Abstract
The increasing volume of organic waste in Indonesia, combined with the reliance on manual raw material selection for bio-organic fertilizer production, has become a major challenge in the modern petrochemical industry due to its susceptibility to human error. This study aims to design, evaluate, and optimize a deep learning-based organic waste detection model using the YOLOv5 algorithm to improve the accuracy and efficiency of automated raw material sorting. The research employed a quantitative experimental approach using the Research and Development (R&D) method. Model performance was enhanced through the Enhanced YOLOv5 approach by implementing Partial Layer Freezing on the backbone (freeze = [0]), hyperparameter optimization, and data augmentation. Performance evaluation was conducted by comparing the Baseline YOLOv5s model with several Enhanced YOLOv5 variants trained at different epochs (10, 15, 25, and 35 epochs) using Precision, Recall, mAP50, mAP50-95, and inference time as evaluation metrics. The experimental results indicate that the Enhanced YOLOv5 Epoch 25 model achieved the optimal performance (sweet spot), with a Precision of 99.53%, Recall of 100%, mAP50 of 99.50%, and an inference speed of 8.59 ms per image (approximately 116 FPS) while maintaining an efficient model size of 17.66 MB. This approach improved inference speed by 38.68% compared to the Baseline YOLOv5s model (14.01 ms) without compromising detection accuracy. Therefore, the Enhanced YOLOv5 Epoch 25 model is highly recommended for integration into conveyor-based automated raw material sorting systems in the petrochemical industry.