Abstract:
Due to high levels of noise and a low number of existing datasets, the classification of
SONAR images in underwater environments is an ill-posed problem. SONAR systems are
currently hamstrung by their lack of real-time processing and the inability to adapt to varying
underwater conditions, which has hindered their use in defence, marine research, and response operations. To address these issues, this project develops a real-time SONAR Image Classification System with noise adaptation and explainability features to make the system usable in decision-critical capabilities. The proposed approach is an extension of the AI model with noise adaptation procedures, which utilizes deep learning to classify the resulting SONAR images in real time. Supervised learning was applied in addition to self-supervised noise adaptation for the generalization of the model across different levels of noise. Exportability functions are also included in the model so that users may carry out an accurate interpretation of the results of the classification made. The model adds explainability functions that use Grad-CAM to visualize and explain classification decisions. Experimental evaluation was carried out using real-life SONAR data sets to determine the measure of classification and time response.
