Authors: M. Anbarasan, Dr. P. Guhan
Abstract: Fruit quality inspection plays a critical role in reducing post-harvest losses and ensuring consumer safety in the agricultural supply chain. Conventional manual inspection techniques are time-consuming, manual and ineffective on larger scales. To address these constraints, the paper introduces a model of identifying fruit quality using deep learning techniques that employ methods of digital image processing. The model exploits two-stage and evaluation procedure including classification and detection operation. We used pre-trained DenseNet networks with transfer learning to divide the fruits into three quality levels of Fresh, Overripe, and Spoiled quality. The method of image preprocessing normalization, filtering, and augmentation were used to increase the model robustness. The DenseNet model had an evaluation accuracy of 97.82%, which was higher as compared to SVM (89.53%) and Random Forest (90.21%) which are the conventional classifiers. Parallel to it, we also tested object detection models such as YOLOv8 to recognize and bound fruits with bounding boxes and label quality. YOLOv8 was revealed to be very fast with an average precision (mAP) of 96.1% and intersection over union (IoU) of 87.3%. It was also calculated that precision, recall, F1-score, and the time of inference were taken across 10 models. Findings confirm the efficiency of deep learning in automating the process of fruit quality determination to consequently deploy real-time applications in separating systems. The presented model is very flexible to other types of agricultural products and compatible with smart farming and automation processes that include retailing. Generally, this work fills the nexus between manual inspection and smart visual systems by making the fruit quality monitoring scalable, consistent, and efficient.
