conv2D’ , has a CNN -based classifier that automatically detects the important features e.g., rotifers, and other anomalies while transfer learning optimises the model’s efficiency. The addition of fine-tuning utilises a pre-trained network and augmentation to boost the accuracy of the resultant model during the feature extraction process. We benchmark the proposed model with further five case models: Dropout Regularisation, Data Augmentation, Visual Geometry Group ( VGG-16) feature extraction model, and VGG-16 feature extraction with weights tuning. The study uses metrics of cross-entropy loss and F-beta score to evaluate the proposed model’s accuracy and register good results with increased accuracy and minimum loss for the proposed deep learning architecture TA-Conv2D , outperforming alternative approaches. The extensive empirical evaluation of this study, model analysis, and valuable insights using multi-label classification and deep image analysis methods can be extended to design high-performance computer-aided detection systems for medical imaging or other image processing tasks