Despeckling-Based Augmentation for Noise-Aware SONAR Image Classification on Edge Devices

Abstract

Real-time sound navigation and ranging (SONAR) image classification on embedded platforms is critical for autonomous underwater monitoring and inspection. However, developing reliable classifiers for such deployments remains challenging due to severe speckle noise, limited annotated training data, and strict computational constraints on edge devices. To address these challenges, we propose a noise-aware training strategy for SONAR image classification that generates additional training samples by jointly exploiting noisy and despeckled image representations. By exposing the classifier to intermediate noise conditions between noisy and despeckled observations, the proposed strategy effectively enlarges the training distribution and promotes stable feature learning across varying noise levels. This formulation improves robustness in data-scarce SONAR imaging scenarios without increasing inference complexity. Experimental results on the 18-class Marine Debris Turntable Dataset improve MobileNet accuracy from 72.5% to 89.8% and SqueezeNet accuracy from 72.4% to 86.4%. The resulting lightweight classification pipeline is deployed on a Raspberry Pi 5 using open neural network exchange (ONNX) Runtime and INT8 quantization, achieving inference latencies as low as 2.29 ms while reducing model size by up to 74.8%. The proposed method further achieves 100% classification accuracy on the 12-class Marine Debris configuration, demonstrating its effectiveness for robust SONAR image classification on resource-constrained edge devices.