Automated Apple Defect Detection Using Transfer Learning with Mobile Net

This study investigates for utilizing the MobileNet with transfer learning for automated detection and classification of flaws in “Chaubattia Anupam apple”. To distinguish between rotting, bruised, damaged and healthy apples, using machine learning and object detection methods, the model was trained on a dataset comprising 2,000 annotated images were preprocessed and improved for performance and consistency. The model's practical efficacy was demonstrated by its 90.4% accuracy for defected apples including (bruised apples, rotten apples and damaged apples), 56.4% accuracy for good apples. It ensured accurate defect identification with a mean Average Precision (mAP) of 92% (IoU 0.50–0.95) and 68% mAP at IoU 0.50. A custom built separator is used to separate the defected apples using MobileNet to deploy on Raspberry pi to send the signal in relay to separator. This approach makes easier with high accuracy, less data required and faster training.