Elastic-Net regularized two-dimensional canonical correlation analysis for robust underwater sonar image classification

Abstract

For sonar image classification, it is critical to obtain effective representations from sensor data impaired by underwater interference. To address the loss of spatial structure inherent in vectorized canonical correlation analysis (CCA), we propose a two–dimensional Elastic Net regularized CCA (2D-ECCA), a sliding window formulation that preserves highlight–shadow geometry while embedding a mixed penalty ${\ell }_1+{\ell }_2$ directly inside the correlation objective. The proposed 2D-ECCA operates directly on local two-dimensional windows, thereby preserving spatial correlations as well as discriminative texture cues. Embedding an Elastic Net penalty in the correlation maximization objective further enhances robustness, suppresses overfitting, and yields more interpretable projections. A dedicated alternating minimization solver combines forward–backward updates for the projection blocks with a closed-form step for the shared latent matrix, guaranteeing monotone descent and complexity per iteration. Extensive experimental analysis on two public datasets and two experimental datasets validate the effectiveness and efficiency of the proposed algorithm. A field trial with an Oculus MT750d forward-looking sonar mounted on an underwater unmanned vehicle further confirms real-time capability (0.04 GFLOPs, 13 ms per frame) and 87.8 % accuracy at 1.2 MHz over the 40 m range. It is suitable for real-time deployment on resource-limited underwater platforms.