A Naive ConvNet Model for Defect Identification and Grading of Mangoes

Abstract

Fruit sorting and defect recognition in fruits are active fields where remote fruit categorization is done without human supervision. Classifying mangoes by differentiating between good and poor ones is the main purpose of helping in the production process. Deep learning methods are being used for object classification and recognition. A Convolutional Neural Network (CNN) based model is used in this paper for defect identification and ripeness detection of mangoes. The models are experimented with by coordinating the convolution and pooling layers and using various activation and loss functions. The outputs, which come from the models, are compared for defect detection and ripeness classification. Mask R-CNN is used for localizing defects which is a transfer learning approach and a pixel-based segmentation technique. It predicted the mask for the defected part in the mango. Accuracy for defect detection is 85% on the validation set and 85.79% on the unseen test set. The final grading has an accuracy of 87.53%. Localization of defect is presented with an average Intersection over Union score which is 77.46%. This work can help in areas such as industries where mango sorting and defect identification are needed.