ADL-Net: attention-driven densenet-LSTM network for underwater images classification

Abstract

Listen

Translate article

Abstract Imaging in marine environment is a challenging task due to several reasons, including light variations, color distortions, noise, and limited datasets, which causes inaccurate target classification problems. This paper presents a methodology for image classification in marine biology and environmental research. We proposed the Attention-Driven DenseNet-LSTM Network (ADL-Net), aiming to improve classification accuracy of underwater imagery. Initially, feature extraction is performed using multi-levels of DenseNet201, which excels in hierarchical feature extraction, offering stability and rapid convergence. In the next stage, two modified attention mechanisms are designed: the squeeze and excitation operations are used to refine channel-specific feature responses, while the convolutional block attention module refines attention for both channels and spatial dimensions. These attention mechanisms help the model to highlight important features and minimize distractions. Additionally, a Bi-directional Long Short-Term Memory layer is integrated to capture spatial dependencies and improve classification robustness. Various training strategies are used to find the best parameter tuning. Our method demonstrates excellent performance when tested on fish datasets: LifeCLEF 2015 and Fish4Knowledge, achieving accuracies of 98.02%, and 99.52%, respectively. These findings underscore ADL-Net's potential for enhancing automated underwater image classification, advancing research in marine ecology.