Efficient Vision for Precision Sericulture: Bridging the Reality Gap in Silkworm Rearing via Interpretable Feature Streams

Vision systems for precision agriculture require high accuracy and efficiency for deployment on resource-constrained platforms. In this work, we introduce a novel dataset and a resource-efficient vision framework for silkworm rearing monitoring and feeding, a domain challenged by severe visual clutter, occlusion, and morphological plasticity. To bypass the prohibitive cost of manual pixel-level annotation, we propose an Environment-Prior Pipeline that extracts interpretable occupancy statistics using domain-specific visual heuristics. When compared against end-to-end Deep Learning baselines, our approach achieves highly competitive classification performance while drastically reducing computational complexity, offering a practical solution for embedded agricultural monitoring.