Edge-AI Empowered Waste Detection

Resumo

This work optimizes the YOLOv11 object detector to enable real-time, high-precision waste sorting on resource-constrained embedded devices. The optimization is carried out in four steps, replacing standard convolution blocks with Ghost Conv layers and swapping the C2PSA and C3K2 modules for lighter alternatives, applying knowledge distillation from a full YOLOv11 model to a compact model to improve the accuracy of the compressed model, fusing neighboring layers to reduce memory access, and quantifying to 16-bit floating point (FP16) while generating builds for ONNX and NCNN. This optimization reduces the number of parameters by 46%. Compared to the full YOLOv11 model, the optimized version runs faster and uses fewer resources on the Raspberry Pi 4. On the TrashNet dataset, the inference latency is reduced from 1,077.1 ms to 329 ms, and further to 296.5 ms with the NCNN FP16 version. On the Drinking Waste Detection dataset, the latency decreases from 1,019.5 ms to 299.4 ms. Memory usage is reduced from 5.23 MB to about 3 MB, and energy consumption drops from 0.74 Wh to 0.29 Wh. In terms of performance, the mAP50 of the PyTorch model remains high, with 95.9% on TrashNet and 99.3% on Drinking Waste, closely matching the original YOLOv11 model's performance. These results show that our optimization approach creates a compact and accurate model suitable for embedded recycling systems running on modest hardware like the Raspberry Pi 4.