Enhancing Apple Fruit Quality Detection with Augmented YOLOv3 Deep Learning Algorithm

Precise apple detection is essential in the food manufacturing industry to provide quality control in production lines for differentiating between fresh and damaged apples. Various apple detection difficulties are found even before harvest in today's environment. However, post-harvest evaluation is still crucial for identifying apple species and assessing quality to expedite food processing procedures. This study presents a sophisticated detection model for multi-class apple recognition to distinguish between regular, damaged, and red delicious apples. The proposed model enhances Augment-YOLOv3 by integrating background removal through GrabCut, thereby improving object localization. Additionally, extra spatial pyramid pooling and a Swish activation function are incorporated to optimize feature retention during training. The YOLOv3 framework is refined using the Darknet53 backbone with feature pyramid network-based spatial pooling, ensuring superior feature extraction before object detection. The final classification layer precisely distinguishes between apple categories. Experimental evaluations reveal that the Augment-YOLOv3 model achieves a mean average precision (mAP) of 98.20%, outperforming conventional YOLOv3 and YOLOv4 models. The study leverages a newly curated Kaggle dataset, utilizing Google Colab with an NVIDIA Tesla K-80 GPU for inference, ensuring precise object localization and robust multi-object detection performance.