Smart Vision System for Enhanced Food Waste Management Using Black Soldier Fly Larvae

Abstract

Food waste remains a critical global challenge requiring sustainable and efficient processing methods. This study presents a low-cost embedded vision system designed to assist Black Soldier Fly Larvae (BSFL) bioconversion by automatically classifying food waste based on moisture content. Using a Raspberry Pi 4 and a convolutional neural network (CNN), the system categorizes waste into dry (<30%), optimal (30–70%), and wet (>70%) classes and actuates a motorized wiper to route feedstock accordingly. A dataset of approximately 500 labeled food waste images was collected under controlled laboratory lighting to train and validate the model. The CNN achieved 98.16% accuracy during training, while performance on unseen validation images reached 85.04%, reflecting moderate generalization given the dataset size and moisture-related visual ambiguity. In pilot conveyor-belt trials—where lighting and motion introduced additional variability—the system maintained ~84–85% accuracy and processed each frame in about 1.2 seconds, enabling functional real-time sorting and reducing manual labor by approximately 89%. Costing under USD 150, the prototype demonstrates the potential of affordable edge-AI systems for moisture-sensitive waste pre-sorting in small-scale BSFL operations. However, its robustness remains constrained by dataset size, controlled imaging conditions, and environmental variability. Future development will require larger and more diverse datasets, improved regularization techniques, and hardware optimization to enhance generalization and support broader deployment aligned with Sustainable Development Goal 12.3.