Year: 2026
Type: preprint
Abstract: Wepresent a 24,984-parameter sea-state classifier trained through recurrent refinement and deployed by direct spatial voting. Across ten archived encoders, voting averages 85.79% accuracy versus 85.65% with recurrence. A selected archived checkpoint with a floating-point encoder and latent-INT8 voting achieves 89.02%, versus MUSeNet’s published 88.7%, at approximately 1/80 of its parameters. Packing preserves the integer vote in 196 bytes. Instructioncycle accounting bounds the voting head at 182.5 µs on a 16-MHz ATmega328P; the encoder requires 23.71 million multiply–accumulate operations per image. A supporting Otter USV simulation links edge-processing budgets and ordinal recognition errors to supervisory response.
Authors: Hanyue Shen, Xinling Liao, Raju Datla
Institutions: Stevens Institute of Technology, King's College London
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225432114