LiteDebrisNet: A Light-Weight Hybrid Architecture of Real-Time Underwater Marine Debris Detection in Shallow Waters of a Coastal Environment

Abstract

Marine litter along passive coastal zones is a growing problem. The Andhra Pradesh coast runs 974 km along the Bay of Bengal with a reported litter density near 1.147 items per square meter, and India emits about 9.3 million tonnes of macroplastic waste each year. Autonomous underwater vehicles offer a scalable way to monitor this litter, but full-size detectors exceed the compute budget of embedded hardware. The LiteDebrisNet, a compact detector that pairs a MobileNetV3-Large backbone with the GELAN detection design of YOLOv9, preceded by a physics-informed color correction module based on the Beer-Lambert attenuation model. On the TrashCAN 1.0 instance dataset (22 classes), evaluated across five random seeds, LiteDebrisNet reaches 86.0 percent mAP at IoU 0.5, an improvement of 1.6 points over a YOLOv9c baseline trained under identical conditions (paired t-test, p below 0.001, 95 percent CI 1.16 to 2.04), while using 4.5 M parameters against 25.3 M, a reduction of 82 percent. The color correction stage adds 8.5 ms per frame, so end-to-end speed is 56.8 FPS on an RTX 3090, lower than both baselines. LiteDebrisNet therefore trades inference speed for accuracy and parameter efficiency. All turbidity results in this work use simulated Beer-Lambert degradation, and timing was measured on a desktop GPU, not embedded hardware.