<![CDATA[Recyclable waste management in Indonesia is often hindered by the inefficiency of manual sorting and lack of public awareness. This study aims to implement the YOLO11n model based on transfer learning to detect six types of recyclable waste and analyze the effect of layer freezing strategies on model accuracy and efficiency. An experimental method was employed utilizing the Recycle Trash Dataset of 2,462 images divided into three training scenarios: Full Transfer Learning, Partial Backbone Freezing (5 layers), and Full Backbone Freezing (10 layers). The results indicated that the Partial Backbone Freezing strategy delivered the best performance with an mAP50 of 0.840 and Precision of 0.823, showing statistical significance ($p<0.05$). The model demonstrated superior performance on the Metal class (mAP50 0.970) but faced significant challenges with the Organic class (Recall 0.496) due to high shape variation and the Plastic class due to material transparency. Testing on Raspberry Pi 5 showed an average inference time of 261 ms with a model size of 5.9 MB. In conclusion, the Partial Backbone Freezing technique applied to YOLO11n has proven to be the most effective in balancing generic feature extraction and semantic adaptation, making it suitable for implementation on resource-constrained edge devices.]]>
Copyrights © 2026
